A tunnel construction process intelligent scheduling optimization method and device based on GAT-GCN-DDQN

Abstract

This invention belongs to the field of tunnel construction scheduling optimization technology, specifically relating to an intelligent scheduling optimization method and device for tunnel construction procedures based on GAT-GCN-DDQN. This method, by considering the comprehensive impact of procedure dependencies, resource allocation constraints, and dynamic changes in the construction environment, constructs an attention feature extraction model based on GAT, a global topology model based on GCN, and a reinforcement learning scheduling optimization model based on DDQN. This invention comprehensively considers key factors such as multi-procedure dependencies, multi-resource constraints, and dynamic disturbances, providing effective guidance for the intelligent generation and real-time optimization of tunnel construction scheduling strategies. Simultaneously, it provides a theoretical foundation and technical support for the intelligent upgrading of tunnel construction management systems and on-site decision support.

Description

Technical Field

[0001] This invention belongs to the field of tunnel construction scheduling optimization technology, specifically relating to a method and device for intelligent scheduling optimization of tunnel construction procedures based on GAT-GCN-DDQN. Background Technology

[0002] Tunnel construction process scheduling is a crucial link in improving project efficiency and resource utilization. It carries the core functions of orderly task progression, rational resource allocation, and construction cycle optimization, serving as a vital bridge connecting construction plans and on-site execution. Traditional methods rely primarily on manual experience and rule setting. This manual command model is not only labor-intensive and inefficient but also highly susceptible to subjective human factors, making it difficult to cope with the complex scheduling needs of large-scale, multi-process, and multi-resource coupling. Especially in dynamic construction environments, existing technologies lack robustness in handling sudden disturbances such as geological changes and equipment failures, and lack real-time and adaptive capabilities. Although some research has introduced intelligent algorithms for optimization, most methods are highly targeted and struggle to simultaneously address the multi-objective requirements of minimizing construction time and maximizing resource utilization. Furthermore, they generally suffer from weak global modeling capabilities, insufficient state representation, and poor strategy generalization, failing to meet the demands of intelligent development in actual construction scheduling. Summary of the Invention

[0003] In order to solve at least one of the above-mentioned technical problems in the prior art, the present invention provides a method and apparatus for intelligent scheduling and optimization of tunnel construction procedures based on GAT-GCN-DDQN.

[0004] This invention is implemented using the following technical solution: an intelligent scheduling and optimization method for tunnel construction procedures based on GAT-GCN-DDQN, comprising: Historical data on tunnel construction procedures were acquired and structured to extract procedure attributes, resource attributes, and process topology dependencies of procedure components, resulting in standardized historical data on tunnel construction procedures. Based on the historical data of standardized tunnel construction procedures, a heterogeneous analysis graph of historical data containing procedure nodes and resource nodes is dynamically constructed, and dynamic feature vectors are assigned to procedure nodes and resource nodes. Based on graph attention network, graph convolutional network and dual-depth Q network, an intelligent scheduling and optimization model for tunnel construction procedures is constructed, and the intelligent scheduling and optimization model for tunnel construction procedures is trained by heterogeneous extraction graph of historical data. By analyzing the progress and resource usage snapshots of the construction site, the physical site status is obtained. The physical site status is then synchronously mapped to the virtual environment to construct a digital twin of the tunnel construction project. The synchronous status is input into a pre-trained intelligent scheduling and optimization model for tunnel construction procedures for real-time reasoning. The model outputs the optimal procedure and resource scheduling instructions for the current moment, as well as the construction schedule plan for subsequent tunnel projects.

[0005] Preferably, in the steps of acquiring historical data on tunnel construction procedures and performing structured processing to extract procedure attributes and resource attributes, the extraction of resource attributes and procedure attributes is a process of abstracting physical construction elements into the smallest computable unit, including: Construction resources are abstracted into complete work teams or sets of machinery with independent operating capabilities, and the maximum parallel operation capacity of the smallest work unit on the construction site is defined as the total resource, which is used to constrain the cumulative usage of the resource by all processes; among them, construction elements include extracted process attributes and resource attributes; construction resources include human resources and physical construction resources; Based on the historical construction logs, the distribution patterns of construction period are mined, and the standard time distribution of each process is set as a probability distribution function that includes a baseline value and a random fluctuation range. A compatibility mapping table between processes and resource types is constructed to clarify the pool of compliant resource candidates that can be used when each process is executed.

[0006] Preferably, based on standardized tunnel construction procedure historical data, a heterogeneous analytical graph containing procedure nodes and resource nodes is dynamically constructed, and dynamic feature vectors are assigned to the procedure nodes and resource nodes, including: Construct a set of process nodes and a set of resource nodes and establish a connection relationship, including a unidirectional temporal dependency edge to characterize the flow of construction technology and a bidirectional supply and demand interaction edge to characterize the relationship between process and resource occupation and release; Construct dynamic feature vectors for process nodes. The dynamic feature vectors for process nodes include normalized remaining duration values, binary identifiers representing whether the process is completed, and concatenated binary identifiers representing whether the process is currently consuming resources. Construct a dynamic feature vector for resource nodes. The dynamic feature vector for resource nodes includes the current available amount of resources, the total capacity of resources, and the load rate value representing the real-time occupancy ratio of resources, which are concatenated.

[0007] Preferably, the intelligent scheduling and optimization model for tunnel construction procedures is trained using heterogeneous extraction maps from historical data, including: The heterogeneous extraction graph of historical data is input into the graph attention network layer of the intelligent scheduling and optimization model for tunnel construction procedures. The multi-head attention mechanism is used to calculate the interaction weights between nodes, and high weight coefficients are assigned to process nodes located on the critical path according to the process dependency. The weighted process node features are input into the graph convolutional network layer, and neighborhood feature aggregation is performed in combination with the graph topology, thus integrating the supply and demand distribution information of resource nodes into the feature representation of process nodes. The global pooling mechanism is used to aggregate the features of all encoded process nodes, mapping the variable-length graph data with dynamic node size into a high-dimensional feature vector with constant dimension representing the global state. High-dimensional feature vectors are input into a dual deep Q-network layer and mapped to a discrete action space containing process execution and waiting. The design incorporates a standardized reward mechanism that integrates total project duration optimization rewards, resource utilization feedback, and multiple constraint penalties. The network parameters of the intelligent scheduling optimization model for tunnel construction procedures are updated iteratively through experience playback and gradient descent, and the model is trained until the strategy converges.

[0008] Preferably, the high-dimensional feature vector is input into a dual deep Q-network layer and mapped to a discrete action space containing process execution and waiting, including: At each decision step, all process tasks are traversed, and processes that have been completed or are in progress are eliminated based on the current status of the process, processes that have not met their prerequisites are eliminated based on the process dependency matrix, and processes that lack the necessary human or mechanical resources are eliminated based on the real-time availability of resources. Generate a subset of valid actions containing only verified process assignment instructions and waiting instructions, and select actions from this subset to ensure that the output scheduling instructions always meet the physical and logical constraints of tunnel construction.

[0009] Preferably, a standardized reward mechanism is designed that integrates total project duration optimization rewards, resource utilization feedback, and multiple constraint penalties, including: Construct a standardized reward function, which includes total project duration reward, resource utilization reward, and constraint penalty terms; Store the quadruple that includes the current global features of the heterogeneous graph, the selected effective actions, the dense step reward value, and the global features of the heterogeneous graph at the next time step into the experience replay pool. The target Q-value based on intensive step reward is calculated using a dual Q-learning mechanism, and the prediction error is minimized through backpropagation algorithm, driving the network parameters to converge toward the Pareto optimal boundary that minimizes the project duration and maximizes resource utilization.

[0010] Preferably, the normalized reward function includes: The total construction period bonus is calculated based on the ratio of the actual total construction period after all work processes are completed to the reference construction period. It is used to guide the model to optimize overall construction efficiency. The formula for the total construction period bonus is as follows: in, The total project duration reward function; The project duration after all processes are completed. This serves as a reference lead time for the process. The weighting coefficient for the total project duration bonus; Resource utilization rewards include: incentive allocation rewards based on the number of currently executing processes to improve parallelism; overload prevention penalties for resource usage exceeding total capacity; and fixed negative waiting penalties imposed on waiting actions. The formula for resource utilization rewards is expressed as: in, For resource utilization reward function, , and The definitions are as follows: in, To incentivize the allocation of reward values; Assign basic weights to incentives; Weighting of rewards based on the number of work processes in the construction project; This indicates the number of processes currently in progress; in, This is the penalty value for overload prevention. Weighting for overload penalties; For resource type serial number; The number of resource types; in, The penalty value is the value to wait for. For waiting penalty weight; Constraint penalty terms are used to penalize scheduling behaviors that violate process dependencies, resource conflicts, and resource capacity limits; the formula for constraint penalty terms is expressed as: in, Penalties for violations of process dependencies; Penalties for violating resource capacity limits; Penalties for resource conflicts; Combining the above three reward mechanisms, the final reward function is as follows: In the formula, Bonus based on total project duration; Incentives for resource utilization; To constrain penalty items.

[0011] Preferably, the objective function formula of the intelligent scheduling optimization model for tunnel construction procedures is expressed as: in, and The weight coefficients of the objective function represent the importance of total project duration and resource utilization rate in the model optimization, respectively, satisfying the following conditions: ; Total construction period; To serve as a reference for the total construction period; based on historical experience and theoretical analysis, the average construction period is selected as a reference basis to provide a benchmark for the normalization of the total construction period, so that the total construction period of tunnel construction projects of different scales or complexities can be compared and analyzed on a unified scale. This represents the average utilization rate of all resources.

[0012] Preferably, a digital twin of the tunnel construction project is constructed; the synchronous state is input into a trained intelligent scheduling and optimization model for tunnel construction procedures for real-time reasoning, and the model outputs the optimal procedure and resource scheduling instructions for the current moment, as well as the construction schedule plan for subsequent tunnel projects, including: The execution status injection operation parses the structured field snapshot data containing real-time process status, execution time, and resource usage records. It iterates through the process list in the data, marks the processes marked as "in progress" as occupied in the virtual environment and corrects the remaining duration. At the same time, it forcibly deducts the corresponding available resources from the virtual resource pool and advances the simulation clock to the field timestamp. Perform dynamic rescheduling reasoning. Based on the synchronized virtual environment state, use the trained model to generate the process assignment instructions that need to be executed immediately at the current moment, and continuously deduce subsequent strategies until the project ends, generating a dynamically adjusted Gantt chart containing the expected start and end times of all unfinished processes.

[0013] This invention is achieved using the following technical solution: an intelligent scheduling and optimization device for tunnel construction procedures based on GAT-GCN-DDQN, comprising: The data acquisition module is used to acquire historical data of tunnel construction procedures, perform structured processing, extract procedure attributes, resource attributes, and process topology dependencies of procedure components, and obtain standardized historical data of tunnel construction procedures. The data processing module is used to dynamically construct a heterogeneous extraction graph of historical data containing process nodes and resource nodes based on standardized historical data of tunnel construction procedures, and to assign dynamic feature vectors to process nodes and resource nodes. The model building and training module is used to build an intelligent scheduling and optimization model for tunnel construction procedures based on graph attention network, graph convolutional network and dual depth Q network, and to train the intelligent scheduling and optimization model for tunnel construction procedures by extracting graphs from historical data heterogeneously. The intelligent scheduling and optimization module is used to analyze the progress and resource usage snapshots of the construction site to obtain the physical site status. The physical site status is synchronously mapped to the virtual environment to construct a digital twin of the tunnel construction project. The synchronous status is input into the trained intelligent scheduling and optimization model of tunnel construction procedures for real-time reasoning. The model outputs the optimal procedure and resource scheduling instructions at the current moment and the construction schedule plan for subsequent tunnel projects.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention, by considering the comprehensive impact of process dependencies, resource allocation constraints, and dynamic changes in the construction environment, constructs an attention feature extraction model based on GAT, a global topology model based on GCN, and a reinforcement learning scheduling optimization model based on DDQN. By comprehensively considering key factors such as multi-process dependencies, multi-resource constraints, and dynamic disturbances, this invention provides effective guidance for the intelligent generation and real-time optimization of tunnel construction scheduling strategies. Simultaneously, it provides a theoretical foundation and technical support for the intelligent upgrading of tunnel construction management systems and on-site decision support. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating an intelligent scheduling and optimization method for tunnel construction procedures based on GAT-GCN-DDQN provided by the present invention.

[0017] Figure 2 This is a logical schematic diagram of an intelligent scheduling and optimization method for tunnel construction procedures based on GAT-GCN-DDQN provided by the present invention.

[0018] Figure 3 This is a schematic diagram of the process-resource heterogeneous extraction graph model in the intelligent scheduling and optimization method for tunnel construction processes based on GAT-GCN-DDQN provided by this invention.

[0019] Figure 4 This is a schematic diagram of the scheduling decision process at the first critical time point in the intelligent scheduling optimization method for tunnel construction procedures based on GAT-GCN-DDQN provided by the present invention.

[0020] Figure 5This is a schematic diagram of the scheduling decision process at the second critical time point in the intelligent scheduling optimization method for tunnel construction procedures based on GAT-GCN-DDQN provided by the present invention.

[0021] Figure 6 This is a schematic diagram of the scheduling decision process at the third critical time point in the intelligent scheduling optimization method for tunnel construction procedures based on GAT-GCN-DDQN provided by the present invention.

[0022] Figure 7 This is a schematic diagram of the scheduling decision process at the fourth critical time point in the intelligent scheduling optimization method for tunnel construction procedures based on GAT-GCN-DDQN provided by this invention.

[0023] Figure 8 This is a schematic diagram of the scheduling decision process at the fifth critical time point in the intelligent scheduling optimization method for tunnel construction procedures based on GAT-GCN-DDQN provided by the present invention.

[0024] Figure 9 This is a schematic diagram of the scheduling decision process at the sixth critical time point in the intelligent scheduling optimization method for tunnel construction procedures based on GAT-GCN-DDQN provided by this invention.

[0025] Figure 10 This is a logical diagram of the dual-depth Q-network algorithm in the intelligent scheduling and optimization method for tunnel construction procedures based on GAT-GCN-DDQN provided by the present invention.

[0026] Figure 11 This is a schematic diagram of the structure of an intelligent scheduling and optimization device for tunnel construction procedures based on GAT-GCN-DDQN provided by the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings. 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 implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] It should be noted that the structures, proportions, sizes, etc., shown in the accompanying drawings of this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportional relationships, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should fall within the scope of the technical content disclosed in the present invention. It should be noted that in this specification, relational terms such as "first" and "second" are only used to distinguish one entity from several other entities, and do not necessarily require or imply any actual relationship or order between these entities.

[0029] To address the technical bottlenecks of traditional tunnel construction sequence scheduling methods in handling large-scale, multi-process, multi-resource coupled, and dynamic construction environments, and because existing methods struggle to simultaneously address the multi-objective optimization problem of minimizing construction period and maximizing resource utilization, this embodiment proposes the following... Figure 1 and Figure 2 The method for intelligent scheduling and optimization of tunnel construction procedures based on GAT-GCN-DDQN, as shown, includes: S1: Obtain historical data of tunnel construction procedures, perform structured processing, extract procedure attributes, resource attributes, and process topology dependencies of procedure components to obtain standardized historical data of tunnel construction procedures.

[0030] Multi-dimensional data collection is conducted to cover key information throughout the entire tunnel construction process. This includes selecting complete records from at least three similar completed projects in historical construction logs to extract core data such as actual time consumption for each process, detailed resource input, and records of geological condition changes; standard process quotas refer to current industry standards and the company's internal construction standards to clarify the theoretical time range and minimum resource allocation threshold for each process; and construction plan data focuses on process flow design, key node definition, and resource allocation planning to ensure data coverage of the three core dimensions required for scheduling decisions: processes, resources, and dependencies.

[0031] To address the characteristics of tunnel construction data—multi-source heterogeneity, high noise, and complex logic—this embodiment employs the following in-depth processing flow. First, by acquiring historical construction logs, standard procedure quotas, and construction plan data for the tunnel project, the focus is on extracting procedure attributes (including procedure identifier, standard time distribution, and resource requirements), resource attributes (including resource identifier, type, and total quantity), and the technological topology dependencies between procedures, providing standardized input for constructing the scheduling model. The standard quota database, based on the "Technical Regulations for Safety in Railway Tunnel Engineering Construction" and internal enterprise quotas, establishes theoretical time intervals and resource allocation benchmark thresholds for standard procedures under various levels of surrounding rock. The construction plan topology data is then analyzed from the construction organization design documents to clarify the procedure connection logic under the "three-step seven-stage excavation method" or the "full-section method." During tunnel construction, different procedures may be performed cyclically. For example, a complete cycle typically includes drilling, blasting, ventilation and smoke dispersal, muck removal and transportation, initial support, invert construction, and secondary lining. These procedures are completed cyclically within the same construction section in a strict sequence, and as the tunnel face advances, they create a spatial "catch-up" relationship between adjacent construction sections. Therefore, data acquisition needs to specifically capture the characteristics of cyclical operations, including: transition time between procedures (such as the waiting time for ventilation and smoke dispersal after blasting, and the connection delay between invert construction and initial support), cumulative impact (such as the impact of continuous cycle on equipment wear, personnel fatigue, or gradual changes in geological conditions on the project schedule), and inter-section dependencies (such as the first...). Only after the initial support of a certain section is completed can the next section be allowed to proceed. Drilling operations began in the section to ensure the safety and stability of the working face.

[0032] To address noise and missing data in the original data, a systematic preprocessing procedure is performed: an outlier detection algorithm is used to remove abnormal data from the logs caused by human error or equipment failure; missing resource usage records and process time data are completed using linear interpolation; numerical data such as process time and resource quantity are standardized and mapped to the [0,1] interval to eliminate the influence of different units and improve the stability and convergence speed of model training.

[0033] To transform complex on-site resources into discrete variables that the model can handle, this embodiment implements a "minimum work unit" abstraction strategy: construction resources are encapsulated into the smallest unit with independent operational capabilities. For example, "1 rock drilling rig + 3 operators" is defined as an "excavation resource unit," and "1 wet shotcrete machine + 2 shotcrete workers" is defined as a "support resource unit." The total amount of "excavation resources" on site is set to 2 (i.e., 2 units are allowed to work in parallel), thereby transforming complex resource scheduling into integer programming constraints. At the same time, a resource-process supply and demand matrix is ​​constructed to clarify the legal resource pool for each process. For example, it is set that the "blasting process" can only request resources from the "blasting team," and the allocation of "slag removal team" is strictly prohibited, providing static constraints for the effective screening of the subsequent action space.

[0034] Based on historical construction logs, patterns in project duration distribution were identified. Statistical analysis methods were used to determine the baseline value and random fluctuation variance of the time consumed by each process. The standard time distribution of each process was then set as a normal distribution function. (In the formula, For process Baseline time consumption To reflect the dynamic characteristics of the construction period being affected by factors such as geological conditions and personnel status during actual construction, a compatibility mapping table between work processes and resource types is constructed to clarify the pool of compliant resources that can be selected for each work process during execution. For example, the tunnel excavation work process can only be allocated to work teams and excavation machinery groups with blasting qualifications to ensure the rationality of resource allocation.

[0035] Finally, process topology dependencies between procedures are extracted through association analysis and topology modeling. Based on the process design in the construction plan and the actual execution order in historical construction logs, a process dependency matrix is ​​constructed. , matrix elements Identify prerequisite dependencies between processes: if processes Must be in the process It can only be started after completion. =1; conversely, It is 0.

[0036] Considering the linear extension characteristics of tunnels, this embodiment employs a two-dimensional modeling logic of "longitudinal segmentation - transverse process" to extract the process topology dependencies: the entire tunnel length is discretized along the axial direction into several standard construction segments, denoted as... , , ..., Within each construction section, a standard lateral sequence of procedures is defined, including drilling operations → charging and blasting → ventilation and smoke dispersal → muck removal and transportation → initial support → invert construction → secondary lining; a procedure dependency matrix is ​​constructed based on the above logic. (dimension is) ,in For the number of segments, (The number of procedures in a construction section cycle) The matrix precisely defines two types of procedure dependency constraints, including (1) intra-section process constraints, that is, within the same section, explosives must be loaded and blasted before slag removal and transportation. (2) Inter-segment safety constraints, i.e., between adjacent segments, the first segment... The initial support of the segment must be in the first stage The drilling operation of the segment is completed before it begins (step advance constraint), and this matrix needs to take into account the characteristics of the process cycle operation.

[0037] Based on the multi-objective requirements of minimizing the construction period and maximizing resource utilization, a mathematical model for the scheduling problem is established, clarifying the core optimization objectives and constraints. The inputs for this step are the original historical logs, standard quotas, and construction plan data, and the outputs are a process attribute table (including process identifier, standard time distribution, and resource requirements), a resource attribute table (including resource identifier, type, and total quantity), and a process dependency matrix, which serve as the input data for step S2.

[0038] S2: Based on the historical data of standardized tunnel construction procedures, a heterogeneous extraction graph of historical data containing procedure nodes and resource nodes is dynamically constructed, and dynamic feature vectors are assigned to procedure nodes and resource nodes.

[0039] A heterogeneous disjunctive graph containing sets of process nodes and resource nodes is constructed, and two types of core topological connections are established: First, unidirectional temporal dependency edges constructed based on the process dependency matrix, pointing from the preceding process to the subsequent process. These edges not only represent the flow of construction technology but also form the topological foundation for calculating the critical path. Considering the segmented characteristics of the tunnel, for example, step-advance constraint edges are added between adjacent construction sections. Second, bidirectional supply-demand interaction edges constructed based on the resource compatibility matrix, connecting process nodes with their corresponding resource nodes. These serve as information bridges representing the process's resource occupancy requests and resource supply status. Based on this topology, the system employs a cascading strategy to assign dynamic feature vectors to nodes. The process node features consist of normalized remaining construction period, binary labels for completion status, and binary labels for execution status. Normalization effectively eliminates gradient instability caused by differences in construction period magnitudes, while the execution status label provides a direct basis for subsequent action masks. For example, the excavation process node features in the initial state are... Resource node characteristics consist of current availability, total capacity, and real-time load rate. The real-time load rate directly maps to the degree of resource congestion in the environment, enabling the model to perceive resource bottlenecks. For example, the initial value of the excavation team node is... .

[0040] S3: Based on graph attention network, graph convolutional network and dual-depth Q network, an intelligent scheduling and optimization model for tunnel construction procedures is constructed, and the intelligent scheduling and optimization model for tunnel construction procedures is trained by heterogeneous extraction graph of historical data.

[0041] In order to extract high-dimensional semantic information for decision-making from the above heterogeneous graph, this invention designs a serially coupled hybrid coding architecture and a global pooling mechanism.

[0042] The heterogeneous disjunctive graph is input into the graph attention network layer. The multi-head attention mechanism is used to compute the mutual influence between process nodes in parallel in different representation subspaces. The weight coefficients between nodes are automatically learned based on the process dependency relationship. Process nodes located on the critical path are given higher attention weights, which strengthens the feature representation of the critical link and suppresses noise interference in non-critical paths.

[0043] The formula for calculating attention weights is as follows: In the formula, A learnable parameter vector; For GAT number Layer weight matrix; and They are nodes and nodes In GAT The feature vector of the layer; Indicates that it will pass through Transformed nodes and nodes The feature vectors are concatenated along a specific dimension; For nodes The set of neighboring nodes.

[0044] The weighted features output by GAT are input into the graph convolutional network layer. Combined with the graph topology, "isotropic" neighborhood feature aggregation is performed. The message transmission between resource flow and process flow is realized through bidirectional supply and demand interaction edges. The scarcity and load distribution information of resource nodes are deeply integrated into the embedding vector of process nodes, so that the process features not only include their own progress status, but also implicitly contain the difficulty of obtaining the required resources.

[0045] Each process node The feature update formula is: In the formula, For neighboring process nodes For the current process node The importance of information transmission; It serves a normalization function; and They are nodes and nodes The degree; Represents the processes in the process dependency matrix With process Relational dependency; For neighboring process nodes In GCN The feature vector of the layer; For GCN Layer weight matrix; This is the activation function.

[0046] To enhance the adaptability of the graph structure to dynamic construction environments and the model's generalization ability, the aforementioned heterogeneous disjunctive graph adopts an event-driven dynamic update mechanism. A schematic diagram of the constructed process-resource heterogeneous disjunctive graph model is shown below. Figure 3 As shown, each time the simulation environment is simulated, the system updates the remaining construction period and available quantity values ​​in the node feature vectors in real time based on the real-time progress feedback and resource status changes at the construction site, and dynamically adjusts the connection state of dependent edges to ensure that the graph structure is always a real-time digital twin of the physical site. Simultaneously, Dropout regularization is introduced during feature extraction to prevent the model from overfitting to noise in specific tunnel cases by randomly blocking the connections of some neurons, thereby significantly improving the robustness of the scheduling strategy under different geological environments and sudden disturbances.

[0047] The global pooling mechanism is used to aggregate the features of all encoded nodes. The permutation invariance of the graph neural network is used to compress and map the variable-length graph data with dynamic node size into a high-dimensional feature vector with constant dimension that represents the global state. For example, a 256-dimensional vector is output, which comprehensively covers the process execution status, resource supply and demand, and process dependencies.

[0048] Furthermore, a dual deep Q-network layer is used to transform the high-dimensional feature vector into optimal scheduling actions, achieving multi-objective optimization that minimizes project duration and maximizes resource utilization. The input to the dual deep Q-network layer is a high-dimensional feature vector. Complete the core architecture design of the model and establish the core architecture of the model. With this as its core state space, the vector deeply integrates multi-dimensional information such as process state, resource load, and topology dependency, ensuring that the agent can fully perceive the construction environment. It constructs a discrete action space consisting of "process execution instructions" and "system wait instructions," where each execution instruction corresponds to a combination of "selecting a specific process + allocating suitable resources" (encoded as a two-dimensional vector). ,in For process index, (For resource indexing), ensuring a high degree of consistency between action coding and the logical mapping of the actual scheduling scenario. Based on this, to address the low search efficiency and ineffective decision-making issues caused by a large-scale action space, this step introduces an action validity screening mechanism based on process topology and resource status. At each decision simulation step, the system traverses all process tasks and executes a strict three-level filtering logic: first, based on the process status identifier, completed or ongoing processes are eliminated to prevent duplicate scheduling; second, based on the process dependency matrix... Remove processes where prerequisites are not met (i.e., those with unmet dependencies). It is 1 and the preceding process (Incomplete), ensuring process logic compliance; finally, comparing real-time resource availability with process requirements, eliminating processes lacking necessary manpower or mechanical resources, ensuring physical resource compliance. The system merges the remaining candidate instructions and waiting instructions after the above filtering to construct a subset of valid actions at the current moment. Furthermore, by using masking technology to force the agent to sample actions only in this subset, physical pruning of the invalid search space is achieved, ensuring that the output instructions always meet the physical and logical constraints of tunnel construction.

[0049] Building upon this, an action validity screening mechanism based on process topology and resource status is introduced: At each decision step, all process tasks are traversed and a three-level filtering is performed. First, completed or in-process processes are eliminated based on process status identifiers. Second, the process dependency matrix is ​​used to filter out completed or in-process processes. Remove processes where prerequisites are not met (i.e., those with unmet dependencies). It is 1 and the preceding process For incomplete tasks (such as those with inter-segment step constraints), the available resources are compared with the requirements, and the insufficient resources are eliminated by referring to the compatibility matrix; a subset of valid actions is generated, and a mask is used to force the agent to sample only in the subset to ensure that the output instructions meet the physical and logical constraints.

[0050] like Figures 4-9 As shown, taking a tunnel construction project with simple process dependencies and resource allocation relationships as an example, the scheduling decision-making process at key time points is illustrated. The execution of this process scheduling process is divided into 6 key time points, assuming... This is the current time point, with time starting from 0 and the unit being days. Below each process node is a similar [image / description]. The 'd' annotation indicates the progress of this process at the current point in time. The integer represents the number of days that the process has been running. This represents the time required to perform the procedure, expressed as a fraction (the ratio has no unit; 'd' in the annotation refers to the denominator). (The unit is "day"). For example, such as Figure 5 As shown, the execution process The required time is 3 days. On the fourth day, this process is just completed, while the adjacent process... Since no constraints have been violated, the process will proceed normally, with a completion rate of 0 / 3d (on the fourth day, the process...). The execution time is less than 1 day. (Integer, rounded down to 0).

[0051] like Figure 4 As shown, when At that time, the process Process Process Execution has commenced and all tasks have been completed for 0 days.

[0052] like Figure 5 As shown, when At that time, the process Once completed, the process... Ready to execute, but due to process Resources needed and ,resource Positive process Occupied, therefore process In a blocked state, resources and Pre-allocated .

[0053] like Figure 6 As shown, when At that time, the process Once completed, the process... The process transitions from the blocked state to the ready state, and there are no other constraints blocking the process at this time. If it runs, then the process It is currently in operation; for the process process Once completed, it should continue, but due to its resource requirements... Already Time pre-allocation to process Therefore, the process Block, and Pre-allocated to process .

[0054] like Figure 7 As shown, when At that time, the process Once completed, the process... The process transitions from the blocked state to the ready state, and there are no other constraints blocking the process at this time. If it runs, then the process It is currently in operation.

[0055] like Figure 8 As shown, when At that time, the process Once the process is completed, that is, the procedure is finished. Once all dependent process routes have been executed, the scheduler instance will now only have the process remaining. and The two processes were not completed, with completion progress of 0 / 2d and 2 / 4d respectively.

[0056] like Figure 9 As shown, when At that time, the process and Once both processes are completed, all processes in this scheduling instance are completed, and the required duration for this project is 12 days.

[0057] Furthermore, a standardized reward function is designed to drive training. This function consists of a total duration reward, a resource utilization reward, and constraint penalties. The total duration reward is calculated based on the deviation ratio between the actual total duration and the reference duration, and is used to guide overall efficiency optimization. The resource utilization reward includes incentive allocation based on the number of parallel operations, overload prevention penalties, and waiting penalties, which are used to improve resource parallelism. The constraint penalties target behaviors that violate dependencies, resource conflicts, and capacity limits, such as weight parameters. , .

[0058] Normalized reward functions include: The total construction period bonus is calculated based on the ratio of the actual total construction period after all work processes are completed to the reference construction period. It is used to guide the model to optimize overall construction efficiency. The formula for the total construction period bonus is as follows: in, The total project duration reward function; The project duration after all processes are completed. This serves as a reference lead time for the process. The weighting coefficient for the total project duration bonus; Resource utilization rewards include: incentive allocation rewards based on the number of currently executing processes to improve parallelism; overload prevention penalties for resource usage exceeding total capacity; and fixed negative waiting penalties imposed on waiting actions. The formula for resource utilization rewards is expressed as: in, For resource utilization reward function, , and The definitions are as follows: in, To incentivize the allocation of reward values; Assign basic weights to incentives; Weighting of rewards based on the number of work processes in the construction project; This indicates the number of processes currently in progress; in, This is the penalty value for overload prevention. Weighting for overload penalties; For resource type serial number; The number of resource types; in, The penalty value is the value to wait for. For waiting penalty weight; Constraint penalty terms are used to penalize scheduling behaviors that violate process dependencies, resource conflicts, and resource capacity limits; the formula for constraint penalty terms is expressed as: in, Penalties for violations of process dependencies; Penalties for violating resource capacity limits; Penalties for resource conflicts; Combining the above three reward mechanisms, the final reward function is as follows: In the formula, Bonus based on total project duration; Incentives for resource utilization; To constrain penalty items.

[0059] The objective function of the intelligent scheduling optimization model for tunnel construction procedures is expressed as follows: in, and The weight coefficients of the objective function represent the importance of total project duration and resource utilization rate in the model optimization, respectively, satisfying the following conditions: ; Total construction period; To serve as a reference for the total construction period; based on historical experience and theoretical analysis, the average construction period is selected as a reference basis to provide a benchmark for the normalization of the total construction period, so that the total construction period of tunnel construction projects of different scales or complexities can be compared and analyzed on a unified scale. This represents the average utilization rate of all resources.

[0060] The model needs to satisfy multiple constraints: (1) Process dependency constraint: In the formula, For process The start time; For process End time; This is the process dependency matrix.

[0061] (2) Resource capacity constraints: In the formula, Indicates a process at a certain moment Resources Usage status; For the first The duration of each process; For resources Total available quantity; For process The start time.

[0062] (3) Scheduling time constraints: In the formula, For process End time; For process Start time; For the first The duration of each process.

[0063] (4) Compatibility constraints between processes and resources: In the formula, for Time process Resources Usage status.

[0064] (5) Resource conflict constraints: In the formula, Number of processes; for Time process Resources Usage status.

[0065] (6) Total project duration constraint: In the formula, The total construction period; For process The end time.

[0066] (7) Resource utilization maximization constraint: In the formula, The average utilization rate of all resources; for Time process Resources Usage status; For the first The duration of each process; For resources Total available quantity; The total construction period; In the formula, the process dependency constraint requires the process to... start time Must be in dependent processes Execution can only begin after completion. Dependencies are established through the process dependency matrix. To indicate, if the process Must be in the process It can only begin after completion. =1, otherwise =0; Resource capacity constraints ensure that the usage of each type of resource at any given time will not exceed its maximum capacity. Therefore, during the scheduling process, the resource requirements of all processes must meet the resource capacity limits; the time constraints of process scheduling define the completion time of the processes. It equals the start time of the process. In addition to its duration The compatibility constraints between processes and resources stipulate that each process can only be executed on a specific resource, and each process can only be assigned to one resource; resource conflict constraints ensure that at any given time... A certain resource It can only be used by one process. This avoids resource conflicts; the total project duration constraint defines the total project duration. It is the process of all steps Completion Time The maximum value in the scheduling is the maximum value, and the goal of scheduling is to minimize the total project duration. The resource utilization maximization constraint calculates the resource utilization rate, and the goal is to maximize the efficiency of resource use by optimizing the scheduling scheme, while keeping the resource usage within the available range.

[0067] The model training employs an experience replay and a dual-network architecture to ensure training stability and convergence. An initial experience replay pool stores four-tuple samples (state, action, reward, next state), and random batch sampling is used to break temporal correlations. During iteration, [the following is used]. The strategy dynamically balances exploration and utilization, with the exploration rate gradually decreasing as the training rounds increase. This transitions from random trial and error to strategy optimization. Simultaneously, a dual deep Q-network algorithm is introduced to decouple action selection and value evaluation, suppressing... Value overestimation; based on mean squared error loss function The Adam optimizer iteratively updates the evaluation network parameters, continuously optimizing the evaluation network to achieve the desired results. Gradually approaching the target value The goal is to enable the agent to learn the best possible strategy. Exploration rate The value decreases from 0.9 to 0.01, and after a fixed number of training epochs until the model converges, the principle diagram of the dual-depth Q-network algorithm is as follows. Figure 10 As shown.

[0068] The system calculates the prediction error based on the mean squared error loss function, and uses the Adam optimizer to minimize the loss function using the stochastic gradient descent algorithm, driving the evaluation network parameter update, and periodically soft-synchronizing the parameters to the target network until the training rounds reach a preset threshold or the cumulative reward fluctuations converge, ultimately generating an intelligent scheduling strategy with strong generalization ability.

[0069] S4: Analyze the progress and resource usage snapshots of the construction site to obtain the physical site status. Synchronously map the physical site status to the virtual environment to construct a digital twin of the tunnel construction project. Input the synchronized status into the trained intelligent scheduling and optimization model for tunnel construction procedures for real-time reasoning. The model outputs the optimal procedure and resource scheduling instructions at the current moment, as well as the construction schedule plan for the subsequent tunnel project.

[0070] The process of building a digital twin and real-time reasoning is essentially about establishing a two-way closed-loop control system between the physical construction site and the virtual decision-making environment. Specifically, it includes three core components: logical mapping based on on-site data snapshots, dynamic rescheduling reasoning, and online continuous optimization. First, this invention uses a multi-source sensing network (including IoT sensors, video surveillance, and construction management log systems) deployed at the construction site to collect real-time snapshots of engineering data and performs a rigorous "state injection" operation to achieve virtual-real synchronization. The system analyzes the real-time status (not started / in execution / suspended / completed) and resource occupancy records of each process in the snapshot, mapping the real state of the physical world back to the virtual simulation environment: for processes marked as "in progress," its occupancy status is forcibly locked in the virtual node, and the remaining construction period is corrected to the difference between the standard construction period baseline value and the executed time. At the same time, the corresponding occupied amount is deducted from the virtual resource pool (i.e., virtual available amount = total capacity - actual on-site occupancy). Finally, the simulation system clock is aligned to the timestamp of the snapshot generation, thereby eliminating the "state deviation" between the virtual environment and the physical site and ensuring that subsequent inference is based on real boundary conditions. Based on a synchronized high-fidelity virtual environment, the system performs dynamic rescheduling inference. The updated heterogeneous graph is input into the trained intelligent scheduling optimization model for tunnel construction procedures, generating a global state representation vector for the current moment. The model quickly outputs the optimal scheduling action (i.e., a combination of procedure assignment and resource allocation instructions) through single-step forward propagation. Building upon this, the system further utilizes the simulation environment's extrapolation capabilities to continuously simulate subsequent construction processes based on the current strategy until project completion, generating a dynamically adjusted Gantt chart containing the estimated start and end times of all unfinished procedures and resource load curves. This mechanism enables the scheduling system to not only provide instructions on "what to do now," but also a global predictive view of "what will happen in the future," assisting managers in making visual decisions. The digital twin system possesses exceptional resilience and continuous evolution capabilities. When unexpected disturbances occur at the construction site, such as geological shifts or equipment failures, the system triggers emergency updates via an event-driven mechanism. Without retraining the model, it simply regenerates the state vector based on the latest snapshot of the abnormal state and performs inference. Within milliseconds, it outputs an adjustment plan adapted to the new constraints, achieving dynamic rescheduling. Simultaneously, the system establishes a closed-loop "data-model" feedback mechanism, periodically collecting actual execution data from the site (including actual schedule deviations and resource consumption efficiency) as incremental training samples. This data is stored in an experience replay pool to fine-tune model parameters, enabling the agent to continuously adapt to specific geological environments and construction rhythms as the project progresses, achieving lifelong learning and continuous optimization of the scheduling strategy.

[0071] The method of this invention effectively integrates process dependencies and resource allocation relationships by constructing a heterogeneous disjunctive graph model of processes and resources; it combines GAT and GCN to achieve dynamic perception of critical path features and efficient expression of global state; and it further introduces the DDQN reinforcement learning mechanism to achieve multi-objective optimization and adaptive updating of scheduling strategies, thus realizing intelligent scheduling in complex construction environments. This method not only effectively balances the demands of high construction efficiency and maximized resource utilization in construction projects, but also adapts to dynamically changing construction sites, ensuring the real-time performance and executability of scheduling strategies.

[0072] Based on the aforementioned methodological framework, the following verification is conducted on common sudden disturbance scenarios in tunnel construction sites, including resource occupation conflicts and construction period delays, demonstrating that the present invention has high adaptive response capability and robustness.

[0073] This example uses a mountain tunnel project (total length 5.2km, construction section) =5), the structured data collected includes seven standard horizontal process sequences: drilling (baseline time 5.0 hours), charging and blasting (2.0 hours), ventilation and smoke dispersal (1.0 hour), muck removal and transportation (5.0 hours), initial support (6.0 hours), invert construction (8.0 hours), and secondary lining (12.0 hours). Resources are categorized into human resources (including excavation teams, support teams, secondary lining teams, and general labor teams, with a total available quantity of 1) and mechanical resources (including rock drilling rigs, excavators, wet spraying machines, and lining rigs, with a total available quantity of 1). Resource allocation to processes is based on a resource compatibility matrix definition. For example, drilling operations require allocation of human resources to excavation teams and general labor teams, as well as rock drilling rig mechanical resources; muck removal and transportation require allocation of human resources to excavation teams and general labor teams, as well as... Including excavator machinery resources, the initial support requires the allocation of support team human resources and wet spraying machine machinery resources. The invert construction requires the allocation of excavation team, support team, secondary lining team, and miscellaneous worker team human resources, as well as rock drilling rigs, excavators, wet spraying machines, and lining trolley machinery resources. The secondary lining requires the allocation of secondary lining team human resources and lining trolley machinery resources. The miscellaneous worker team, as a flexible resource, can assist multiple processes and replace some of the needs of professional teams, realizing dynamic reuse of resources. In this example, the allocated machinery resources are the main resources. There are other resources in the actual construction project, but they are not within the main considerations affecting the scheduling decision. The process dependency relationship is reflected by the process dependency matrix A (dimension is (5×7)×(5×7)=35×35). The matrix part example is shown in Table 1 (only the dependency submatrix of segment 1-segment 2 is shown here). Indicate process Dependent process ).

[0074] Table 1. Example of Part A of the Process Dependency Matrix (Segment 1-Segment 2 Submatrices) When the project simulation reached 48.5 hours, resource conflicts and schedule delays occurred: Section 1 (invert arch construction) was in operation, occupying excavators and other machinery resources, with 6.0 hours remaining (2 hours slower than the expected baseline); Section 2 (muck removal and transportation) was on standby, waiting for excavators to be released; drilling rigs and lining rigs were currently idle; Section 1 (initial support) had just been completed, and Section 1 (secondary lining) was awaiting completion of Section 1 (invert arch); other early-stage procedures were completed. This disturbance caused downstream process blockages, and simulated geological abrupt changes led to a shift in the time distribution of the procedures.

[0075] Based on the digital twin construction mechanism in step S4, the disturbance information is processed synchronously: a state injection operation is performed to update the resource node characteristics in the virtual environment, such as updating the excavator load rate to full load (occupied by the construction of section 1_invert arch), and setting the rock drilling rig and lining rig to idle state; combining the on-site snapshot data and the standard time distribution of the process fitted in step S1, the remaining time of section 1_invert arch construction is corrected, and the probabilistic impact of geological changes on subsequent processes (such as section 1_secondary lining) is considered; the process dependency matrix is ​​adjusted. The values ​​in the middle are used to strengthen the step distance advance constraint between segment 1 and segment 2 (e.g., segment 1_invert arch must precede segment 2_slag removal); the heterogeneous disjunctive graph is dynamically reconstructed through an event-driven mechanism, and the GAT-GCN encoder input in step S2 generates a new global state vector. This ensures that the virtual model accurately captures resource usage and changes in process status.

[0076] The updated Input the trained intelligent scheduling optimization model for tunnel construction procedures and perform a single-step forward propagation. The model outputs the adjusted optimal scheduling action: prioritize releasing the excavator and assigning it to the section 2_muck removal and transportation procedure (requiring excavation team, general labor team manpower, and excavator machinery resources); utilize the idle rock drilling rig to start the section 3_drilling operation (requiring excavation team, general labor team manpower, and rock drilling rig machinery resources); postpone the section 1_secondary lining procedure (requiring secondary lining team manpower and lining rig machinery resources) until resources are restored, and pre-allocate the lining rig to the subsequent secondary lining procedure; generate a complete scheduling plan through continuous simulation, including the expected start and end times of all unfinished procedures, the segmented load distribution, and resource allocation curves.

[0077] If the intelligent decision-making mechanism of this invention is not adopted, and instead traditional manual experience or static CPM methods are relied upon, uneven resource allocation and process blockage may be further amplified in disturbed scenarios. For example, if the occupancy of excavators is not adjusted in time, downstream processes (such as section 2_slag removal and transportation) will be delayed for a long time, and the overall construction period may be extended by more than 20%. At the same time, the resource utilization rate will decrease because idle resources (such as rock drilling rigs) are not effectively mobilized to processes that can be executed in parallel (such as section 3_drilling operation, which depends on section 2_ventilation and smoke dissipation and whose resource requirements are compatible with rock drilling rigs).

[0078] Furthermore, in a simulation environment, various comparative and ablation experiments were conducted on datasets of different sizes, and the decision results are shown in Table 2: Table 2 Among them, the Heuristic-EAS method is based on a heuristic scheduling algorithm (earliest start first algorithm). It combines the process dependencies, resource constraints and dynamic scheduling requirements of tunnel construction, improves the critical path analysis and resource allocation mechanism, and introduces task priority ranking based on relaxation time and event-driven time advancement strategy, making the scheduling more adaptable to complex construction scenarios. It serves as a benchmark experiment.

[0079] The experiment used the Heuristic-EAS algorithm as a baseline (resource utilization rate of 1), and analyzed the project duration and normalized resource utilization rate obtained from the algorithm's decisions in each experimental group. Evaluate the performance of each algorithm. The formula for calculating the normalized resource utilization rate of each algorithm is as follows: in, The resource utilization evaluation index (normalized resource utilization rate) is used for each algorithm. The average resource utilization rate of the algorithm being evaluated. This represents the average resource utilization rate of the benchmark algorithm.

[0080] By comparing the project duration and normalized resource utilization of five methods—Heuristic-EAS, DDQN, GCN-DDQN, GAT-GCN-DQN, and GAT-GCN-DDQN—on datasets of different sizes (5×10, 6×6, 10×20, 20×15, 20×40, and 30×20, with the number of work processes multiplied by the number of resource types), the superior performance of the GAT-GCN-DDQN method in tunnel construction work process scheduling was fully verified. In the experimental group with the 5×10 dataset, the project duration obtained by the GAT-GCN-DDQN decision was 152.40 hours, which was 33.16% shorter than the baseline experiment of 228.00 hours achieved by Heuristic-EAS; in the 30×20 dataset, the duration was 569.24 hours, which was 26.11% shorter than the baseline experiment of 770.40 hours. Furthermore, the normalized resource utilization rate obtained by the GAT-GCN-DDQN decision meets the expected requirements, reaching 1.76 and 1.65 in the experimental groups of the 10×20 and 20×15 datasets, respectively, which is better than other methods.

[0081] In tunnel construction scheduling decisions, Heuristic-EAS is limited by static rules designed based on human experience, resulting in a lack of scheduling flexibility. DDQN, through reinforcement learning, optimizes decisions based on agents guided by multi-objective reward functions, but it lacks sufficient modeling of the state and actions in complex scheduling environments, leading to insufficient target decisions regarding resource utilization, such as a score of only 1.03 on the 20×40 dataset. GCN-DDQN introduces graph convolutional networks to aggregate global topological features, improving multi-objective decisions regarding project duration and resource utilization, achieving a project duration of 187.20 hours and a resource utilization rate of 1.49 on the 10×20 dataset. GAT-GCN-DQN, combined with GAT dynamic weight allocation, enhances the model's graph feature extraction capabilities, reducing the project duration to 720.00 hours in the 6×6 dataset experimental group. However, DQN tends to overestimate the Q-value and suffers from training instability, resulting in limited optimization effects. GAT-GCN-DDQN integrates GAT, GCN, and DDQN. The agent, guided by a carefully designed multi-objective reward function, combines GAT's dynamic weight allocation and GCN's global feature aggregation, with the DDQN algorithm leading to optimized decision-making. This ultimately results in a scheduling decision that meets the requirements of minimizing project duration and maximizing resource utilization. Experimental results demonstrate that GAT-GCN-DDQN performs exceptionally well in complex process dependencies and multi-objective optimization, providing an efficient, stable, and highly adaptable solution for intelligent scheduling of tunnel construction processes.

[0082] like Figure 11 As shown, the present invention also provides an intelligent scheduling and optimization device for tunnel construction procedures based on GAT-GCN-DDQN, comprising: The data acquisition module 210 is used to acquire historical data of tunnel construction procedures, perform structured processing, extract procedure attributes, resource attributes, and process topology dependencies of procedure components, and obtain standardized historical data of tunnel construction procedures. Data processing module 220 is used to dynamically construct a heterogeneous extraction graph of historical data containing process nodes and resource nodes based on standardized tunnel construction process historical data, and to assign dynamic feature vectors to process nodes and resource nodes. The model building and training module 230 is used to build an intelligent scheduling and optimization model for tunnel construction procedures based on graph attention network, graph convolutional network and dual depth Q network, and to train the intelligent scheduling and optimization model for tunnel construction procedures by extracting graphs from historical data heterogeneously. The intelligent scheduling optimization module 240 is used to analyze the progress and resource occupancy snapshots of the construction site to obtain the physical site status. The physical site status is synchronously mapped to the virtual environment to construct a digital twin of the tunnel construction project. The synchronous status is input into the trained intelligent scheduling optimization model for tunnel construction procedures for real-time reasoning. The model outputs the optimal procedure and resource scheduling instructions at the current moment and the construction schedule plan for subsequent tunnel projects.

[0083] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A GAT-GCN-DDQN-based tunnel construction process intelligent scheduling optimization method, characterized in that, include: Historical data on tunnel construction procedures were acquired and structured to extract procedure attributes, resource attributes, and process topology dependencies of procedure components, resulting in standardized historical data on tunnel construction procedures. Based on the historical data of the standardized tunnel construction procedures, a heterogeneous extraction graph of historical data containing procedure nodes and resource nodes is dynamically constructed, and dynamic feature vectors are assigned to the procedure nodes and the resource nodes. A smart scheduling and optimization model for tunnel construction procedures is constructed based on graph attention network, graph convolutional network and dual depth Q network, and the model is trained by extracting graphs from the historical data. The construction site's progress and resource usage snapshots are analyzed to obtain the physical site status. This physical site status is then synchronously mapped to a virtual environment to construct a digital twin of the tunnel construction project. The synchronous status is input into a trained intelligent scheduling and optimization model for tunnel construction procedures for real-time reasoning. The model outputs the optimal procedure and resource scheduling instructions for the current moment, as well as the construction schedule plan for subsequent tunnel projects.

2. The GAT-GCN-DDQN-based tunnel construction process intelligent scheduling optimization method according to claim 1, characterized in that, In the steps of acquiring historical data on tunnel construction procedures and performing structured processing to extract procedure and resource attributes, the extraction of resource and procedure attributes is the process of abstracting physical construction elements into the smallest computable unit, including: Construction resources are abstracted into complete work teams or sets of machinery with independent operating capabilities, and the maximum parallel operation capacity of the smallest work unit on the construction site is defined as the total resource, which is used to constrain the cumulative usage of the resource by all processes; among them, construction elements include extracted process attributes and resource attributes; construction resources include human resources and physical construction resources; Based on the historical construction logs, the distribution patterns of construction period are mined, and the standard time distribution of each process is set as a probability distribution function that includes a baseline value and a random fluctuation range. A compatibility mapping table between processes and resource types is constructed to clarify the pool of compliant resource candidates that can be used when each process is executed.

3. The GAT-GCN-DDQN-based tunnel construction process intelligent scheduling optimization method according to claim 1, characterized in that, Based on the standardized tunnel construction procedure historical data, a heterogeneous historical data extraction graph containing procedure nodes and resource nodes is dynamically constructed, and dynamic feature vectors are assigned to the procedure nodes and resource nodes, including: Construct a set of process nodes and a set of resource nodes and establish a connection relationship, including a unidirectional temporal dependency edge to characterize the flow of construction technology and a bidirectional supply and demand interaction edge to characterize the relationship between process and resource occupation and release; Construct a dynamic feature vector for process nodes, wherein the dynamic feature vector for process nodes includes a normalized remaining duration value, a binary identifier representing whether the process is completed, and a concatenated binary identifier representing whether the process is currently consuming resources. Construct a dynamic feature vector for resource nodes, which includes a concatenation of the current available amount of resources, the total capacity of resources, and the load rate value representing the real-time occupancy ratio of resources.

4. The intelligent scheduling and optimization method for tunnel construction procedures based on GAT-GCN-DDQN according to claim 1, characterized in that, The intelligent scheduling and optimization model for tunnel construction procedures is trained using the heterogeneous extraction graph of the historical data, including: The heterogeneous extraction graph of the historical data is input into the graph attention network layer of the intelligent scheduling and optimization model of tunnel construction process. The multi-head attention mechanism is used to calculate the interaction weight between nodes, and high weight coefficients are assigned to process nodes located on the critical path according to the process dependency relationship. The weighted process node features are input into the graph convolutional network layer, and neighborhood feature aggregation is performed in combination with the graph topology, thus integrating the supply and demand distribution information of resource nodes into the feature representation of process nodes. The global pooling mechanism is used to aggregate the features of all encoded process nodes, mapping the variable-length graph data with dynamic node size into a high-dimensional feature vector with constant dimension representing the global state. The high-dimensional feature vector is input into the dual deep Q network layer and mapped to a discrete action space containing process execution and waiting. The design incorporates a standardized reward mechanism that integrates total project duration optimization rewards, resource utilization feedback, and multiple constraint penalties. The network parameters of the intelligent scheduling optimization model for tunnel construction procedures are updated iteratively through experience playback and gradient descent, and the model is trained until the strategy converges.

5. The intelligent scheduling and optimization method for tunnel construction procedures based on GAT-GCN-DDQN according to claim 4, characterized in that, The high-dimensional feature vector is input into the dual deep Q-network layer and mapped to a discrete action space containing process execution and waiting, including: At each decision step, all process tasks are traversed, and processes that have been completed or are in progress are eliminated based on the current status of the process, processes that have not met their prerequisites are eliminated based on the process dependency matrix, and processes that lack the necessary human or mechanical resources are eliminated based on the real-time availability of resources. Generate a subset of valid actions containing only verified process assignment instructions and waiting instructions, and select actions from this subset to ensure that the output scheduling instructions always meet the physical and logical constraints of tunnel construction.

6. The intelligent scheduling and optimization method for tunnel construction procedures based on GAT-GCN-DDQN according to claim 4, characterized in that, The design incorporates a standardized reward mechanism that integrates total project duration optimization rewards, resource utilization feedback, and multiple constraint penalties, including: Construct a standardized reward function, which includes a total project duration reward, a resource utilization reward, and a constraint penalty term; The quadruple, which includes the current global features of the heterogeneous graph, the filtered effective actions, the dense step reward value, and the global features of the heterogeneous graph at the next moment, is stored in the experience replay pool. The target Q-value based on intensive step reward is calculated using a dual Q-learning mechanism, and the prediction error is minimized through backpropagation algorithm, driving the network parameters to converge toward the Pareto optimal boundary that minimizes the project duration and maximizes resource utilization.

7. The intelligent scheduling and optimization method for tunnel construction procedures based on GAT-GCN-DDQN according to claim 6, characterized in that, The normalized reward function includes: The total construction period bonus is calculated based on the ratio of the actual total construction period after all work processes are completed to the reference construction period. It is used to guide the model to optimize overall construction efficiency. The formula for the total construction period bonus is as follows: in, The total project duration reward function; The project duration after all processes are completed. This serves as a reference lead time for the process. The weighting coefficient for the total project duration bonus; Resource utilization rewards include: incentive allocation rewards based on the number of currently executing processes to improve parallel operation; overload prevention penalties for resource usage exceeding total capacity; and fixed negative waiting penalties imposed on waiting actions; the formula for the resource utilization rewards is expressed as: in, For resource utilization reward function, , and The definitions are as follows: in, To incentivize the allocation of reward values; Assign basic weights to incentives; Weighting of rewards based on the number of work processes in the construction project; This indicates the number of processes currently in progress; in, This is the penalty value for overload prevention. Weighting for overload penalties; For resource type serial number; The number of resource types; in, The penalty value is the value to wait for. For waiting penalty weight; The constraint penalty term is used to penalize scheduling behaviors that violate process dependencies, resource conflicts, and resource capacity limits; the formula for the constraint penalty term is expressed as: in, Penalties for violations of process dependencies; Penalties for violating resource capacity limits; Penalties for resource conflicts; Combining the above three reward mechanisms, the final reward function is as follows: In the formula, Bonus based on total project duration; Incentives for resource utilization; To constrain penalty items.

8. The intelligent scheduling and optimization method for tunnel construction procedures based on GAT-GCN-DDQN according to claim 1, characterized in that, The objective function formula of the intelligent scheduling and optimization model for tunnel construction procedures is expressed as follows: in, and The weight coefficients of the objective function represent the importance of total project duration and resource utilization rate in the model optimization, respectively, satisfying the following conditions: ; Total construction period; To serve as a reference for the total construction period; based on historical experience and theoretical analysis, the average construction period is selected as a reference basis to provide a benchmark for the normalization of the total construction period, so that the total construction period of tunnel construction projects of different scales or complexities can be compared and analyzed on a unified scale. This represents the average utilization rate of all resources.

9. The intelligent scheduling and optimization method for tunnel construction procedures based on GAT-GCN-DDQN according to claim 1, characterized in that, Construct a digital twin of the tunnel construction project; input the synchronous state into the trained intelligent scheduling and optimization model of the tunnel construction process for real-time reasoning, and the model outputs the optimal process and resource scheduling instructions at the current moment, as well as the construction schedule plan for the subsequent tunnel project, including: The execution status injection operation parses the structured field snapshot data containing real-time process status, execution time and resource usage records, traverses the process list in the data, marks the process marked as "in progress" as occupied in the virtual environment and corrects the remaining duration, forcibly deducts the corresponding available resources in the virtual resource pool, and advances the simulation clock to the field timestamp. Perform dynamic rescheduling reasoning. Based on the synchronized virtual environment state, use the trained model to generate the process assignment instructions that need to be executed immediately at the current moment, and continuously deduce subsequent strategies until the project ends, generating a dynamically adjusted Gantt chart containing the expected start and end times of all unfinished processes.

10. A tunnel construction process intelligent scheduling and optimization device based on GAT-GCN-DDQN, characterized in that, include: The data acquisition module is used to acquire historical data of tunnel construction procedures, perform structured processing, extract procedure attributes, resource attributes, and process topology dependencies of procedure components, and obtain standardized historical data of tunnel construction procedures. The data processing module is used to dynamically construct a heterogeneous extraction graph of historical data containing process nodes and resource nodes based on the standardized tunnel construction process historical data, and to assign dynamic feature vectors to the process nodes and the resource nodes. The model building and training module is used to build an intelligent scheduling and optimization model for tunnel construction procedures based on graph attention network, graph convolutional network and dual depth Q network, and to train the intelligent scheduling and optimization model for tunnel construction procedures using the heterogeneous extraction graph of the historical data. The intelligent scheduling and optimization module is used to analyze the progress and resource occupancy snapshots of the construction site to obtain the physical site status. The physical site status is synchronously mapped to the virtual environment to construct a digital twin of the tunnel construction project. The synchronous status is input into the trained intelligent scheduling and optimization model of the tunnel construction process for real-time reasoning. The model outputs the optimal process and resource scheduling instructions at the current moment and the construction schedule plan for the subsequent tunnel project.

Images