Rail transit full-scene intelligent construction collaborative control method and system

Abstract

The present application relates to a kind of track traffic full-scene wisdom construction collaborative control method and system, constructs " unified data pedestal-dynamic digital twin engine-collaborative control platform " three-layer closed loop architecture, through unified data pedestal real-time acquisition and fusion multiple source heterogeneous data, based on space-time alignment protocol is handled synchronously;Real-time digital twin model is constructed using dynamic digital twin engine, and the way of graph neural network and physical constraint model fusion is used to dynamically evolve prediction for construction scene;Based on real-time digital twin model, through collaborative control platform, multiple agent reinforcement learning algorithm is used to carry out autonomous path planning, task allocation and fault prediction for multiple construction equipment, realize the collaborative work between equipment and risk front control.The present application can break data barrier, realize high-fidelity real-time twin and have autonomous collaborative control capability, significantly improve construction efficiency, reduce safety risk and reduce manual intervention.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of rail transit construction technology, and in particular to a rail transit full-scene intelligent construction collaborative control method and system. BACKGROUND

[0002] In the past decade, China's rail transit has entered a period of rapid construction, with an average of over 1,000 kilometers of new operational mileage added each year. Shield tunnels, viaducts, and underground stations are being developed simultaneously, and the construction environment is becoming increasingly complex. To cope with extreme geological conditions such as ultra-deep burial, large diameter, high water pressure, and high permeability, construction parties have introduced new equipment such as BIM (Building Information Modeling), GIS (Geographic Information System), IoT (Internet of Things) sensors, UWB (Ultra-Wideband) personnel positioning, and automated shield machines, intelligent track-laying vehicles, and intelligent grouting robots, in an attempt to replace traditional manual operations with digital and mechanized means, reduce safety risks, and improve construction efficiency. However, these technologies are implemented independently by different manufacturers in different stages and sections, with significant differences in underlying data formats, spatial and temporal references, communication protocols, and update frequencies, resulting in "information silos" on site. Data needs to be processed again through manual export, table comparison, and WeChat group forwarding, with information delays typically ranging from tens of minutes to several hours, making it impossible to support minute-level or even second-level decision-making, and collaborative efficiency remains low.

[0003] In the field of visualization and simulation, the industry has attempted to build digital twin systems, but existing solutions mostly remain at the "3D visualization + periodic refresh" stage, with model updates relying on manual import of progress reports or timed batch processing, with frequencies measured in days or even weeks, making it difficult to reflect the transient changes of key parameters such as tunnel convergence deformation, shield posture deviation, groundwater seepage, and equipment oil temperature in real time. At the same time, there is a lack of bidirectional feedback mechanism between the twin model and the physical entity, making it impossible to calibrate and trace errors online during the construction process, resulting in "unreal twins and inaccurate simulations". When encountering sudden risks such as fault fracture zones, pressure water gushing, and existing pipeline conflicts, managers can only rely on experience to conduct post-mortem analysis on offline models, making it impossible to identify risk evolution trends in advance, let alone optimize construction parameters and resource scheduling through simulation and rehearsal, and safety risks remain prominent.

[0004] In addition, although large construction equipment is generally equipped with PLC (Programmable Logic Controller) and local sensors, it can realize automatic advancement of the established track, but the control logic is still mainly based on single machine scripts, and lacks group decision-making and collaborative optimization capabilities based on artificial intelligence. The real-time state cannot be shared between each shield machine, track laying machine and hoisting robot, and when obstacles, geological mutations or task changes are encountered, the equipment can only stop and wait for manual instructions, resulting in low equipment utilization and high energy consumption. Moreover, the existing system lacks predictive maintenance models for early faults such as bearing wear, hydraulic leakage and tool wear, and often "after-the-fact maintenance", resulting in unplanned downtime and directly reducing overall efficiency. In summary, the field of rail transit construction urgently needs a full-scene intelligent construction solution that can break down data barriers, achieve high-fidelity real-time twinning and have autonomous collaborative control capabilities, in order to comprehensively improve construction safety, engineering quality and construction efficiency, and meet the new demands of high-quality rail transit construction in large-scale and complex environments. SUMMARY

[0005] In order to solve the problems of fragmented information, twinning distortion and lack of device collaboration in existing rail transit construction, the present application provides a full-scene intelligent construction collaborative control method and system for rail transit, which realizes millisecond-level spatio-temporal alignment of multi-source heterogeneous data, second-level update of high-fidelity twinning model and integrated fusion of multi-machine group intelligence decision-making by building a three-layer closed-loop architecture of "unified data base-dynamic digital twinning engine-collaborative control platform".

[0006] The present application provides a full-scene intelligent construction collaborative control method for rail transit, comprising the following steps:

[0007] S1, real-time collection and fusion of multi-source heterogeneous data from BIM model, GIS system, IoT sensor, construction equipment PLC and personnel positioning system through unified data base, and synchronous processing of the data based on spatio-temporal alignment protocol;

[0008] S2, based on the synchronized data, constructing a real-time digital twinning model of the construction site using a dynamic digital twinning engine, and dynamically evolving and predicting the construction scene using a graph neural network and a physical constraint model fusion method;

[0009] S3, based on the real-time digital twinning model, using a multi-agent reinforcement learning algorithm on the collaborative control platform to perform autonomous path planning, task allocation and fault prediction for multiple construction equipment, realizing collaborative operation and risk pre-control between equipment.

[0010] Correspondingly, in addition, the present application also provides a full-scene intelligent construction collaborative control system for rail transit, comprising:

[0011] The unified data base module is used for collecting and fusing multi-source heterogeneous data from a BIM model, a GIS system, an IoT sensor, a construction equipment PLC and a personnel positioning system in real time through the unified data base module, and performing synchronous processing on the data based on a space-time alignment protocol;

[0012] The dynamic digital twin engine module is used for constructing a real-time digital twin model of a construction site based on the synchronized data by using the dynamic digital twin engine module, and performing dynamic evolution prediction on a construction scene in a manner of fusing a graph neural network and a physical constraint model.

[0013] The collaborative control platform module is used for performing autonomous path planning, task allocation and fault prediction on multiple construction equipments by using a multi-agent reinforcement learning algorithm of the collaborative control platform based on the real-time digital twin model, so as to realize collaborative operation and risk pre-control among the equipments.

[0014] In addition, the application discloses a computer readable storage medium, which stores computer instructions, and the instructions are executed by a processor to realize the control method in any of the above embodiments.

[0015] The application provides a track traffic full-scene intelligent construction collaborative control method and system, which has the following technical effects compared with the prior art:

[0016] (1) The unified data base of the application adopts a self-designed STAP space-time alignment protocol, and realizes real-time fusion and conflict resolution of data across protocols, coordinates and frequencies by flowing access of BIM geometric semantics, GIS geographic coordinates, IoT sensing time sequence, equipment PLC control words and personnel UWB positioning, and taking UTC timestamp+space hash coding as a unique key.

[0017] (2) The dynamic digital twin engine is based on a graph neural network-physical constraint coupling model, abstracts a tunnel, a track, equipment, an environment and personnel as a dynamic topology graph node, learns a construction evolution rule online, synchronously outputs a prediction field of structure deformation, equipment pose and risk probability, and updates a frequency of no less than 1Hz, and through a bidirectional feedback interface, prediction errors are fed back to physical sensors to form a self-correcting twin closed loop.

[0018] (3) The collaborative control platform of the present invention takes multi-agent reinforcement learning as its core, models each piece of equipment such as shield tunneling, track laying, hoisting, and transportation as an independent intelligent agent, drives the group strategy network to be pre-trained in the simulation environment with a three-dimensional weighted reward function of construction period, energy consumption, and safety, and then sends it to the vehicle controller through the edge computing box to realize the group collaborative operation of path self-optimization, task self-allocation, and fault self-prediction. Finally, it achieves the integration of real-time perception, dynamic modeling, intelligent decision-making and autonomous execution of the entire process of rail transit construction, which significantly improves construction efficiency, reduces safety risks and reduces human intervention. Attached Figure Description

[0019] Figure 1 A flowchart of the intelligent construction collaborative control method for rail transit in all scenarios provided in this embodiment of the invention;

[0020] Figure 2 This is an architecture diagram of a smart construction collaborative control system for rail transit across all scenarios, provided in an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of the on-site deployment of the intelligent construction collaborative control system for rail transit in all scenarios provided in this embodiment of the invention;

[0022] Figure 4 This is a schematic diagram of the risk prediction interface of the intelligent construction collaborative control system for rail transit in all scenarios provided in an embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] The application discloses a track traffic full-scene intelligent construction collaborative control method, which comprises the following steps:

[0026] S1, real-time collection and fusion of multi-source heterogeneous data from BIM models, GIS systems, IoT sensors, construction equipment PLCs and personnel positioning systems through a unified data base, and synchronous processing of the data based on a space-time alignment protocol;

[0027] S2, construction site real-time digital twin model construction by using a dynamic digital twin engine based on the synchronized data, and dynamic evolution prediction of the construction scene by using a graph neural network and a physical constraint model fusion mode;

[0028] S3, autonomous path planning, task allocation and fault prediction of multiple construction equipment by using a multi-agent reinforcement learning algorithm through a collaborative control platform based on the real-time digital twin model, so that collaborative operation and risk pre-control among the equipment are realized.

[0029] In the step S1, the space-time alignment protocol adopts a UTC timestamp + space hash coding mechanism; specifically, the step S1 comprises:

[0030] S11, data collection: firstly, data is collected from various data sources (BIM models, GIS systems, IoT sensors, construction equipment PLCs and personnel positioning systems). The BIM model provides geometric and attribute information of construction design; the GIS system provides geographic spatial coordinates; the IoT sensor (such as a pressure sensor, a temperature sensor and a vibration sensor) collects environmental and equipment state data; the construction equipment PLC (programmable logic controller) collects equipment operation parameters; and the personnel positioning system tracks the position of personnel in real time.

[0031] S12, synchronous processing of the data based on the space-time alignment protocol: in order to ensure data fusion under a unified space-time reference, the following substeps are used for processing:

[0032] S121, UTC timestamping: an accurate UTC timestamp is added to each collected data to ensure the consistency of the time dimension. This step is realized by using a high-precision clock source, so that the time reference of all data is unified, and time errors caused by clock asynchronization are avoided.

[0033] S122, space indexing of the data based on space hash coding of a construction area: the construction area is divided into grids, and each grid is allocated a unique hash code. After data collection, the data is mapped to the corresponding grid according to the geographic position, so that the space index is formed. For example, for tunnel construction, the tunnel is segmented and coded, and the position data of personnel and equipment is mapped to the corresponding grid according to the specific position in the tunnel, so that fast space positioning and retrieval are realized.

[0034] S123, aligning and interpolating compensation of spatio-temporal conflict data using a sliding window mechanism: a sliding time window is set, and the window size is dynamically adjusted according to the data update frequency. When there is a conflict in time or space (such as multiple different values appearing at the same time at the same location), weighted average or interpolation calculation is performed on the data in the sliding window to compensate for the conflict data and ensure the accuracy and continuity of the data. For example, if the same sensor has multiple readings in a short period of time, the average of the readings in the window can be taken as the final value; for spatial conflicts (such as device location data and personnel location data conflicting in the same grid), the final position is calculated according to the weight distribution of device priority or personnel safety level.

[0035] S13, data fusion and output: after the above spatio-temporal alignment processing, the data of each data source is fused into unified format streaming data, which is output in real time to the unified data foundation through Kafka and other message queues for subsequent dynamic digital twin engine and collaborative control platform. The fused data has time consistency, clear spatial index, and conflict data has been compensated, providing a solid foundation for real-time modeling and intelligent decision-making in subsequent construction scenarios.

[0036] As can be seen, step S1 realizes efficient collection and accurate synchronization processing of multi-source heterogeneous data, and real-time fusion of BIM, GIS, IoT, PLC, UWB and other system data, eliminating protocol, coordinate, frequency differences and breaking down information silos. After spatio-temporal alignment, the data has time consistency and clear spatial index, and the conflict data is compensated by a sliding window to ensure accurate and continuous data. This lays a solid data foundation for rail transit full-scene intelligent construction collaborative control, enabling subsequent digital twin modeling and intelligent decision-making to be based on real-time, accurate and integrated data, greatly improving construction collaborative efficiency and intelligent level, and providing strong support for construction safety and quality.

[0037] The updating process of the dynamic digital twin engine in step S2 includes: constructing a dynamic graph structure with construction entities as nodes and construction relationships as edges; extracting construction topology evolution features using graph neural networks; introducing a physical constraint model to limit the boundary of construction mechanics behavior; outputting the next time construction state prediction result and feeding back to the collaborative control platform.

[0038] Specifically, step S2 includes:

[0039] S21, dynamic graph structure construction:

[0040] S211、Construction entity and relationship definition: Take various entities in the construction site (such as tunnels, tracks, equipment, personnel, etc.) as nodes of the graph, and the relationships between construction entities (such as the interaction between equipment and tracks, the cooperation between personnel and equipment, etc.) as edges of the graph. For example, the tunneling relationship between a shield machine and a tunnel, the laying relationship between a track-laying machine and a track, etc.

[0041] S212、Data mapping: Map the synchronized data obtained from the unified data base into the graph structure. Each node contains the geometry, attribute, and state information of the entity, and each edge contains the interaction information between entities. For example, the shield machine node contains information such as its position, attitude, and thrust, and the edge between the shield machine and the tunnel contains information such as the tunneling speed and soil pressure.

[0042] S22、Graph neural network (GNN) feature extraction:

[0043] S221、Graph neural network construction: A graph neural network (GNN) model is used. GNN aggregates the information of nodes and their neighbors to extract the topological evolution features of the construction scene. For example, by aggregating the information of a shield machine and its surrounding environment (such as the soil layer and the tunnel structure), the mechanical features in the tunneling process are extracted.

[0044] Specifically, the graph convolution network (GCN) construction under the subway construction scene of the present application includes the following steps:

[0045] S2211、Initialization of graph structure

[0046] Node feature matrix construction: Construct a node feature matrix X, which has the same number of rows as the total number of construction entities, including shield machines, tunnel segments, track segments, and construction personnel, and the same number of columns as the total number of feature dimensions. The shield machine node feature dimension covers the tunneling depth, cutterhead speed, and soil volume, the tunnel segment feature dimension involves the lining segment state and water leakage condition, the track segment feature dimension includes the track laying progress and track gauge accuracy, and the construction personnel node feature dimension has the type of work and the length of work.

[0047] S2212、Adjacency matrix construction: Construct an adjacency matrix A, which has the same number of rows and columns as the total number of construction entities, and the element A ij represents whether there is an association between node i and node j.

[0048] For example, the association between a shield machine and a tunnel segment: if the shield machine i is tunneling the tunnel segment j, then A ij = 1. This indicates that there is a direct tunneling relationship between the shield machine and the tunnel segment, and the tunneling parameters of the shield machine (such as thrust, torque, speed, etc.) directly affect the construction state of the tunnel segment (such as deformation, water leakage condition, etc.). The association between a construction personnel and a shield machine: if the construction personnel i is operating the shield machine j, then A ij= 1. This indicates that there is an operational relationship between the construction personnel and the shield machine, and the construction personnel's operational behavior (such as adjusting parameters, maintaining equipment, etc.) directly affects the running state of the shield machine. The association between the track segment and the tunnel segment: if the track segment i is laid in the tunnel segment j, then A ij = 1. This indicates that there is a laying relationship between the track segment and the tunnel segment, and the laying progress and quality of the track segment directly affect the safety and functionality of the tunnel segment. The cooperative association between construction personnel: if there is a cooperative relationship (such as cooperative work, information transmission, etc.) between construction personnel i and construction personnel j, then A ij = 1. This indicates that there is a cooperative relationship between the construction personnel, and the cooperative behavior directly affects the construction efficiency and quality. The indirect association between the shield machine and the track segment: if the tunneling work of shield machine i directly affects the laying conditions (such as tunnel inner diameter, flatness, etc.) of track segment j, then A ij = 1. This indicates that there is an indirect construction influence relationship between the shield machine and the track segment, and the tunneling parameters and construction quality of the shield machine directly affect the laying effect of the track segment. The adjacent association between the tunnel segment and the tunnel segment: if the tunnel segment i and the tunnel segment j are adjacent in space (such as connected in front and back or side by side), then A ij = 1. This indicates that there is a spatial adjacent relationship between the tunnel segments, and the construction states (such as deformation, water leakage, etc.) of adjacent tunnel segments affect each other.

[0049] Through the above definitions, the adjacency matrix A can accurately represent the various complex relationships between the construction entities in the subway construction scene, providing accurate structural information for feature extraction and construction scene evolution prediction of the graph convolution network (GCN).

[0050] In one specific embodiment, assume that there are the following entities in the construction scene: shield machine 1 (node 1), shield machine 2 (node 2), tunnel segment 1 (node 3), tunnel segment 2 (node 4), track segment 1 (node 5), construction personnel 1 (node 6), and construction personnel 2 (node 7). According to the above definitions, the adjacency matrix A is as follows:

[0051]

[0052] Among them, node 1 (shield machine 1) has a tunneling relationship with node 3 (tunnel section 1) and an operating relationship with node 6 (construction personnel 1). Node 2 (shield machine 2) has a tunneling relationship with node 4 (tunnel section 2) and an operating relationship with node 7 (construction personnel 2). Node 3 (tunnel section 1) has a tunneling relationship with node 1 (shield machine 1), an adjacent relationship with node 4 (tunnel section 2), and a laying relationship with node 5 (track section 1). Node 4 (tunnel section 2) has a tunneling relationship with node 2 (shield machine 2), an adjacent relationship with node 3 (tunnel section 1), and a laying relationship with node 5 (track section 1). Node 5 (track section 1) has a laying relationship with nodes 3 (tunnel section 1) and 4 (tunnel section 2). Node 6 (construction personnel 1) has an operating relationship with node 1 (shield machine 1) and a cooperative relationship with node 7 (construction personnel 2). Node 7 (construction personnel 2) has an operating relationship with node 2 (shield machine 2) and a cooperative relationship with node 6 (construction personnel 1). Through this detailed adjacency matrix construction process, the complex relationships between various construction entities in the subway construction scene can be fully and accurately reflected, providing a solid foundation for subsequent graph convolution network (GCN) processing.

[0053] S222, graph convolution layer design:

[0054] S2221, first layer graph convolution layer:

[0055] Feature transformation: linear transformation is performed on the node feature matrix X, and the formula is H (0) =XW (0) , where W (0) is the first layer weight matrix, which is a parameter matrix obtained through training and learning.

[0056] Aggregation of neighbor information: the weighted sum of the neighbor features of each node is calculated to obtain a new feature representation, and the formula is H (1) =σ(AH (0) W (1) ), where W (1) is the second layer weight matrix, and σ is the ReLU activation function. This step enables the shield machine node to integrate tunnel section geological information, the tunnel section node to integrate shield machine tunneling parameters, and the personnel node to integrate cross-entity information such as shield machine operating status.

[0057] S2222, second layer graph convolution layer:

[0058] Feature transformation: linear transformation is performed based on the first layer output H (1) , and the formula is H (2) =σ(AH (1) W (2) ), where W (2)This is the third layer of the weight matrix. This process deeply extracts the characteristics of the construction scene, so that the characteristics of the tunnel boring machine node are no longer just the fusion of local tunneling parameters, but further combined with comprehensive information such as multiple surrounding tunnel sections and personnel collaboration. The characteristics of the tunnel section node also integrate more information such as the impact of tunnel boring machine operations and the laying status of related track sections. The characteristics of personnel node deeply reflect the interactive operation mode with the tunnel boring machine and tunnel section.

[0059] Aggregate neighbor information: Continue to aggregate neighbor information through adjacency matrix A to enhance the feature representation of the interaction and influence between construction entities.

[0060] S2223, Third layer convolutional layer:

[0061] Feature transformation: utilizing the second layer output H (2) A final linear transformation is performed to obtain the high-level feature representation of the construction entity, expressed by the formula H. (3) =H (2) W (3) W (3) This is the third layer of the weight matrix. The shield machine characteristics output at this stage accurately depict its comprehensive construction status, the tunnel section characteristics reflect the structural condition after being affected by various construction activities, and the personnel characteristics highlight their operational roles and efficiency in complex construction environments.

[0062] S2224. Feature Extraction: At each time step, the GNN model extracts features from the graph structure, outputting a feature vector for each node. These feature vectors reflect the current state of the construction entity and its interaction with the surrounding environment. For example, the feature vector of a tunnel boring machine node includes information such as tunneling speed, earth pressure, and cutterhead wear.

[0063] Through the aforementioned three-layer graph convolutional network, the feature matrix H of each construction entity is... (3) The output is updated to a high-level feature representation that includes information about itself and its surrounding environment. For tunnel boring machine nodes, the output features include high-level features such as tunneling speed, earth pressure, and cutterhead wear; for tunnel section nodes, the output features include high-level features such as deformation and geological stability. The final feature matrix H is then used to represent this high-level feature representation. (3) As output, these features are used for subsequent construction scenario evolution prediction. They will be used to predict the construction status at the next moment, including the position, attitude, thrust, and torque of the tunnel boring machine, the deformation and geological stability of the tunnel section, and the position and task status of personnel.

[0064] S23. Introduction of Physical Constraint Model:

[0065] S231, Physical constraint definition: Introduce a physical constraint model to limit the boundary of construction mechanical behavior. The physical constraint model corrects the output of the graph neural network according to the physical properties and mechanical laws of the construction entity. For shield tunneling, the physical constraint model limits the thrust and torque of the shield machine according to the principles of soil mechanics, ensuring that the tunneling process complies with physical laws.

[0066] In subway tunnel construction, shield machines need to operate in complex geological conditions. Assuming that a subway tunnel passes through a silt layer, which has high permeability and low internal friction angle, the thrust and torque of the shield machine must be strictly controlled during tunneling to prevent soil collapse, ground settlement, and loss of control of the shield machine's attitude.

[0067] According to the principles of soil mechanics, the thrust F of the shield machine during tunneling should satisfy the limit equilibrium condition of the soil. The thrust constraint formula is:

[0068]

[0069] The torque T of the shield machine during tunneling is related to the shear strength of the soil. The torque constraint formula is:

[0070]

[0071] where the internal friction angle of the soil , the cohesion of the soil c, the unit weight of the soil γ, the diameter of the shield machine cutterhead D, and the tunneling speed v of the shield machine.

[0072] During actual construction, high-precision thrust and torque sensors are installed on the shield machine to collect real-time data on thrust and torque during tunneling. At the same time, geological radar and ground settlement monitoring systems are used to monitor the stability of the soil in real time. The collected thrust and torque data are input into the physical constraint model in real time, and the model calculates whether the current thrust and torque exceed the constraint values according to the above formulas. If the thrust or torque exceeds the constraint value, the model will automatically issue an alarm and adjust the tunneling parameters, such as reducing the thrust or torque, adjusting the tunneling speed, etc. Based on real-time monitoring data and feedback from the physical constraint model, the tunneling parameters of the shield machine are dynamically adjusted. For example, if an increasing trend in ground settlement is detected, the model will automatically reduce the thrust of the shield machine and increase the grouting amount to stabilize the soil. Through the application of the physical constraint model, the safe tunneling of the shield machine in complex geological conditions can be effectively ensured, reducing construction risks and improving construction efficiency.

[0073] S232, Boundary constraint: the physical constraint model corrects the features extracted by the graph neural network, ensuring that the prediction results conform to the physical laws. For example, if the thrust of the shield machine predicted by the graph neural network exceeds the physical constraint range, the physical constraint model will adjust it to conform to the actual construction conditions. By correcting the features extracted by the graph neural network through the physical constraint model, it is ensured that the prediction results conform to the actual construction conditions, avoiding unreasonable construction parameter settings caused by model bias. This correction mechanism can effectively improve construction safety, reduce construction risk, and ensure stable tunneling of the shield machine under complex geological conditions.

[0074] S24, state prediction and feedback:

[0075] S241, state prediction: combining the features extracted by the graph neural network and the correction results of the physical constraint model, the dynamic digital twin engine outputs the prediction results of the next time construction state. The prediction results include the position, attitude, state, etc. of the construction entity, as well as the risk probability of the construction scene. For example, predicting the position, attitude and soil pressure of the shield machine at the next time, and the risk probability of tunnel deformation.

[0076] Suppose the features extracted by GNN include the current position (x, y, z), attitude (roll, pitch, yaw), thrust F GNN and torque T GNN . According to the previous physical constraint model, the upper limits of the physical constraints of thrust and torque are calculated respectively. The dynamic digital twin engine combines the features extracted by GNN and the correction results of the physical constraint model to output the prediction results of the next time construction state. The prediction results include the position, attitude, state, etc. of the shield machine, as well as the risk probability of the construction scene.

[0077] S242, feedback to collaborative control platform: real-time feedback of prediction results to collaborative control platform to provide data support for subsequent intelligent decision-making. The collaborative control platform adjusts the control strategy of the construction equipment according to the prediction results, realizing dynamic optimization of the construction process.

[0078] As can be seen, step S2 realizes real-time digital twin modeling and dynamic evolution prediction of the construction site. The construction of dynamic graph structure ensures the integrity and interactivity of the construction scene, the features extracted by the graph neural network reflect the complexity of the construction process, and the introduction of the physical constraint model ensures the accuracy of the prediction results. Finally, the prediction results are fed back to the collaborative control platform in real time, providing strong support for intelligent decision-making in the construction process.

[0079] The training and execution process of the multi-agent reinforcement learning algorithm includes:

[0080] S31, model each construction equipment as an independent agent, define its state space, action space and reward function; the reward function is composed of three-dimensional indexes of construction progress, energy consumption level and construction risk.

[0081] Specifically, each construction equipment (such as a shield machine, a track laying machine, a hoisting machine, a transport vehicle, etc.) is modeled as an independent agent. Each agent has its state space, action space and reward function. The state space includes state information such as the position, attitude, speed, energy consumption, remaining life of the equipment. For example, the state space of the shield machine includes the tunneling depth, the cutterhead speed, the thrust, the torque, etc. The action space includes all actions that the equipment can perform. For example, the action space of the shield machine includes adjusting the thrust, adjusting the torque, changing the tunneling speed, etc. The reward function is composed of three-dimensional indexes of construction progress, energy consumption level and construction risk. The construction progress measures the speed and efficiency of the equipment in completing the task. For example, a positive reward is obtained for completing the task ahead of schedule, and a negative reward is obtained for delaying the completion of the task; the energy consumption level measures the energy consumption of the equipment in completing the task. For example, the lower the energy consumption, the higher the reward; the construction risk measures the risks encountered by the equipment during construction, such as equipment failure, geological disasters, etc. For example, the lower the risk, the higher the reward.

[0082] Specifically, the reward function formula is as follows:

[0083]

[0084] R represents the total reward function. w1, w2, w3 are the weights of construction progress, energy consumption level and construction risk, respectively, satisfying w1+w2+w3=1. R schedule is the reward item of construction progress. R energy is the reward item of energy consumption level. R risk is the reward item of construction risk.

[0085] The construction progress reward item reflects the speed and efficiency of the equipment in completing the task. Assume that each equipment has a preset task completion time T target , and the actual completion time is T actual . The construction progress reward item is defined as:

[0086]

[0087] If the actual completion time is less than or equal to the preset time, the reward value is between [0, 1], and the earlier the completion, the higher the reward; if the actual completion time exceeds the preset time, the reward value is 0, indicating delayed completion.

[0088] The energy consumption level reward item reflects the energy consumed by the equipment in completing the task. Assume that the ideal energy consumption required by the equipment to complete the task is E target , and the actual energy consumption is E actualThe energy consumption level reward item is defined as:

[0089]

[0090] If the actual energy consumption is less than or equal to the ideal energy consumption, the reward value is between [0, 1], and the lower the energy consumption, the higher the reward; if the actual energy consumption exceeds the ideal energy consumption, the reward value is 0, indicating that the energy consumption is too high.

[0091] The construction risk reward item reflects the risks encountered by the equipment during the construction process. It is assumed that the types of risks that the equipment may encounter during the construction process include geological disaster risk R geological and equipment failure risk R mechanical . The construction risk reward item is defined as:

[0092]

[0093] α is the weight of the geological disaster risk, satisfying 0≤α≤1. R geological is the probability of the geological disaster risk, with a value range of [0, 1], and 1 indicates that the geological disaster risk is extremely high. R mechanical is the probability of the equipment failure risk, with a value range of [0, 1], and 1 indicates that the equipment failure risk is extremely high. The higher the risk probability, the lower the reward value, and the minimum is 0.

[0094] In practical applications, the weights w1, w2, w3 and α need to be adjusted according to the specific needs of the construction project. For example, if the project has very strict requirements on the construction period, the weight of w1 can be increased. If the project has strict requirements on energy consumption control, the weight of w2 can be increased. If the project is in a complex geological condition, the weight of α can be increased to pay more attention to the geological disaster risk. The reward value reflects the optimization degree of the equipment in completing the task, considering the construction period, energy consumption level and construction risk. Through this complex reward function design, it ensures that the multi-agent reinforcement learning algorithm can effectively balance multiple key factors in the actual construction scene and achieve the optimal construction control strategy.

[0095] S32, using a centralized training-distributed execution (CTDE) framework to pre-train the policy network in a simulation environment.

[0096] Specifically, a centralized training-distributed execution (CTDE) framework is used; in the centralized training phase, the policy networks of all agents are jointly trained in a simulation environment to learn the optimal collaborative strategy. In the distributed execution phase, each agent independently executes actions according to the trained policy network.

[0097] The training process is as follows:

[0098] S321, a high-fidelity construction simulation environment is constructed, including dynamic models of shield machines, track laying machines, hoisting machines, transport vehicles and other equipment, as well as geological conditions of the construction site, task requirements and the like.

[0099] S322, each agent initializes a policy network, and the input of the policy network is the state space and the output is the action space.

[0100] S323, in the simulation environment, all agents select actions according to the current state, obtain new states and rewards after executing the actions. The policy network of each agent is updated through a multi-agent reinforcement learning algorithm to maximize the cumulative reward.

[0101] Specifically, the MADDPG algorithm is used to update the policy network of each agent, including the following steps:

[0102] S3231, initialize the policy network (Actor) and value network (Critic) of each agent, initialize the target network, and synchronize them with the main network:

[0103] The goal of the policy network is to learn a policy π i , so that the action of agent i can maximize the cumulative reward. The update of the policy network is based on the policy gradient, and its update formula is:

[0104]

[0105] Where: is the policy network parameter of agent i. is the learning rate, which controls the step size of parameter update.

[0106] is the performance function, defined as:

[0107]

[0108] Where: is the observation state of all agents. is the action of all agents. is the value function of agent i, which represents the expected cumulative reward of taking action in state .

[0109] The goal of the value network is to accurately evaluate the goodness of the current policy. The update of the value network is based on the mean square error (MSE) loss function, and its update formula is:

[0110]

[0111] Where: is the value network parameter of agent i. is the learning rate of the value network. N is the batch size.

[0112] is the target Q-value, defined as:

[0113]

[0114] where: is the reward obtained by taking action in state . is the discount factor, representing the discount rate of future rewards. is the new state after performing action . is the action taken by all agents in new state .

[0115] To improve the stability of training, MADDPG uses a target network to calculate the target Q-value. The parameters of the target network are synchronized with the main network through a soft update policy, and its update formula is:

[0116]

[0117]

[0118] where: and are the target network parameters of the policy network and the value network, respectively. is the soft update coefficient, usually a very small positive number such as 0.001, controlling the speed of updating the target network parameters.

[0119] S3232, training loop:

[0120] S32321, state observation and action selection: Each agent (such as a shield machine, a crane, a transport vehicle) observes its own state (such as position, speed, energy consumption, etc.) and part of the environment state (such as geological conditions, task progress, etc.). According to the policy network, the agent selects an action (such as adjusting the thrust, changing the speed, hoisting materials, etc.).

[0121] S32322, reward calculation: According to the reward function, calculate the reward obtained by each agent after performing the action. The reward function considers the progress of the construction period, the level of energy consumption and the construction risk.

[0122] S32322, experience storage and sampling: Store the experience (state, action, reward, new state) in the replay buffer. Randomly sample a small batch of experience from the replay buffer for network update.

[0123] S32322, Network Update: Update the value network using the extracted experiences, minimizing the mean squared error between predicted Q-values and target Q-values. Evaluate the policy gradient using the updated value network, and update the policy network. Use soft updates to synchronize the target network.

[0124] S3233, Collaboration and Optimization:

[0125] Agents adjust actions according to the updated policy network, achieving collaborative control. Through continuous iteration, agents learn optimal strategies in the simulated environment and apply these strategies to actual construction. In the subway construction scenario, the MADDPG algorithm enables effective collaboration between agents, such as the cooperation between the shield machine and the crane: the shield machine adjusts the tunneling parameters according to the tunneling depth and geological conditions, and the crane adjusts the lifting tasks according to the shield machine progress. Transport vehicle path planning: the transport vehicle dynamically adjusts the transport route according to the real-time state of the construction site, avoiding congestion and improving transport efficiency.

[0126] S3234, Risk Avoidance:

[0127] Agents adjust strategies according to risk predictions, such as reducing the tunneling speed of the shield machine in high-risk geological sections and reducing the lifting height of the crane in strong wind weather.

[0128] S33, Deploy the trained policy network to actual construction equipment to achieve real-time collaborative control.

[0129] S331, Deploy the trained policy network to actual construction equipment. Each device runs the policy network through an edge computing box, real-time receives the state information of the device, and outputs the optimal action.

[0130] S332, During the construction process, each agent independently executes actions according to the output of the policy network, achieving autonomous path planning, task allocation, and fault prediction. For example, the shield machine adjusts the tunneling parameters according to the predicted geological conditions, the track layer adjusts the laying progress according to the shield machine tunneling progress, and the crane adjusts the lifting tasks according to the device requirements.

[0131] S333, By real-time monitoring of device status and construction environment, potential risks are predicted and measures are taken in advance. For example, if the shield machine cutter is predicted to be severely worn, maintenance is scheduled in advance; if the tunnel deformation risk increases, adjust the tunneling parameters or take reinforcement measures.

[0132] Through the above multi-agent reinforcement learning algorithm, autonomous collaborative control of multiple construction equipment is realized, which significantly improves construction efficiency, reduces energy consumption and construction risk. Real-time collaboration between equipment reduces waiting time and task conflicts, improving overall construction progress; by optimizing the operating parameters of the equipment, the energy consumption of the equipment is reduced, and the energy utilization efficiency is improved; by monitoring and predicting potential risks in real time, measures are taken in advance to reduce the probability of equipment failure and geological disasters, improving construction safety.

[0133] In addition, the present application also provides a track traffic full-scene intelligent construction collaborative control system, comprising:

[0134] The unified data base module is used for real-time acquisition and fusion of multi-source heterogeneous data from BIM models, GIS systems, IoT sensors, construction equipment PLCs and personnel positioning systems through the unified data base module, and synchronous processing of the data based on a space-time alignment protocol;

[0135] The dynamic digital twin engine module is used for constructing a real-time digital twin model of the construction site based on the synchronized data using the dynamic digital twin engine module, and dynamically evolving and predicting the construction scene in a graph neural network and physical constraint model fusion manner;

[0136] The collaborative control platform module is used for autonomous path planning, task allocation and fault prediction of multiple construction equipment through the collaborative control platform using a multi-agent reinforcement learning algorithm based on the real-time digital twin model, realizing collaborative operation and risk pre-control between equipment.

[0137] The track traffic full-scene intelligent construction collaborative control system provided by the embodiments of the present application can perform the control method of any of the above embodiments.

[0138] In another embodiment, the present application discloses a computer readable storage medium, which stores computer instructions, and the instructions are executed by a processor to realize the control method in any of the above embodiments.

[0139] It should be understood that the various forms of flow shown above can be used to reorder, add or delete steps. For example, each step described in the present application can be executed in parallel, sequentially or in a different order, as long as the desired results of the technical solutions of the present application can be achieved, which is not limited herein.

[0140] The above detailed description does not limit the scope of the application. Various modifications, combinations, sub-combinations and alternatives can be made to the detailed description. Any modification, equivalent replacement and improvement etc. made within the spirit and principle of the application shall be included in the scope of the application.

Claims

1. A rail transit full-scene intelligent construction collaborative control method, characterized in that, The method comprises the following steps: S1, real-time collection and fusion of multi-source heterogeneous data from BIM models, GIS systems, IoT sensors, construction equipment PLCs and personnel positioning systems through a unified data base, and synchronous processing of the data based on a space-time alignment protocol; S2, based on the synchronized data, construction site real-time digital twin models are constructed using a dynamic digital twin engine, and a graph neural network is used in combination with a physical constraint model to dynamically evolve and predict the construction scene; S3, based on the real-time digital twin models, multi-agent reinforcement learning algorithms are used on a collaborative control platform to autonomously plan paths, assign tasks and predict faults for multiple construction equipment, achieving collaborative operation and risk pre-control between equipment; Step S1 specifically comprises: S11, data collection: real-time collection of data from various data sources; S12, synchronous processing of the data based on a space-time alignment protocol: to ensure data fusion under a unified space-time reference, the following sub-steps are used for processing: S121, UTC timestamp: an accurate UTC timestamp is added to each collected data to ensure consistency in the time dimension; S122, spatial indexing of data based on space hash coding of the construction area: the construction area is divided into grids, and each grid is assigned a unique hash code; after data collection, the data is mapped to the corresponding grid according to its geographic location to form a spatial index; S123, alignment and interpolation compensation of space-time conflict data using a sliding window mechanism: a sliding time window is set, and the window size is dynamically adjusted according to the data update frequency; when there is a conflict in time or space, the data within the sliding window is used for weighted averaging or interpolation calculation to compensate for the conflict data and ensure the accuracy and continuity of the data; if the same sensor has multiple readings in a short period of time, the average of the readings within the window is taken as the final value; for spatial conflicts, the final position is calculated after weight distribution according to the equipment priority or personnel safety level; S13, data fusion and output: after space-time alignment processing, the data from various data sources is fused into a unified format of streaming data, which is output in real time to the unified data base through a Kafka message queue for use by the subsequent dynamic digital twin engine and collaborative control platform; Step S2 specifically comprises: S21, definition of entities and relationships in the construction site, taking the entities as nodes of the graph and the relationships between the entities as edges of the graph, and mapping the data into a graph structure to provide a basis for subsequent feature extraction and prediction; S211, construction entity and relationship definition: various entities in the construction site are taken as nodes of the graph, and the relationships between the construction entities are taken as edges of the graph; S212, data mapping: the synchronized data obtained from the unified data base is mapped into a graph structure; each node contains geometric, attribute and state information of the entity, and each edge contains interaction information between the entities; S22, the information of the nodes and their neighbors is aggregated using a graph neural network, and topological evolution features of the construction scene are extracted through multi-layer graph convolution operations to generate high-level feature representations reflecting the state of the construction entities and their interactions; S23, the output of the graph neural network is corrected by introducing a physical constraint model to ensure that the prediction result conforms to the physical law of actual construction, and the thrust and torque of the shield machine are limited; S24, combining the features extracted by the graph neural network and the correction result of the physical constraint model, the dynamic digital twin engine outputs the prediction result of the construction state at the next time, and the prediction result is fed back to the collaborative control platform in real time to provide data support for intelligent decision-making; The step S22 specifically comprises: S221, graph neural network construction: a graph neural network (GNN) model is adopted; the topological evolution characteristics of the construction scene are extracted by aggregating the information of the nodes and their neighbors; the mechanical characteristics in the tunneling process are extracted by aggregating the information of the shield machine and its surrounding environment; S222, graph convolution layer design: S2221、first layer graph convolution layer: linear transformation is performed on the node feature matrix X, and the formula is H (0) =XW (0) , where W (0) is a first layer weight matrix, which is a matrix of parameters learned through training; By calculating the weighted sum of each node neighbor feature, a new feature representation is obtained, and the formula is H (1) =σ(AH (0) W (1) ), wherein W (1) is the second layer weight matrix, and σ is the ReLU activation function; the shield machine node fuses the tunnel segment geological information, the tunnel segment node is integrated into the shield machine tunneling parameter, and the personnel node combines the shield machine running state cross-entity information; S2222, the second layer of graph convolution layer: Based on the first layer output H (1) , a linear transformation is performed, and the formula is H (2) =σ(AH (1) W (2) ), where W (2) is the third layer weight matrix; aggregate neighbor information: continue to aggregate neighbor information through the adjacency matrix A, and strengthen the feature representation of the interaction between construction entities; S2223, the third layer of graph convolution layer: The second layer output H (2) is used to perform a final linear transformation to obtain a high-level feature representation of the construction entity, which is given by H (3) = H (2) W (3) , where W (3) is a third layer weight matrix. S2224, feature extraction: the GNN model extracts features from the graph structure at each time step, and outputs the feature vector of each node, which reflects the state of the construction entity at the current time and its interaction with the surrounding environment; The step S23 specifically comprises: S231, physical constraint definition: a physical constraint model is introduced to limit the boundary of the construction mechanics behavior; the physical constraint model corrects the output of the graph neural network according to the physical properties and mechanical laws of the construction entity; for shield tunneling, the physical constraint model limits the thrust and torque of the shield machine according to the principles of soil mechanics; According to the principle of soil mechanics, the thrust F of the shield machine during tunneling should satisfy the limit equilibrium condition of the soil; the thrust constraint formula is: The torque T of the shield machine during tunneling is related to the shear strength of the soil. The torque constraint formula is: Where, the internal friction angle of the soil is ϕ, the cohesion of the soil is c, the specific gravity of the soil is γ, the diameter of the shield machine cutter is D, and the tunneling speed of the shield machine is v; S232, boundary limitation: the physical constraint model corrects the features extracted by the graph neural network to ensure that the prediction result conforms to the physical law; if the thrust of the shield machine predicted by the graph neural network exceeds the physical constraint range, the physical constraint model will adjust it to meet the actual construction conditions.

2. The rail transit full-scene intelligent construction collaborative control method of claim 1, wherein The step S24 specifically comprises: S241, state prediction: combining the features extracted by the graph neural network and the correction result of the physical constraint model, the dynamic digital twin engine outputs the prediction result of the construction state at the next time; the prediction result includes the position, attitude and state information of the construction entity, and the risk probability of the construction scene; S242, feedback to the collaborative control platform: the prediction result is fed back to the collaborative control platform in real time to provide data support for subsequent intelligent decision-making; the collaborative control platform adjusts the control strategy of the construction equipment according to the prediction result to realize the dynamic optimization of the construction process.

3. The rail transit full-scene intelligent construction collaborative control method of claim 2, wherein The step S3 specifically comprises: The training and execution process of the multi-agent reinforcement learning algorithm includes: S31, model each construction equipment as an independent agent, define its state space, action space and reward function; the reward function is composed of the weighted three-dimensional indexes of construction period, energy consumption level and construction risk; S32, pre-train the policy network in the simulation environment using the centralized training-distributed execution framework; S33, deploy the trained policy network to the actual construction equipment to realize real-time collaborative control.

4. The rail transit full-scene intelligent construction collaborative control method of claim 3, wherein The step S32 specifically comprises: S321, a high-fidelity construction simulation environment is constructed, including dynamic models of shield machines, track-laying machines, hoisting machines, transport vehicles, and geological conditions of the construction site and task requirements; S322, each agent initializes a policy network, and the input of the policy network is a state space and the output is an action space; S323, in the simulation environment, all agents select actions according to the current state, obtain new states and rewards after executing the actions, and update the policy network of each agent through a multi-agent reinforcement learning algorithm to maximize the cumulative reward.

5. The rail transit full-scene intelligent construction collaborative control method of claim 4, wherein The step S33 specifically comprises: S331, the trained policy network is deployed to actual construction equipment; each device runs the policy network through an edge computing box, receives state information of the device in real time, and outputs an optimal action; S332, during the construction process, each agent independently executes an action according to the output of the policy network, realizes autonomous path planning, task allocation and fault prediction, the shield machine adjusts the tunneling parameters according to the predicted geological conditions, the track-laying machine adjusts the laying progress according to the tunneling progress of the shield machine, and the hoisting machine adjusts the hoisting task according to the equipment requirements; S333, potential risks are predicted by monitoring the device state and the construction environment in real time, and measures are taken in advance; if it is predicted that the shield machine cutter is severely worn, maintenance is arranged in advance; if it is predicted that the risk of tunnel deformation increases, the tunneling parameters are adjusted or reinforcement measures are taken.

6. A rail transit full-scene intelligent construction collaborative control system, adopting the rail transit full-scene intelligent construction collaborative control method according to any one of claims 1-5, characterized in that, It comprises: A unified data base module is used to collect and fuse multi-source heterogeneous data from BIM models, GIS systems, IoT sensors, construction equipment PLCs and personnel positioning systems in real time through the unified data base module, and the data is synchronously processed based on a space-time alignment protocol; A dynamic digital twin engine module is used to construct a real-time digital twin model of the construction site based on the synchronized data using the dynamic digital twin engine module, and a graph neural network and a physical constraint model are fused to predict the dynamic evolution of the construction scene; A collaborative control platform module is used to realize autonomous path planning, task allocation and fault prediction of multiple construction equipment through the collaborative control platform using a multi-agent reinforcement learning algorithm based on the real-time digital twin model, and to realize collaborative operation and risk pre-control between devices.

Images