A Smart Image Sensing and Control Method and System for Electromechanical Construction in Data Centers

Abstract

This application discloses an intelligent graph-based perception and control method and system for electromechanical construction in data centers, belonging to the fields of building engineering informatization and artificial intelligence technology. The method includes: collecting multi-source construction data; predicting construction progress; monitoring quality and safety; fusing data and verifying standards; multi-engine collaborative perception; generating an analysis result set; and feedback for model retraining. The system adopts a bottom-up five-layer architecture, including a multi-source data acquisition layer, a data fusion and governance layer, a graph-based temporal collaborative perception layer, a federated knowledge management layer, and an application service layer. This application solves the technical problems of missing process dependencies, difficulties in cross-project knowledge transfer, small sample recognition bottlenecks, and rigid data access in existing technologies. It can be used for intelligent control of the entire electromechanical construction process in data centers, improving the accuracy of progress prediction and monitoring efficiency.

Description

Technical Field

[0001] This application belongs to the field of building engineering informatization and artificial intelligence technology, specifically involving an intelligent control method and system for the whole process management of data center electromechanical engineering construction. Background Technology

[0002] Data center electromechanical construction encompasses over ten professional systems, including power supply and distribution, HVAC, and intelligent low-voltage systems. It involves hundreds of overlapping processes and multi-disciplinary collaboration, making it the most complex and challenging stage in data center construction. With the comprehensive advancement of the "Eastern Data, Western Computing" project, the scale of data center construction is experiencing explosive growth, placing higher demands on accurate progress prediction, real-time quality monitoring, and scientific resource allocation for electromechanical construction. However, complex non-linear dependencies exist between the processes in data center electromechanical construction—parallel construction, resource competition, spatial conflicts, and regulatory constraints are intertwined. Traditional experience-based, extensive management is insufficient to meet the multi-objective control requirements of modern data center construction regarding schedule, quality, and cost. Therefore, developing intelligent sensing and control technologies for the entire data center electromechanical construction process is of significant practical importance for ensuring project quality, shortening construction cycles, and reducing construction costs.

[0003] Construction schedule and quality control technologies have roughly gone through three stages of development. In the early stages, project management mainly relied on the Critical Path Method (CPM) and Work Breakdown Structure (WBS) for schedule planning, supplemented by manual inspections and experience-based judgment for quality acceptance. This method treated processes as a linear sequence, unable to adapt to the dynamic changes on the construction site. With the popularization of Building Information Modeling (BIM) and Internet of Things (IoT) technologies, construction management entered a data-driven stage. Various sensors, video surveillance, and BIM models enabled the collection and visualization of multi-source data. However, schedule prediction and quality inspection at this stage still relied mainly on rule engines and statistical analysis, lacking the ability to deeply model the complex dependencies between processes. In recent years, deep learning technology has been gradually introduced into the field of construction management. Researchers have begun to explore applying models such as Long Short-Term Memory Networks (LSTM) and Spatiotemporal Graph Convolutional Networks (ST-GCN) to construction schedule prediction, applying improved YOLO models to defect detection on construction sites, and attempting to use federated learning for secure sharing of engineering data. However, these technologies are mostly direct applications of general-purpose AI algorithms to construction scenarios, and have not yet formed a systematic solution specifically for the electromechanical construction of data centers.

[0004] Existing technologies face a technical challenge when applied to data center electromechanical construction: the lack of a graph-based perception and fusion management system covering the entire construction process. Data center electromechanical construction involves multi-dimensional non-linear dependencies among its processes, including logical dependencies, resource competition, spatial conflicts, and regulatory constraints. Construction data is scattered across heterogeneous data sources such as Building Information Modeling (BIM), IoT sensors, video surveillance, and construction logs, while construction specifications exist in unstructured text form. Existing technologies employ linear sequence models for progress prediction, general object detection models for quality inspection, federated learning frameworks for cross-project knowledge transfer, and rule engines for compliance verification. However, these technologies lack a unified graph structure semantic space and data closure mechanism, hindering the collaborative work of process dependency modeling, multi-source data fusion, regulatory constraint embedding, and cross-project knowledge transfer. This makes it difficult to achieve holistic intelligent management of the entire data center electromechanical construction process. Summary of the Invention

[0005] To address the problems existing in the background technology, this application provides an intelligent image perception and control method and system for electromechanical construction of data centers.

[0006] This application provides an intelligent map-based perception and control method for electromechanical construction in data centers, comprising the following steps: Step S1 involves collecting multi-source construction data and fusing it to obtain fused multi-source construction data. The multi-source construction data includes at least Building Information Modeling (BIM) data, real-time data from IoT sensors, construction logs, video surveillance streams, drawing files, and laser point cloud data. It is accessed through multiple data acquisition interfaces and supported by a unified data acquisition adapter, utilizing communication protocols such as Hypertext Transfer Protocol (HIP), Message Queue Telemetry Transmission Protocol (MQTT), Modbus Protocol, OPC Unified Architecture Protocol (OPC), and Real-Time Streaming Protocol (RTS). The fusion processing is performed by the data fusion and governance layer, including an event-driven extraction, transformation, and loading pipeline architecture to extract, transform, and load the multi-source heterogeneous data. This includes industrial basic class parsing, portable document format drawing optical character recognition, video keyframe extraction, and semantic alignment based on a standard knowledge graph. The loaded data is written to different storage engines according to its type: time-series data is written to InfluxDB, relational data to PostgreSQL, graph data to Neo4j, document data to MinIO object storage, and vector data to Milvus. The fused multi-source construction data is then passed to the graph-time-series collaborative perception layer for analysis and judgment.

[0007] Step S2: Based on the fused multi-source construction data, a graph neural network progress prediction engine is used to predict the construction progress. Specifically, this includes: constructing a process dependency graph; obtaining the process list for the electromechanical construction project of the target data center; using each process as a node; constructing a directed heterogeneous graph containing logical dependency edges, resource competition edges, spatial conflict edges, and specification constraint edges; each node containing a 64-dimensional attribute vector; constructing a specification knowledge graph; and based on the specification knowledge graph, using the TransE model to embed specification clauses, related processes, and constraint triples into a distributed representation in a low-dimensional vector space, and concatenating the specification constraint vectors to the attribute vectors of the corresponding process nodes; using a relational graph convolutional network to perform message passing on the process dependency graph; using independent weight matrices for message aggregation of the four types of edges to obtain a deep representation of each process node; introducing a Transformer-style multi-head temporal attention mechanism to calculate temporal attention weights; connecting multiple task-specific prediction heads to perform critical path duration prediction, duration deviation probability distribution prediction, and risk assessment. The system employs multi-task joint prediction for risk warning level assessment and resource conflict warning. The total loss function for multi-task joint training is the weighted sum of the loss terms of each task. Cross-project knowledge transfer is achieved through federated learning, which is divided into a local training phase and a global aggregation phase. These two phases are completed collaboratively by the graph temporal collaborative perception layer and the federated knowledge management layer, respectively. In the local training phase, the graph neural network progress prediction engine in the graph temporal collaborative perception layer trains the model using the project's historical construction data on each project's local server. The trained model parameters are then uploaded to the federated knowledge management layer. In the global aggregation phase, the federated knowledge management layer adds differential privacy noise to the model parameters uploaded by each project and performs federated average aggregation. The aggregated global model is then distributed to the local graph neural network progress prediction engine of each project for personalized fine-tuning. Knowledge distillation and model pruning techniques are used to compress the trained complex teacher model into a student model, adapting it for edge computing device deployment.

[0008] Step S3: Based on the fused multi-source construction data, a neural radiation field and building information model (BIM) image recognition engine is used to perform real-time quality and safety monitoring of the construction site. Specifically, this includes: acquiring multi-view images of the construction site using ultra-high-definition network cameras deployed at key work surfaces; using these multi-view images as input to the neural radiation field; the neural radiation field network takes the coordinates of a 3D point and the viewing direction as input, and outputs the volume density and color of that point; and reconstructing the 3D scene using voxel rendering technology. Corresponding control points are selected in both the reconstructed 3D scene and the BIM model, and the rigid body transformation matrix is ​​solved using an iterative nearest-point algorithm to achieve spatial alignment. Virtual training samples are generated in the reconstructed 3D scene through perspective transformation, lighting perturbation, and occlusion simulation. Five hundred and forty virtual samples are generated for each real image. These virtual samples carry spatial semantic labels from the BIM model and have geometric consistency. An improved YOLOv11 network is used to train a mixture of real and virtual samples. kbone employs the C3k2 mechanism for multi-scale feature extraction and embeds a pyramid spatial attention module. The Neck layer uses a bidirectional feature pyramid network. The detection head adds a P2 detection head with a resolution of 104x104 on top of P3 and P4, and injects building information model semantic vectors into the detection head. The trained model is deployed on edge computing devices, and adaptive cropping is performed based on the preset region of interest in the building information model. The TensorRT engine is used for half-precision floating-point quantization and eight-bit integer calibration inference, with a single-frame inference time of less than 30 milliseconds. Based on the detection results and the semantic information of the building information model, a tiered warning system is implemented: green for normal status, yellow for general alarms, and red for emergency warnings. In case of an emergency warning, the system triggers an audible and visual alarm, saves video evidence, highlights the defect location in the building information model, and simultaneously pushes warning information to on-site management and safety personnel.

[0009] Step S4 involves real-time compliance verification of the fused multi-source construction data based on the aforementioned standard knowledge graph. Specifically, this includes: storing the standard knowledge graph in a graph database, and storing it in the form of triples representing standard clauses, related processes or components, and constraints; automatically verifying whether preconditions and acceptance standards are met upon completion of a process, comparing the detection values ​​output by the image recognition engine with the allowable deviation thresholds in the standard knowledge graph, and automatically generating a compliance acceptance form during the completion phase.

[0010] Step S5 involves executing multi-engine collaborative perception and closed-loop decision-making. The graph neural network progress prediction engine, the neural radiation field and building information model-associated image recognition engine, and the federated resource optimization engine form a bidirectional data closed loop via an event bus and message queue. In the forward control flow, the resource conflict warning signal and critical path deviation risk level output by the graph neural network progress prediction engine are sent to the image recognition engine in real time via the event bus. The image recognition engine dynamically adjusts the preset region of interest weights of the cameras accordingly, prioritizing the allocation of computing power to the building information model spatial region corresponding to high-risk processes. In the reverse feedback flow, the actual installation completion degree identified by the image recognition engine is semantically mapped to generate an actual progress tensor, which is then sent back to the progress prediction engine in real time. This serves as an external observation value for the temporal attention mechanism, driving the model to perform rolling re-prediction. In the global constraint flow, the normative knowledge graph simultaneously provides dynamic constraint benchmarks to both the graph neural network progress prediction engine and the neural radiation field and building information model-associated image recognition engine.

[0011] Step S6: Based on the bidirectional data loop, generate an analysis result set. The analysis of construction progress prediction results includes analyzing the predicted duration of the critical path, the probability distribution of duration deviations, risk warning levels, and resource conflict warning signals; the analysis of quality and safety monitoring results includes analyzing defect types, detection confidence levels, defect locations, and references to relevant standards and regulations; the aforementioned analysis results are then aggregated to form an analysis result set, which includes process information, quality defect information, and corresponding handling suggestions for different risk warning levels.

[0012] Step S7 involves feeding the analysis result set back to the graph neural network progress prediction engine and the neural radiation field and building information model-associated image recognition engine, correcting and retraining each engine model, and synchronizing the global model parameters aggregated in the cloud to the student model deployed on the edge computing device, ensuring parameter consistency between the student model and the global model. Specifically, this includes: extracting and converting the analysis result set according to the training data format of each engine model; comparing the converted data with the historical training data of each engine model; optimizing and adjusting the parameters of each engine model based on the comparison results; retraining the graph neural network progress prediction engine and the neural radiation field and building information model-associated image recognition engine using the optimized parameters; validating the retrained model using a validation set and adjusting the parameters; and evaluating the validated model using a test set, with evaluation metrics including prediction accuracy, recall, and F1 score.

[0013] Preferably, in step S2, the inter-layer propagation rule of the relational graph convolutional network is: in For nodes In the The hidden state of the layer, where R is a set of relation types. For nodes There are relation types The set of neighboring nodes, The normalization constant is For relationship In the The learnable weight matrix of the layer, with different weight matrices used for different relation types. This is the self-connection weight matrix. It is a linear rectification activation function.

[0014] Preferably, in step S2, the formula for the federal average aggregation is: in The total number of participating projects, For the first Number of local samples for each project Total number of samples for all projects , for the first The project in the The local model parameters with Gaussian noise added are uploaded in rounds.

[0015] Based on the above method, this application also provides an intelligent graph perception and control system for electromechanical construction in data centers. The system adopts a five-layer architecture from bottom to top, namely, a multi-source data acquisition layer, a data fusion and governance layer, a graph temporal collaborative perception layer, a federated knowledge management layer, and an application service layer.

[0016] The multi-source data acquisition layer is equipped with various data acquisition interfaces for accessing building information models, real-time data from IoT sensors, construction logs, video surveillance streams, drawing files, and laser point cloud data. It supports communication protocols such as Hypertext Transfer Protocol, Message Queue Telemetry Transfer Protocol, Modbus Protocol, OPC Unified Architecture Protocol, and Real-Time Streaming Protocol through a unified data acquisition adapter.

[0017] The data fusion and governance layer, based on an event-driven extraction, transformation, and loading pipeline architecture, extracts, transforms, and loads multi-source heterogeneous data. It includes an industrial basic parsing engine, a portable document format drawing optical character recognition module, a video keyframe extraction module, a specification semantic alignment module, and a data quality verification module. The specification semantic alignment module automatically identifies the specification clauses referenced in the data and maps them to corresponding nodes in the knowledge graph, based on the associations between specification clauses and processes stored in the specification knowledge graph, thus achieving automatic association between data and specifications. The loaded data is written to different storage engines according to its type.

[0018] The graph-time-coordinated perception layer comprises three parallel engines: a graph neural network progress prediction engine that uses a relational graph convolutional network and temporal attention mechanism to predict construction progress in multiple tasks and is responsible for the local training phase of federated learning, training the model on local servers of each project using historical construction data and uploading the model parameters; a neural radiation field and building information model-related image recognition engine that uses neural radiation field 3D reconstruction and an improved YOLOv11 object detection network to monitor the quality and safety of the construction site in real time; and a federated resource optimization engine that uses a multi-objective evolutionary algorithm to optimize resource allocation and receives global model parameters from the federated knowledge management layer to drive resource scheduling decisions. These three engines interact bidirectionally through an event bus and message queue, forming a positive control flow of prediction-driven perception and a negative feedback flow of perception verifying prediction.

[0019] The federated knowledge management layer is deployed on a cloud server and forms a front-end and back-end collaborative relationship with the graph neural network progress prediction engine in the graph temporal collaborative perception layer. It is responsible for receiving the model parameters uploaded by each project, using a federated averaging algorithm combined with a differential privacy mechanism for secure aggregation, and then distributing the aggregated global model to the local graph neural network progress prediction engine of each project for personalized fine-tuning, thereby realizing cross-project knowledge transfer.

[0020] The application service layer provides data query and command issuance functions to the outside world through a representational state transfer interface and a GraphQL interface. It includes an intelligent progress management module, a quality and safety monitoring module, a resource scheduling optimization module, a risk warning module, and a visualization dashboard module. The risk warning module calls the corresponding warning push channel according to the warning level, including application push, SMS push, and audible and visual alarm control.

[0021] Preferably, the data fusion and governance layer further includes a data quality verification module, which performs integrity checks, consistency checks, and outlier detection on the extracted and transformed data, and marks and filters out data that does not meet the quality standards.

[0022] Compared with the prior art, the method and system provided in this application have the following advantages: This application addresses the systemic technical problem mentioned in the background art, namely the lack of a graph-aware fusion management and control system covering the entire construction process. Starting from the construction of a unified graph structure semantic space and data closed-loop mechanism, it achieves synergistic improvements in multiple technical dimensions, specifically reflected in the following aspects: First, this application solves the problems of missing process dependency modeling and insufficient progress prediction accuracy in the prior art by constructing a process dependency graph containing four types of edges: logical dependency, resource competition, spatial conflict, and normative constraints, using a relational graph convolutional network for message passing, and embedding constraint vectors from a normative knowledge graph.

[0023] Second, this application achieves secure aggregation and personalized fine-tuning of cross-project model parameters through a federated learning framework, enabling knowledge transfer while protecting the data privacy of each project. This solves the problems of difficult cross-project knowledge transfer and long cold start cycles for new projects in existing technologies.

[0024] Third, this application generates geometrically consistent virtual training samples through three-dimensional reconstruction of neural radiation fields, and combines them with the improved YOLOv11 detection head with semantic injection of building information model, which solves the bottleneck problem of small sample recognition in professional scenarios of electromechanical construction in data centers and improves the recognition accuracy of professional scenarios such as pipeline installation deviation.

[0025] Fourth, this application achieves rapid access to new data sources within two hours and automatic compliance verification of specifications through event-driven microservice architecture and automatic semantic mapping driven by specification knowledge graph, thus solving the problem of rigid data access and governance in existing technologies.

[0026] Fifth, this application constructs a complete closed-loop control link from data acquisition to prediction, monitoring, decision-making, execution, and feedback by using the forward control flow, reverse feedback flow, and global constraint flow between the three engines, thereby realizing the active regulation of monitoring data to power distribution output and the continuous optimization of the model.

[0027] The improvements mentioned above are not independent of each other, but rather support and coordinate with each other through a unified graph-sensing fusion control system, jointly achieving a systematic technical effect of intelligent control of the entire process of data center electromechanical construction. Attached Figure Description

[0028] The specific embodiments of this application will be further described in detail below with reference to the accompanying drawings. It should be noted that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this application. Various modifications and equivalent substitutions made by those skilled in the art without departing from the spirit and scope of the technical solution of this application should fall within the protection scope of this application.

[0029] Figure 1 The schematic diagram illustrates the overall flowchart of the intelligent image perception and control method for data center electromechanical construction provided in at least one embodiment of this application; Figure 2 The flowchart illustrating the progress prediction of the graph neural network in the intelligent graph perception and control method for data center electromechanical construction provided in at least one embodiment of this application is shown in the illustration. Figure 3 This illustration schematically shows a flowchart of the image recognition and monitoring of the association between the neural radiation field and the building information model in the intelligent image perception and control method for electromechanical construction of data centers provided in at least one embodiment of this application; Figure 4This illustration schematically shows a multi-engine collaborative perception and closed-loop decision-making structure in the intelligent image perception and control method for data center electromechanical construction provided in at least one embodiment of this application; Figure 5 The illustration shows the overall architecture of the intelligent image perception and control system for data center electromechanical construction provided in at least one embodiment of this application. Detailed Implementation

[0030] The technical solutions of this application will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application. It should be noted that relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0031] This application provides an intelligent graph-based perception and control method and system for electromechanical construction in data centers. It is understood that, in this field, data center electromechanical construction encompasses more than ten professional systems, including power supply and distribution systems, HVAC systems, and low-voltage intelligent systems, involving hundreds of cross-operations and multi-disciplinary collaborations. It represents the most complex and challenging stage in data center construction. These processes not only have sequential logical dependencies but also multi-dimensional non-linear dependencies such as resource competition, spatial conflicts, and constraints from mandatory national standards. Existing technologies treat process sequences as simple time series, failing to characterize these complex dependencies, resulting in low accuracy in schedule prediction and data silos between different projects, leading to poor model generalization capabilities. This application integrates Building Information Modeling (BIM), the Internet of Things (IoT), computer vision, graph neural networks, and federated learning technologies to achieve intelligent prediction of construction progress, real-time monitoring of quality and safety, and optimized resource allocation.

[0032] The executing entity of this intelligent image perception and control method can be a computer, a smart terminal, a cloud server, an edge computing device, or other intelligent processing device with logical operation and image processing functions.

[0033] Example 1 This application provides an intelligent map-based perception and control method for electromechanical construction in data centers. Figure 1 This schematically illustrates the overall flowchart of the intelligent diagram perception and control method for data center electromechanical construction provided in at least one embodiment of this application, such as... Figure 1 As shown, the intelligent diagram perception and control method for data center electromechanical construction includes steps S1 to S7.

[0034] Step S1 involves collecting multi-source data and fusing it to obtain fused multi-source construction data. The multi-source data collection includes access to six data sources through various data acquisition interfaces: Building Information Modeling (BIM), real-time data from RFID or IoT sensors, construction logs, ultra-high-definition video surveillance streams, portable document format drawings, and laser point clouds. The acquisition layer supports fourteen communication protocols, including Hypertext Transfer Protocol, Message Queuing Telemetry Transmission Protocol, Modbus Protocol, OPC Unified Architecture Protocol, and Real-Time Streaming Protocol, through a unified data acquisition adapter. The fused multi-source construction data is then transmitted to the graph-time-series collaborative perception layer for analysis and judgment.

[0035] The fusion processing is executed by the data fusion and governance layer, including an event-driven extraction, transformation, and loading pipeline architecture for extracting, transforming, and loading multi-source heterogeneous data. This fusion processing encompasses industrial basic class parsing, optical character recognition of portable document format drawings, video keyframe extraction, and semantic alignment of the knowledge graph based on the national standard GB 50462-2024 "Construction and Acceptance Standard for Data Center Infrastructure". The loaded data is then written to different storage engines according to its type: time-series data is written to InfluxDB, relational data to PostgreSQL, graph data to Neo4j, document data to MinIO object storage, and vector data to Milvus.

[0036] The graph-time-series collaborative perception layer comprises three parallel engines: a graph neural network progress prediction engine that uses a relational graph convolutional network and temporal attention mechanism to predict construction progress in multiple tasks; a neural radiation field and building information model-related image recognition engine that uses neural radiation field 3D reconstruction and YOLOv11 target detection network to monitor construction site quality and safety in real time; and a federated resource optimization engine that uses federated learning and multi-objective evolutionary algorithms to achieve cross-project knowledge transfer and resource optimization. The federated knowledge management layer, deployed on a cloud server, receives model parameters trained locally by each project, uses a federated averaging algorithm combined with differential privacy mechanisms for secure aggregation, and then distributes the aggregated global model to each project for personalized fine-tuning. The application service layer provides five functional modules: intelligent progress control, quality and safety monitoring, resource scheduling optimization, risk warning, and visualization dashboards. These three engines interact bidirectionally via an event bus and message queue, forming a positive control flow of prediction-driven perception and a negative feedback flow of perception verifying prediction, achieving system-level collaborative perception and closed-loop decision-making.

[0037] Step S2: Use a graph neural network progress prediction engine to predict the construction progress. Figure 2 This illustration schematically shows a flowchart of the graph neural network progress prediction in the intelligent graph perception and control method for data center electromechanical construction provided in at least one embodiment of this application; as shown... Figure 2 As shown, the graph neural network progress prediction specifically includes steps S21 to S27.

[0038] Step S21: Construct a process dependency graph. Obtain the process list for the electromechanical construction project of the target data center, and construct a directed heterogeneous graph with each process as a node. Each node in the node set V contains a 64-dimensional attribute vector, specifically including the work breakdown structure code, resource requirement type and quantity, planned and completed duration, building information model spatial coordinates, professional type (including electrical, HVAC, low-voltage electrical, fire protection, and more than ten others), a list of preceding processes, and resource input intensity. The edge set E contains four types of edges: logical dependency edges representing the sequential relationship between processes, resource competition edges representing two processes sharing the same type of finite resource, spatial conflict edges established based on geometric collision detection in the building information model, and normative constraint edges established based on the mandatory clauses of the national standard GB 50462-2024 "Construction and Acceptance Standard for Data Center Infrastructure". Node type T is used to distinguish different professional fields.

[0039] The specific construction methods for the four types of edges are as follows. The construction method for logical dependency edges is: based on the preceding process field in the process list, if the preceding process of process j includes process i, then construct a directed logical dependency edge from node i to node j. The construction method for resource contention edges is: traverse all process pairs... If process i and process j share the same type of finite resource and their resource demand time windows overlap, then an undirected resource competition edge is constructed, with the edge weight... Calculate using the following formula:

[0040] in The overlap duration of the resource demand time windows for process i and process j. For the shorter of the two processes, The required quantity of resources shared between the two processes. This represents the total amount of this type of resource. It is a balance coefficient between time overlap and resource scarcity.

[0041] The method for constructing spatial conflict edges is as follows: based on geometric collision detection of the building information model, bounding box intersection detection is performed on each pair of BIM components associated with process i and process j. If the bounding boxes intersect in three-dimensional space and the construction time windows overlap, then a spatial conflict edge is constructed. The collision tolerance threshold is preset according to the accuracy level of the building information model components. That is, when the minimum distance between the bounding boxes of two components is less than the threshold, it is determined to be a spatial conflict.

[0042] The method for constructing normative constraint edges is as follows: traverse the mandatory clause nodes in the normative knowledge graph. For each mandatory clause, query its associated process nodes. Only construct normative constraint edges for clauses with a violation risk level of medium risk or above. The violation risk level is determined based on the punitive expressions in the clause. Clauses containing expressions such as "must" or "strictly prohibited" are high-risk, clauses containing expressions such as "should" or "should not" are medium-risk, and clauses containing expressions such as "appropriate" or "inappropriate" are low-risk. Only construct normative constraint edges corresponding to high-risk and medium-risk clauses.

[0043] Step S22, embedding the standard knowledge graph. A standard knowledge graph is constructed based on the national standard GB 50462-2024 "Construction and Acceptance Standard for Data Center Infrastructure". This knowledge graph is a graphical knowledge representation structure stored in the form of triples of standard clauses, related procedures, and constraints, used to express the semantic relationships between standard clauses and construction procedures. The TransE model is used to embed the entities and relations in the standard knowledge graph into a distributed representation in a low-dimensional vector space. The training objective of the TransE model is to minimize the following loss function:

[0044] in For the set of positive sample triples, For the set of negative sample triples, For the interval hyperparameter, Use L1 or L2 distance metric function. After training, for each process node, concatenate its associated canonical constraint vector into the attribute vector of the corresponding process node to form an enhanced node representation that integrates canonical semantics.

[0045] Step S23, message passing in the relational graph convolutional network. Multiple rounds of message passing are performed based on the process dependency graph and the enhanced node representation. The inter-layer propagation rule of the relational graph convolutional network is as follows:

[0046] in For node i at the th The hidden state of the layer The set of relation types includes four categories: logical dependency, resource contention, space conflict, and normative constraint. For the relationship type between node i and node i The set of neighboring nodes, The normalization constant is For relation r in the th case The learnable weight matrix of the layer, with different weight matrices used for different relation types. This is the self-connection weight matrix. It is a linear rectified activation function. After L-layer message passing, a deep representation of the multi-hop neighbor information fused from each process node is obtained.

[0047] Step S24, Temporal Attention Calculation. Based on the node representation output by the relational graph convolutional network, a Transformer-style multi-head temporal attention mechanism is introduced. For each process node i, its relational graph convolutional network output sequence within the historical time window is used as input, and the temporal attention weights are calculated according to the following formula:

[0048] in This is the query vector for the current process status. and These are the key vectors and value vectors of the process sequence within a historical time window. As the key vector dimension, this embodiment sets the number of multi-heads to eight, achieving a balance between accuracy and computational efficiency.

[0049] Step S25, Multi-task Joint Prediction. Based on the temporal attention output, multiple task-specific prediction heads are connected to achieve a four-task joint output. Critical path duration prediction is a regression task, outputting the estimated total duration. Duration deviation probability distribution is a classification task, divided into three levels: deviation less than 2.5%, deviation between 2.5% and 5%, and deviation greater than 5%. Risk warning level assessment is a three-class classification task, divided into high risk, medium risk, and low risk levels. Resource conflict warning is an edge-level prediction task, outputting the conflict probability for each resource-competing edge; a warning is triggered when the probability is greater than 0.7. The total loss function for multi-task joint training is:

[0050] in, This represents the total loss value. For regression loss, the mean squared error loss function is used to measure the deviation between the predicted and actual project durations on the critical path. For classification loss, the cross-entropy loss function is used to measure the difference between the predicted result of the project schedule deviation probability distribution and the actual deviation level. For risk loss, a binary cross-entropy loss function is used to measure the difference between the predicted risk warning level and the actual risk level. For conflict loss, a binary cross-entropy loss function is used to measure the difference between the predicted resource conflict edge and the actual conflict state. These are the weight coefficients for the corresponding loss terms, used to balance the contribution of each task in the joint training. The value of is determined through a grid search, and the sum of the four weights equals 1. The specific method of the grid search is as follows: ... (The sentence is incomplete and requires more context to translate accurately). Under the constraints, a grid traversal is performed on the four weight coefficients. For each weight combination, the accuracy index of each task is calculated on the validation set. The normalized weighted sum of the mean absolute error of critical path duration prediction, duration deviation classification accuracy, risk warning F1 score, and resource conflict prediction accuracy is used as the comprehensive evaluation index. This optimizes the comprehensive index and minimizes the normalized variance of each task's index, avoiding an imbalance where one task has excessively high accuracy while others have excessively low accuracy. Preferably, an uncertainty weighting method is used for dynamic weight adjustment. In the early stages of training, the weights are automatically adjusted based on the uncertainty of the loss of each task. Once the loss stabilizes, the optimal fixed weights determined by the grid search are switched to the optimal weights.

[0051] Step S26, Federated Learning Cross-Project Training. Each participating project trains a relational graph convolutional network model on its local server using its project's construction history data; this is the local training phase. Each project then submits the locally trained model parameters... Uploaded to the central server of the federated knowledge management system. The central server first adds parameters to the local models uploaded by each project to meet the requirements. - Differential privacy Gaussian noise, assuming the parameters after adding noise are... ,in With a mean of 0 and a noise scale of Gaussian noise, Value selection and privacy budget Relatedly, the differential privacy budget allocation mechanism is as follows: Let the total privacy budget be... The training rounds are T. The total budget is evenly distributed across each round using the combinatorial theorem, and the privacy budget for each round is... Noise scale With each round of privacy budget and the probability of target failure The functional relationship is The privacy budget tracking module records the privacy budget consumed in each round. When the cumulative consumption reaches a preset percentage of the total budget, it stops adding noise and outputs a prompt. Then, a federated averaging algorithm is used to adjust the parameters after adding noise. Aggregate according to the following formula:

[0052] in The total number of participating projects, Let be the number of local samples for the k-th project. The total number of samples for all projects. The parameters of the local model with Gaussian noise added for the k-th project uploaded in round t+1. The aggregated global model. The data is then distributed to each project site, where local data is used for five to ten rounds of personalized fine-tuning. This is the personalized fine-tuning stage, which aims to ensure model convergence while avoiding overfitting.

[0053] Step S27, Lightweight Deployment. The complex, trained teacher model is compressed into a student model using knowledge distillation and model pruning techniques. The specific process of knowledge distillation is as follows: the teacher model and student model undergo simultaneous forward propagation, and the distillation loss function for the student model is:

[0054] in For temperature parameters, For the teacher model at temperature The softened output probability distribution is as follows. For student models at temperature The softened output probability distribution is as follows. Let KL divergence be the KL divergence. Output the student model with the real labels. Cross-entropy loss between The weights represent the distillation terms. The teacher model is a complete relational graph convolutional network, while the student model is a lightweight model with reduced hidden layer dimensions and some attention heads removed. Model pruning employs structured channel pruning, which prunes channels with weight norms below a threshold after sparsifying the output channels of the convolutional layers during training, adapting to edge computing device deployment.

[0055] The ratio of parameters between the teacher model and the student model is based on general experience in the field of knowledge distillation, achieving significant parameter compression while ensuring the expressive power of the student model.

[0056] Step S3: Use a neural radiation field and building information model-linked image recognition engine to perform real-time quality and safety monitoring at the construction site. Figure 3 This illustration schematically shows a flowchart of image recognition and monitoring of the association between neural radiation field and building information model in the intelligent image perception and control method for data center electromechanical construction provided in at least one embodiment of this application; as shown Figure 3 As shown, the image recognition and monitoring of the neural radiation field and the building information model specifically includes steps S31 to S36.

[0057] Step S31: Multi-view image acquisition and 3D reconstruction of the neural radiation field. Multi-view images of the construction site are acquired using ultra-high-definition network cameras deployed at key work surfaces. These multi-view images are used as input to the neural radiation field. Neural radiation field network. Using three-dimensional spatial point coordinates and observation direction Given the input, output the volume density at that point. and color :

[0058] in, This represents a neural network mapping implemented using a multilayer perceptron. The set of learnable parameters of the network. The volume density at a point in space. For a spatial point in the direction of observation The color presented below. Volume density. It is a perspective-independent quantity, color. It is a viewpoint-dependent quantity, and this design enables neural radiation fields to model complex optical effects such as specular highlights and reflections.

[0059] The neural radiation field is trained by randomly sampling rays from multi-view images with known camera poses, integrating the 3D points on each ray to render pixel colors, and using the mean square error between the rendered color and the true pixel color as the loss function for optimization. Specifically, a set of rays is randomly sampled from the training images. For each ray of light By integrating the volume density and color of each point output by the neural radiation field network along the direction of the ray, the predicted color of the pixel corresponding to that ray is obtained. The predicted color is compared with the color of the pixel in the real image. The mean square error between them is used as the optimization objective:

[0060] in This represents the squared L2 norm, which is the sum of the squared differences between the components of the color vector. The network parameters are iteratively updated using the backpropagation algorithm by minimizing this loss function. This enables neural radiation fields to render high-fidelity scene images from any virtual perspective.

[0061] Step S32: Building Information Model Registration and Spatial Alignment. Corresponding control points are selected in both the 3D scene reconstructed from the neural radiation field and the building information model. The iterative nearest-point algorithm is used to solve the rigid body transformation matrix, minimizing the registration error between the two sets of point clouds.

[0062] Step S33: Virtual training sample generation. Virtual training samples are generated in the 3D scene reconstructed from the neural radiation field through viewpoint transformation, lighting perturbation, and occlusion simulation. Viewpoint transformation includes thirty-six horizontal viewpoints, lighting perturbation includes five lighting conditions, and occlusion simulation includes three occlusion levels. Five hundred and forty virtual samples are generated from each real image. These virtual samples carry spatial semantic labels from the Building Information Model (BIM) and possess geometric consistency. The design of generating five hundred and forty virtual samples from each real image is based on the fact that thirty-six horizontal viewpoints multiplied by five lighting conditions multiplied by three occlusion levels equals five hundred and forty combinations, ensuring sufficient sample diversity while controlling the computational cost of data generation and training.

[0063] Step S34: Training the improved YOLOv11 network. A YOLOv11 object detection network is constructed. Its backbone uses the C3k2 mechanism for multi-scale feature extraction and embeds a pyramid spatial attention module. The neck layer uses a bidirectional feature pyramid network. In addition to P3 and P4, a P2 detection head is added for small-sized object detection, and a 128-dimensional building information model semantic vector is injected into the detection head. Training is performed using a mixture of real samples and virtual samples from the neural radiation field.

[0064] The specific composition of the 128-dimensional building information model semantic vector includes: a 32-dimensional component type unique thermal code, used to represent component categories such as pipes, supports, and equipment; a 32-dimensional spatial coordinate normalization vector, containing the x, y, and z coordinates of the component center point in the BIM coordinate system and its bounding box size, with each dimension normalized to the [0, 1] interval; a 32-dimensional specification requirement value vector, containing specification constraint values ​​such as allowable deviation of horizontality, allowable deviation of verticality, and insulation layer thickness requirements; and a 32-dimensional process status vector, containing the construction stage code of the current process and the deviation between the estimated completion time and the actual completion time.

[0065] The fusion method between the BIM semantic vector and image features is feature concatenation and gating fusion. Specifically, the image feature vector output by the YOLOv11 detection head is... With BIM semantic vectors The features are concatenated to obtain a 384-dimensional fused feature vector. Attention weights are calculated via a gating mechanism:

[0066] in The weight matrix is ​​a learnable matrix. The activation function is Sigmoid. The fused features are:

[0067] in For element-wise multiplication, This is a linear transformation matrix that projects BIM semantic vectors onto the image feature dimension. Feature fusion. After being reduced to 256 dimensions by a 1×1 convolutional layer, the results are fed into the subsequent classification and regression branches to output the detection results.

[0068] Step S35, Real-time Inference via Edge Computing. The trained YOLOv11 model is deployed on an edge computing device. Adaptive cropping is performed based on the preset region of interest in the Building Information Model. After decoding the real-time streaming protocol video stream, the TensorRT engine is used for half-precision floating-point quantization and eight-bit integer calibration inference, and the detection results are output.

[0069] Step S36: Graded early warning system linked with Building Information Model (BIM). Graded early warnings are issued based on the detection results and semantic information from the BIM model. A normal state is indicated by a green indicator; the system continues monitoring without triggering an early warning. General alarms are indicated by a yellow indicator; the system logs the alarm and pushes the alarm information to the management application. The alarm information includes the BIM model's location coordinates and a 3D screenshot of the neural radiation field. Emergency early warnings are indicated by a red indicator; the system immediately triggers an audible and visual alarm, automatically saves 30 seconds of video evidence before and after the alarm, highlights the defect location in the BIM model, and simultaneously pushes the warning information to on-site management personnel and safety officers. The warning information includes the defect type, detection confidence level, BIM component ID, spatial coordinates, reference to relevant regulations, a 3D screenshot of the neural radiation field, and suggested remedial measures.

[0070] Step S4 involves performing real-time compliance verification on the fused multi-source construction data based on the aforementioned standard knowledge graph. The standard knowledge graph-driven compliance verification specifically includes steps S41 to S42.

[0071] Step S41: Standardize the storage of the knowledge graph. The standardized knowledge graph is stored using the Neo4j graph database. The knowledge graph is stored in the form of triples of standard clauses, related processes or components, and constraints, including mandatory clause nodes, process or component nodes, and constraint nodes. It supports standard clause retrieval, related process query, and compliance reasoning based on the Cypher query language.

[0072] Step S42: Automatic compliance verification. Real-time compliance verification is performed on the fused multi-source construction data based on the aforementioned standard knowledge graph. Upon completion of a process, the system automatically verifies whether preconditions and acceptance standards are met. The detection values ​​output by the image recognition engine are compared with the allowable deviation thresholds in the standard knowledge graph. A compliance acceptance form is automatically generated at the completion stage.

[0073] Step S5: Multi-engine collaborative perception and closed-loop decision-making. Figure 4The flowchart illustrating the multi-engine collaborative perception and closed-loop decision-making process in the intelligent image perception and control method for data center electromechanical construction provided in at least one embodiment of this application is shown; Figure 4 As shown, the multi-engine collaborative perception and closed-loop decision-making includes a forward control flow, a reverse feedback flow, and a global constraint flow. The graph neural network progress prediction engine, the neural radiation field and building information model (BIM) image recognition engine, and the federated resource optimization engine form a bidirectional data closed loop through an event bus and message queue. In the forward control flow, the resource conflict warning signal and critical path deviation risk level output by the graph neural network progress prediction engine are sent in real-time to the neural radiation field and BIM image recognition engine via the event bus. Based on this, the image recognition engine dynamically adjusts the preset region of interest weights of the cameras, prioritizing the allocation of computing power to the BIM spatial region corresponding to high-risk processes. In the reverse feedback flow, the actual installation completion degree identified by the neural radiation field and BIM image recognition engine is semantically mapped to generate an actual progress tensor, which is then sent back to the graph neural network progress prediction engine in real-time. This serves as an external observation for the temporal attention mechanism, driving the model to perform rolling re-prediction. In the global constraint flow, the specification knowledge graph simultaneously provides dynamic constraint benchmarks to both the graph neural network progress prediction engine and the neural radiation field and BIM image recognition engine.

[0074] In the forward control flow, the resource conflict warning signal and critical path deviation risk level output by the graph neural network progress prediction engine are sent to the image recognition engine in real time via the event bus. The specific mapping rule is as follows: the progress prediction engine outputs a list of conflicting processes. ,in A unique identifier for BIM components in conflicting work processes. For the conflict time window, The severity of the conflict is scored, with a range of values. After receiving the list of conflicting work processes, the image recognition engine queries the BIM model to obtain the spatial areas corresponding to each conflicting process. The sampling weight of the region of interest for the camera within this spatial area is increased from the default value of 1.0 to... ,in This is a weighting enhancement coefficient that increases the sampling frequency in high-risk areas.

[0075] In the reverse feedback stream, the actual installation completion rate identified by the image recognition engine is semantically mapped to generate an actual progress tensor. The specific mapping rule is as follows: the image recognition engine outputs the detection completion rate of each BIM component. The actual completion rate of process j is obtained by taking the weighted average of the completion rates of all components associated with process j. ,in Let j be the set of components associated with process j. Let i be the process schedule weight. The actual completion rate of each process will be... Completion rate of the plan The difference Arranged as schedule deviation tensors The data is fed back to the progress prediction engine in real time. The external observations used as the time-series attention mechanism are concatenated with the hidden states of the process nodes and then input into the time-series attention module for rolling re-prediction.

[0076] The types of events that the three engines communicate with each other via the event bus include: resource conflict warning events, quality defect detection events, model parameter synchronization events, and specification constraint update events. Each type of event includes an event identifier, timestamp, data payload, and priority fields. Priorities are divided into four levels: urgent, high, medium, and low, with urgent events enjoying the highest transmission priority in the event bus.

[0077] Step S6: Analyze the construction progress prediction results and quality and safety monitoring results to obtain an analysis result set. The analysis of the construction progress prediction results includes analyzing the predicted duration of the critical path, the probability distribution of duration deviations, risk warning levels, and resource conflict warning signals. The analysis of the quality and safety monitoring results includes analyzing defect types, detection confidence levels, defect locations, and references to relevant specifications. The aforementioned analysis results are then aggregated to form an analysis result set. This analysis result set includes process information, quality defect information, and corresponding handling suggestions for different risk warning levels.

[0078] Step S7 involves feeding the analysis result set back into the graph neural network progress prediction engine and the neural radiation field-building information model (BIM) image recognition engine to correct, adjust, and retrain each engine model. Specifically, feeding back the analysis result set includes: extracting and converting the analysis result set according to the training data format of each engine model; comparing the converted data with the historical training data of each engine model; optimizing and adjusting the parameters of each engine model based on the comparison results; retraining the graph neural network progress prediction engine and the neural radiation field-building information model (BIM) image recognition engine using the optimized parameters; validating the retrained model using a validation set and adjusting the parameters; and evaluating the validated model using a test set, with evaluation metrics including prediction accuracy, recall, and F1 score.

[0079] Based on the above method, this application also provides an intelligent map-based perception and control system for electromechanical construction in data centers. Figure 5 This illustration shows the overall architecture of a data center electromechanical construction intelligent diagram perception and control system provided in at least one embodiment of this application. For example... Figure 5As shown, the system adopts a five-layer architecture from bottom to top, namely, a multi-source data acquisition layer, a data fusion and governance layer, a graph time-series collaborative perception layer, a federated knowledge management layer, and an application service layer. The working process of each module is described in detail below.

[0080] The multi-source data acquisition layer serves as the interface between the system and external data sources, and its operating principle is as follows: Internally, the acquisition layer maintains a unified pool of data acquisition adapters. This pool pre-configures implementations of fourteen communication protocols, including Hypertext Transfer Protocol (HTTP) adapters, Message Queuing Telemetry Transport Protocol (MQTP) adapters, Modbus protocol adapters, OPC Unified Architecture Protocol (OPC) adapters, and Real-Time Streaming Protocol (RTS) adapters. Upon system startup, the acquisition layer reads a pre-configured data source registry, which records the connection parameters, protocol type, acquisition frequency, and data structure definition for each data source. The acquisition layer initiates connection requests to various data sources in parallel according to the configured acquisition frequency. For Building Information Modeling (BIM) data sources, the acquisition layer reads model files through the industrial basic class file interface and parses the geometric and attribute information of components. For RFID or IoT sensors, the acquisition layer establishes a bidirectional communication channel through the Modbus protocol or the OPC unified architecture protocol, listens to the data frames reported by the sensors in real time, and performs protocol decoding. For ultra-high-definition video surveillance streams, the acquisition layer pulls the video stream through the real-time streaming protocol and caches it in a circular buffer for subsequent processing. For portable document format drawings, the acquisition layer detects the addition or modification events of drawing files through the file system listening interface. For laser point cloud data, the acquisition layer receives the 3D point cloud files output by the scanner through the point cloud data protocol interface. The acquisition layer encapsulates raw data from different sources and in different formats into a message containing a data source identifier, acquisition timestamp, data format tag, and original content, and pushes it to the event bus, thereby decoupling the acquisition layer from downstream data fusion and governance layers.

[0081] The data fusion and governance layer processes the raw data pushed by the acquisition layer through an event-driven extraction, transformation, and loading pipeline. This layer contains an event listener that continuously monitors raw data messages on the event bus. When a new message arrives, the event listener routes the message to the corresponding processing engine based on the data source identifier and format tag in the message header. For Building Information Modeling (BIM) data in the industrial basic class format, the industrial basic class parsing engine extracts the geometric information and spatial coordinates of the components. Specifically, it parses the exchange structure of the industrial basic class file, traverses each building component entity, reads its spatial location attributes, extracts its geometric shape representation, and determines its professional category and functional classification based on the entity type. Finally, it converts the extracted results into a unified spatial entity format within the system. For portable document format drawing data, the optical character recognition (OCR) engine first preprocesses the drawing image, including grayscale conversion, binarization, and tilt correction. Then, it uses an OCR algorithm to recognize text annotations, dimension numbers, and specification clause numbers in the drawing and stores the recognition results in a structured manner. For video stream data, the keyframe extraction module extracts image frames from the video stream at preset time intervals and inputs the image frames into the neural radiation field depth map generation module. The specification semantic alignment module receives the output results from the aforementioned engines. Based on the associations between specification clauses and procedures stored in the specification knowledge graph, it automatically identifies the specification clauses referenced in the data and maps them to corresponding nodes in the knowledge graph, thus achieving automatic association between data and specifications. The governance layer also includes a data quality verification module, which performs integrity checks, consistency checks, and outlier detection on the extracted and transformed data, marking and filtering data that does not meet quality standards. After fusion and governance processing, the data is written to the corresponding storage engine according to its type.

[0082] The graph-time-series collaborative perception layer is the core computing layer of the system, comprising three engines that run in parallel and exchange information in real time via an event bus. The graph neural network progress prediction engine is responsible for the local training phase of federated learning, training the model on the local servers of each project using historical construction data and uploading the model parameters to the federated knowledge management layer; the federated resource optimization engine receives global model parameters issued by the federated knowledge management layer to drive resource scheduling decisions.

[0083] The internal operating principle of the graph neural network progress prediction engine is as follows. Upon engine startup, it first loads the process dependency graph data for the current project from the graph database, including the node set, edge set, and node attribute vectors. The process dependency graph construction module organizes the loaded data into a tensor format that can be processed by a relational graph convolutional network. Specifically, it normalizes the 64-dimensional attribute vector of each process node, constructs sparse tensors from the adjacency matrices of the four types of edges, and encodes the node type as a one-hot vector. The specification knowledge graph embedding module queries the specification clauses associated with each process from the specification knowledge graph, encodes the specification clauses into vector representations using the TransE model, and concatenates them with the attribute vectors of the corresponding process nodes to form an enhanced node feature matrix. The relational graph convolutional network propagation module inputs the enhanced node feature matrix and four types of adjacency tensors into the relational graph convolutional network. Following the aforementioned inter-layer propagation rules, it performs three layers of convolutional operations. In each layer, the four types of edges use independent weight matrices for message aggregation. The aggregation results of the four types of messages are then summed and passed through a linear rectified activation function to output the node representation for that layer. After the three-layer operation, a deep node representation incorporating multi-hop neighbor information is output. The temporal attention module reads historical progress data for each process from the temporal database, aligns it with the deep node representation output by the relational graph convolutional network in the time dimension, constructs the state sequence of each node within the historical time window, and then inputs it into a multi-head temporal attention mechanism to calculate the attention weights for each historical moment, outputting a weighted aggregated context vector. The multi-task prediction output module inputs the context vector into four task heads respectively. The critical path duration prediction head outputs the duration prediction value through a fully connected layer regression; the duration deviation classification head outputs a three-level deviation probability distribution through a Softmax classifier; the risk warning head outputs a high-risk probability value through a binary classifier; and the resource conflict prediction head outputs the conflict probability for each resource competition edge. All of the above calculations are performed in parallel on the graphics processor, achieving millisecond-level forward inference.

[0084] The internal operating principle of the neural radiation field and building information model (BIM) associated image recognition engine is as follows: Upon engine startup, pre-trained neural radiation field network weights and YOLOv11 network weights are loaded first. The multi-view image acquisition module inputs the acquired multi-view images and corresponding camera pose parameters into the neural radiation field network. The network outputs an implicit representation of the 3D scene using voxel rendering technology. The BIM registration module extracts the spatial coordinates of corresponding control points from the BIM model and performs iterative nearest-point registration with the coordinates of corresponding points in the neural radiation field reconstructed scene, solving for the rigid body transformation matrix to achieve scene alignment. The virtual sample generation module batch-renders virtual images from multiple perspectives in the aligned scene by changing the virtual camera pose and lighting parameters. Simultaneously, it queries the BIM model for the category labels and spatial coordinates of visible components from the corresponding perspectives, automatically generating labeled data. The training process of the YOLOv11 network includes four steps: forward propagation, loss calculation, backpropagation, and weight update. The network adopts a multi-scale feature extraction structure, using detection heads with different receptive fields for large, medium, and small-scale targets. In the detection head stage, the semantic vectors of the Building Information Model (BIM) are fused with image features through a concatenated feature layer, and the fused result is output after convolutional layer fusion. In the inference stage, the engine acquires real-time image frames from the video stream, crops them after the region of interest (ROI), and inputs them into the YOLOv11 network to perform one forward propagation, outputting the bounding box coordinates, class confidence, and normative bias.

[0085] The internal operating principle of the federated resource optimization engine is as follows. The engine subscribes to risk warning signals output by the progress prediction engine and quality inspection results output by the image recognition engine from the event bus. When a new warning signal is received, the engine initiates a multi-objective optimization process. First, it reads the currently available resource list from the resource database, including equipment type, quantity, current occupancy status, and future availability time window. Then, based on the conflicting work order list and conflicting time window contained in the warning signal, it constructs a multi-objective optimization model for resource allocation.

[0086] The mathematical description of the multi-objective optimization model for resource allocation is as follows. The decision variables are... This indicates whether process j is allocated for execution within time window t. , The constraints include: resource capacity constraints. That is, within each time window t, the total demand of all executing processes for resource type k does not exceed the maximum supply of that resource type; process logic constraints. That is, the prerequisite for process j to be executed in time window t is that its predecessor process i has been completed in a previous time window; time window constraint That is, process j can only be executed within the allowed time window.

[0087] The specific mathematical expressions for the three optimization objectives are: minimizing the construction period. That is, minimizing the end time of the last completed process and maximizing resource utilization. This means maximizing the ratio of total resource usage to total resource supply across all time windows, and minimizing costs. This means minimizing the total execution cost of all processes. Let j be the resource type k required by process j. Let $\frac{j}{ ...

[0088] The engine uses the NSGA-III multi-objective evolutionary algorithm to solve for the Pareto optimal solution set. The algorithm's encoding scheme is as follows: each individual is represented as an N× A two-dimensional binary matrix, matrix elements The value can be 0 or 1. The crossover operation uses a two-point crossover, swapping the process allocation states of the same time window in two parent individuals on a process-by-process basis. The mutation operation randomly flips the element values ​​in the matrix. Each generation outputs a non-dominated solution set, selecting the solution with the highest comprehensive score as the recommended solution, and outputting it to the application service layer via the event bus.

[0089] The federated knowledge management layer is deployed on a cloud server, forming a front-end and back-end collaborative relationship with the graph neural network progress prediction engine in the graph temporal collaborative perception layer. Its operating principle is as follows: The management layer maintains a global model parameter library, which stores the global relational graph convolutional network model parameters after secure aggregation. During federated learning training, the management layer first issues the initial global model and training instructions to each participating project. After each project completes a specified number of training rounds locally, it encrypts and uploads the model parameters to the management layer. Upon receiving the uploaded parameters from each project, the management layer's secure aggregation module executes the following aggregation process: it performs format validation and outlier filtering on the model parameters of each project, excluding parameter updates with significant deviations; it calculates a weighted average of the filtered parameters based on the sample size of each project; it adds Gaussian noise that meets the differential privacy definition to the weighted average result, ensuring that the aggregated parameters cannot be used to reverse-engineer the original data of any single project; and it stores the noise-added parameters as a new global model. The management layer also maintains a model version management module, recording the timestamp of each aggregation round, the list of participating projects, and the performance metrics of the aggregated model. When a new project joins the federated learning system, the management's project registration module assigns an identity and key to the new project and distributes the latest global model to the new project's local machine, enabling the new project to use the knowledge of historical projects for a cold start.

[0090] The parameter synchronization mechanism between the student model deployed on the edge computing device and the global model in the cloud is as follows: After each round of federated average aggregation, the federated knowledge management layer generates a global model version number and records the parameter hash value, sending a version update event to each project's edge computing device via the event bus. Upon receiving the version update event, the edge computing device sends a parameter difference request to the federated knowledge management layer within its idle time window after the completion of the current inference task. The federated knowledge management layer calculates the parameter difference between the latest global model and the current student model on the edge device, transmitting only the difference to reduce the amount of data transmitted over the network. After receiving the parameter difference, the edge computing device uses the latest global model as the teacher model and local inference data as the training data to perform incremental knowledge distillation to update the student model parameters. After the update is complete, it sends a confirmation message back to the federated knowledge management layer and records the new version number. If the edge device is under high load, synchronization is delayed until the off-peak hours at night. This mechanism ensures that the lightweight model deployed at the edge can be synchronized promptly after the global model in the cloud is updated, while avoiding the impact of synchronization operations on real-time inference performance.

[0091] The application service layer provides data query and command issuance functions through expressive state passing and GraphQL interfaces. The intelligent progress management module extracts the critical path duration prediction and deviation probability distribution from the output of the graph neural network progress prediction engine, displaying them on a visualization dashboard in the form of Gantt charts and trend curves. The quality and safety monitoring module extracts defect types, detection confidence levels, and defect locations from the output of the neural radiation field and building information model (BIM) image recognition engine, overlaying defect markers onto the BIM view. The resource scheduling optimization module receives recommended scheduling schemes from the federated resource optimization engine, displaying resource allocation suggestions in the form of tables and time-series diagrams. The risk warning module listens for warning signals on the event bus and calls the corresponding warning push channels based on the warning level, including application push, SMS push, and audible and visual alarm control. The visualization dashboard module integrates the BIM 3D rendering engine, video stream decoding player, and chart rendering components, presenting various data uniformly on the user interface.

[0092] The aforementioned layers communicate asynchronously and decoupled through an event bus and message queue. Forward data flow is passed upwards from the acquisition layer to the application layer, while reverse control flow is passed from the application layer to the acquisition and execution layers via an instruction distribution channel, forming a closed-loop control system. The data acquisition in step S1 (acquisition layer), the prediction and identification in steps S2 to S5 (perception layer), the generation of the analysis result set in step S6, and the feedback retraining in step S7 are respectively completed collaboratively at the system level by the multi-source data acquisition layer, the graph-time-series collaborative perception layer, and the application service layer.

[0093] All modules of the system are deployed independently using a microservice architecture. Modules communicate synchronously via a descriptive state passing interface and asynchronously via a Kafka message queue. There are no shared memory or file system dependencies between modules, ensuring that a failure in a single module will not cause the entire system to crash. Furthermore, each module can be horizontally scaled independently based on computational load. Upon system startup, each microservice reports its address and health status to the service registry. The API gateway obtains the actual access addresses of each service through a service discovery mechanism, enabling dynamic routing and load balancing between services.

[0094] The intelligent image perception and control method and system for data center electromechanical construction disclosed in this application feeds back construction progress prediction results and quality and safety monitoring results to each engine model for correction, adjustment, and retraining. This feedback creates a closed loop between the entire control method and the training of each neural network model. Compared with existing technologies where each module works independently and operates in an open-loop manner, the continuous autonomous learning of each engine model, coupled with the feedback and correction, not only facilitates the continuous optimization and correction of each neural network model but also makes the construction progress prediction and quality and safety monitoring results more accurate and closer to the actual construction site.

[0095] Example 2 This embodiment uses a UPS installation project in a Class A data center as an example to describe the method of this application in detail. The steps in this embodiment are basically the same as those in Embodiment 1, except that the specific implementation object of steps S1 to S7 is a UPS installation project. In this embodiment, steps S1 to S7 respectively perform the following operations.

[0096] The data center project requires the installation of twenty 500 kVA UPS units, with a planned construction period of thirty days. Construction faces uncertainties such as equipment delivery delays and cross-disciplinary work with other disciplines.

[0097] In step S1, the intelligent image perception and control system of this application is deployed at the project site. The multi-source data acquisition layer accesses the following data sources: importing a building information model in IFC format from the BIM platform, including the three-dimensional geometric information and component attributes of the UPS installation area; deploying six ultra-high-definition network cameras in the construction area, covering five key work surfaces: the foundation acceptance area, equipment handling passage, installation area, cable laying area, and commissioning area; accessing the project management system to obtain construction logs and resource input records; installing IoT sensors on the hoisting equipment to collect the equipment usage status in real time; and importing construction drawings and design change orders in portable document format. The data fusion and governance layer extracts, transforms, and loads the above multi-source data. The industrial basic parsing engine extracts component information related to UPS installation, and the standard semantic alignment module automatically identifies mandatory clauses related to UPS installation based on the standard knowledge graph.

[0098] In step S21, the process dependency graph construction module of the graph neural network progress prediction engine decomposes the UPS installation into five processes: basic acceptance, equipment handling, installation in place, cable laying, and commissioning and acceptance. A directed heterogeneous graph is constructed using these five processes as nodes. Each node's 64-dimensional attribute vector includes the work breakdown structure code, resource requirement type and quantity, planned duration, building information model spatial coordinates, and the type of profession it belongs to. The edge set contains four types of edges: four logical dependency edges, two resource competition edges, three spatial conflict edges, and one normative constraint edge.

[0099] In step S22, the specification knowledge graph embedding module takes a mandatory clause as an example. The clause is used as the head entity, the battery installation is used as the tail entity, and the internal resistance test must be completed as the relation. This forms a triplet input TransE model, and the specification constraint vector is concatenated into the attribute vector of the in-place installation process node.

[0100] In step S23, the relational graph convolutional network propagation module performs three layers of message passing on the constructed process dependency graph. In the first layer of propagation, each node aggregates the feature information of its direct neighbor nodes. In the second layer of propagation, the information spreads to two-hop neighbors. After the third layer of propagation, each process node obtains a deep representation that integrates all relevant information.

[0101] In step S24, when the timing attention module makes a prediction on the 15th day, it finds that the equipment delivery delay information of the previous three days has the highest attention weight on the current schedule prediction, indicating that the delay event is a key factor affecting the current progress.

[0102] In step S25, the multi-task prediction output module outputs the following prediction results on the 15th day: the critical path duration prediction value is 35 days, the probability of duration deviation being greater than 5% is 68%, the risk warning level is high risk, and the resource conflict warning detects that there is resource competition for hoisting equipment on the 18th day and the probability of conflict is 0.82%.

[0103] In step S26, each of the three historical data center projects contains over 100,000 construction logs. Each project trains a relational graph convolutional network model on its local server for ten rounds and then uploads the model parameters. The central server uses a federated averaging algorithm to aggregate the parameters. Before aggregation, Gaussian noise that satisfies differential privacy is added. The aggregated global model is then distributed to the current project for five rounds of personalized fine-tuning. The prediction accuracy of the model trained through federated learning is significantly better than that of the independently trained single-project model without federated learning.

[0104] In step S5, after the neural network progress prediction engine outputs a high-risk warning at 9:00 AM on the 15th day, this warning signal triggers a collaborative workflow in real time via the event bus. The forward control flow sends the building information model spatial coordinates contained in the warning signal and the conflict time window of the 18th day to the image recognition engine. Based on this, the image recognition engine dynamically adjusts the region of interest weights of the four cameras near the hoisting operation area, increasing the sampling frequency from 1 Hz to 10 Hz. The image recognition engine extracts keyframes from the video stream for 3D reconstruction of the neural radiation field. The YOLOv11 detection network identifies the concrete surface strength characteristics of the HVAC embedded hanger. Through semantic comparison with the building information model, it is determined that the hanger has not yet fully solidified, and its strength has only reached 70% of the design value. This information is fed back to the progress prediction engine in real time via the reverse feedback flow. The progress prediction engine combines the resource conflict warning information and the hanger strength data fed back by image recognition to perform rolling predictions. The federated resource optimization engine uses the NSGA-III multi-objective evolutionary algorithm to generate an optimized scheduling scheme, recommending that the electrical engineering department prioritize using the crane on the 15th day and that the HVAC department postpone for two days.

[0105] After actual construction, the UPS installation for this project took 29 days. The revised project duration predicted on day 15 in this application is quite close. Compared to traditional Gantt chart-based methods, the prediction accuracy of this application is significantly improved. Federated learning significantly accelerates the convergence speed of the new project model, effectively saving labor costs.

[0106] Example 3 This embodiment uses the installation quality monitoring of precision air conditioning pipelines in a data center chilled water system as an example to illustrate the specific application of the image recognition monitoring method of this application. The system architecture and functions of each module in this embodiment are the same as in Embodiment 1, the difference being that the specific implementation object of steps S31 to S36 is pipeline installation quality monitoring.

[0107] The project requires monitoring the levelness and insulation quality of a 2,000-meter main chilled water pipeline. Eight ultra-high-definition cameras are deployed in the construction area, covering 120 monitoring points.

[0108] In step S31, the image recognition engine collects two hundred multi-view images of the construction site, including mobile phone photos and drone aerial photos. The neural radiation field network uses these images as input to perform three-dimensional reconstruction.

[0109] In step S32, the iterative nearest point algorithm is used to register the 3D scene reconstructed from the neural radiation field with the building information model, with a registration error of less than one centimeter.

[0110] In step S33, in the three-dimensional scene reconstructed by the neural radiation field, 540 virtual samples are generated from each real image through perspective transformation of 36 horizontal viewpoints, perturbation of five lighting conditions, and simulation of three degrees of occlusion, for a total of 10,800 virtual samples are generated. These virtual samples have geometric consistency and all carry spatial semantic labels of building information model.

[0111] In step S34, the YOLOv11 network is trained using the following configuration: the backbone employs the C3k2 mechanism, applying 3x3 and 5x5 convolutional kernels at different scales for multi-scale feature extraction. A pyramid spatial attention module is embedded in the backbone, and the Neck layer uses a bidirectional feature pyramid network. In addition to P3 and P4, a P2 detection head is added to detect small targets such as pipe clamps and bolts. The Building Information Model (BIM) semantic injection module extracts the allowable deviation of pipe design level and insulation layer thickness requirements from the BIM model, encoding them as a 128-dimensional BIM semantic vector, which is then concatenated and fused with image features. Training is performed using a mixture of 2,000 real samples and 10,800 virtual samples from neural radiation fields. The training results achieve high levels of average precision and small target detection accuracy on the test set.

[0112] In step S35, the edge computing real-time inference module is deployed on four NVIDIA Jetson AGX Orin devices. It performs adaptive cropping based on 120 regions of interest preset by the building information model, reducing about 90% of invalid computation. After decoding the real-time streaming protocol video stream, it uses the TensorRT engine to perform half-precision floating-point quantization and eight-bit integer calibration inference.

[0113] In step S36, the system detected a pipe level deviation of seven millimeters at monitoring point 15, exceeding the specified threshold of five millimeters. The total time from image acquisition to early warning push is less than thirty milliseconds. The graded early warning module determines this detection result as an emergency warning, immediately triggers an audible and visual alarm, automatically saves thirty seconds of video evidence before and after the detection, and highlights the defect location in the building information model.

[0114] In terms of implementation effectiveness, monitoring efficiency is significantly improved compared to traditional manual inspection. This application enables one person to monitor in real time, replacing multiple workers' inspections over several days. The defect detection rate is significantly better than traditional inspection, and rework costs are effectively reduced.

[0115] Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art and related fields without creative effort should fall within the scope of protection of this application. Structures, devices, and operating methods not specifically described and explained in this application shall be implemented in accordance with conventional means in the art unless otherwise specified and limited.

Claims

1. A method for intelligent image perception and control of electromechanical construction in data centers, characterized in that, Includes the following steps: Step S1: Collect multi-source construction data and perform fusion processing on the multi-source construction data to obtain fused multi-source construction data; Step S2: Based on the multi-source construction data after fusion processing, a graph neural network progress prediction engine is used to predict the construction progress. Step S3: Based on the multi-source construction data after fusion processing, a neural radiation field and building information model associated image recognition engine is used to perform real-time quality and safety monitoring of the construction site. Step S4: Perform real-time compliance verification on the fused multi-source construction data based on the standardized knowledge graph; Step S5: Perform multi-engine collaborative perception and closed-loop decision-making. The graph neural network progress prediction engine, the neural radiation field and building information model associated image recognition engine, and the federated resource optimization engine form a two-way data closed loop through an event bus and message queue. Step S6: Based on the bidirectional data loop, generate an analysis result set; Step S7: Feedback the analysis result set to the graph neural network progress prediction engine and the neural radiation field and building information model associated image recognition engine, correct, adjust and retrain each engine model, and synchronize the global model parameters after cloud federation to the edge computing device.

2. The intelligent image perception and control method for data center electromechanical construction according to claim 1, characterized in that, Step S2 specifically includes: constructing a process dependency graph, treating each process as a node, and constructing a directed heterogeneous graph containing logical dependency edges, resource competition edges, spatial conflict edges, and normative constraint edges; constructing a normative knowledge graph, and embedding normative constraint vectors into the process node representation based on the normative knowledge graph; using a relational graph convolutional network to perform message passing on the process dependency graph to obtain a deep representation of each process node; introducing a multi-head temporal attention mechanism to calculate temporal attention weights; and connecting multiple task-specific prediction heads to perform multi-task joint prediction.

3. The intelligent image perception and control method for data center electromechanical construction according to claim 2, characterized in that, The normative constraint vectors are embedded in the following way: a normative knowledge graph is constructed based on national standards, which is stored in the form of triples of normative clauses, associated processes, and constraint conditions; the entities and relations in the normative knowledge graph are embedded into a distributed representation in a low-dimensional vector space using the TransE model; for each process node, its associated normative constraint vector is concatenated into the attribute vector of the corresponding process node to form an enhanced node representation that integrates normative semantics; the multi-task joint prediction includes four tasks: critical path duration prediction, duration deviation probability distribution prediction, risk warning level assessment, and resource conflict warning; the total loss function of the four-task joint training is the weighted sum of the loss terms of each task.

4. The intelligent map perception and control method for data center electromechanical construction according to claim 2, characterized in that, Step S2 further includes: each participating project uses its own construction history data to train a relational graph convolutional network model on its local server, uploads the locally trained model parameters to the central server, adds differential privacy noise to the model parameters uploaded by each project and performs federated averaging aggregation, and distributes the aggregated global model to each project for personalized fine-tuning.

5. The intelligent map perception and control method for data center electromechanical construction according to claim 1, characterized in that, Step S3 specifically includes: acquiring multi-view construction site images for three-dimensional reconstruction of the neural radiation field; registering the reconstructed three-dimensional scene with the building information model; generating virtual training samples in the reconstructed three-dimensional scene through viewpoint transformation, lighting perturbation, and occlusion simulation; training the real and virtual samples using an improved YOLOv11 network; deploying the trained model on an edge computing device for real-time inference; and executing graded early warning based on the detection results and linking it with the building information model.

6. The intelligent map perception and control method for data center electromechanical construction according to claim 5, characterized in that, In step S3, the improved YOLOv11 network's backbone uses the C3k2 mechanism for multi-scale feature extraction and embeds a pyramid spatial attention module. The neck layer uses a bidirectional feature pyramid network. The detection head adds a P2 detection head on top of P3 and P4, and injects building information model semantic vectors into the detection head. The graded early warning includes: a green indicator for normal status, where the system continues monitoring without triggering an early warning; a yellow indicator for general alarms, where the system logs and pushes alarm information to the management application; and a red indicator for emergency early warnings, where the system triggers an audible and visual alarm, saves video evidence, and links the building information model to highlight the defect location.

7. The intelligent map perception and control method for data center electromechanical construction according to claim 1, characterized in that, The fusion processing in step S1 specifically includes: establishing a unified data model for multiple data entities and assigning a globally unique identifier to each entity; extracting data from various data sources through a protocol adapter, performing format parsing, optical character recognition, video keyframe extraction, and standard semantic alignment, and then writing it into the corresponding multimodal storage engine; deploying each core function as an independent microservice, and automatically generating extraction, transformation, and loading jobs through configuration files when a new data source is accessed; the compliance verification in step S4 specifically includes: storing the standard knowledge graph in a graph database, and storing the standard knowledge graph in the form of standard clauses, related processes or components, and constraint triples; automatically verifying whether the preconditions are met and whether the acceptance criteria are met when the process is completed, comparing the detection value with the allowable deviation threshold in the standard knowledge graph, and automatically generating a compliance acceptance form at the completion stage.

8. An intelligent image perception and control system for electromechanical construction in a data center, characterized in that, The system adopts a bottom-up five-layer architecture, as follows: The multi-source data acquisition layer is equipped with various data acquisition interfaces for accessing building information models, real-time data from IoT sensors, construction logs, video surveillance streams, drawing files, and laser point cloud data. The data fusion and governance layer uses an event-driven extraction, transformation, and loading pipeline architecture to extract, transform, load, and standardize semantic alignment of multi-source heterogeneous data. The graph-time-series collaborative perception layer includes a graph neural network progress prediction engine, a neural radiation field and building information model association image recognition engine, and a federated resource optimization engine. The graph neural network progress prediction engine, the neural radiation field and building information model association image recognition engine, and the federated resource optimization engine interact bidirectionally with the message queue through an event bus. The graph neural network progress prediction engine is responsible for the local training phase of federated learning. The federated knowledge management layer, deployed on a cloud server, forms a front-end and back-end collaborative relationship with the graph neural network progress prediction engine. It is responsible for receiving the model parameters uploaded by each project, using a federated averaging algorithm combined with a differential privacy mechanism for secure aggregation, and then distributing the aggregated global model to the local graph neural network progress prediction engine of each project for personalized fine-tuning. The application service layer provides functions such as progress control, quality monitoring, resource scheduling optimization, risk warning, and visualization.

9. The intelligent image perception and control system for data center electromechanical construction according to claim 8, characterized in that, The data fusion and governance layer includes an industrial basic parsing engine, a portable document format drawing optical character recognition module, a video keyframe extraction module, a standard semantic alignment module, and a data quality verification module. The standard semantic alignment module automatically identifies the standard clauses referenced in the data and maps them to the corresponding nodes in the knowledge graph based on the association between the standard clauses and processes stored in the standard knowledge graph.

10. The intelligent image perception and control system for data center electromechanical construction according to claim 8, characterized in that, The application service layer provides data query and command issuance functions to the outside world through the expressive state transfer interface and the GraphQL interface. The application service layer includes an intelligent progress management module, a quality and safety monitoring module, a resource scheduling optimization module, a risk warning module, and a visualization dashboard module. The risk warning module calls the corresponding warning push channel according to the warning level, including application push, SMS push, and sound and light alarm control.

Images