Space resource dynamic allocation method based on AI model and BIM operation and maintenance

Abstract

The present application belongs to the technical field of resource allocation, and particularly relates to a space resource dynamic allocation method based on an AI model and BIM operation and maintenance, comprising: obtaining a to-be-processed task and a spatial position identifier thereof, and constructing a task feature vector set; inputting the task feature vector set into a modal recognition network model to obtain a plurality of primary scheduling modes containing spatial coupling components and resource demand components through mapping; iteratively updating coupling weights between the primary scheduling modes and high-level scheduling modes until convergence, determining a current resource allocation mode based on the converged high-level scheduling modes, and determining the association strength of each task relative to the current resource allocation mode based on the coupling weights; generating a scheduling instruction set containing joint scheduling paths and independent scheduling instructions according to the association strength and in combination with the spatial position identifier, and converting the scheduling instruction set into executable control instructions; and the present application realizes fine and adaptive dynamic resource allocation based on BIM and real-time situation.

Description

Technical Field

[0001] This invention belongs to the field of resource allocation technology, and in particular relates to a method for dynamic allocation of spatial resources based on AI models and BIM operation and maintenance. Background Technology

[0002] In BIM and IoT-based park energy management systems, to achieve refined control of HVAC, lighting, and power distribution systems, the system needs to rely on real-time operational data collected at the second level (such as fan current, water pipe pressure, and lighting circuit power) and perform linkage analysis with static equipment topology and rated parameters in the BIM model. It then uses optimization algorithms to generate control commands such as fan speed setting and lighting group dimming. However, in actual operation, due to sensor accuracy drift, network packet loss, and asynchronous equipment response, the real-time data stream often contains anomalies such as instantaneous spikes, short-term missing values, and persistent biases. For example, the current data of an air conditioning fan may exhibit periodic spikes due to signal interference, while the rated power parameter of the fan in the BIM model may not be updated with the actual aging state. Existing systems, when calculating energy consumption optimization strategies, typically only perform simple filtering on the raw data. This fails to effectively identify and compensate for such complex data anomalies with equipment-related characteristics, and also fails to embed dynamic characteristics such as the efficiency decay of chiller units over time and the nonlinear relationship between lighting fixture light decay and energy consumption into the optimization objective function. This leads to a situation where the system might calculate theoretically energy-saving fan speed-reduction commands based on distorted fan current data and outdated rated power models. However, in actual execution, because the actual performance of the fan has deteriorated, maintaining the required airflow causes it to operate in an inefficient range for extended periods, triggering frequent adjustments to associated water valves. Ultimately, this results in a decrease in overall system energy efficiency and equipment oscillation. Therefore, how to enable energy optimization algorithms to generate robust, executable, and non-instantaneous energy-saving control strategies under conditions of multi-source heterogeneous data anomalies and unmodeled equipment dynamic characteristics has become a specific technical bottleneck restricting the improvement of the actual efficiency of current systems. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a dynamic spatial resource allocation method based on AI models and BIM-based operation and maintenance. The method includes: acquiring tasks to be processed and their spatial location identifiers, constructing a task feature vector set; inputting the task feature vector set into a modal recognition network model to map multiple primary scheduling modes containing spatial coupling components and resource demand components; iteratively updating the coupling weights between the primary and higher-level scheduling modes until convergence; determining the current resource allocation mode based on the converged higher-level scheduling mode, and determining the correlation strength of each task relative to the current resource allocation mode based on the coupling weights; generating a set of scheduling instructions containing joint scheduling paths and independent scheduling instructions based on the correlation strength and spatial location identifiers, and converting it into executable control instructions. This invention solves the problems of static data, delayed response, and low resource coordination efficiency in traditional operation and maintenance scheduling, achieving refined and adaptive dynamic resource allocation based on BIM and real-time situational awareness.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A method for dynamic allocation of spatial resources based on AI models and BIM operation and maintenance includes:

[0006] Obtain execution feature data of multiple tasks to be processed within the target area and spatial location identifiers corresponding to each task, and construct a task feature vector set based on the execution feature data;

[0007] The task feature vector set is input into the modality recognition network model to map multiple primary scheduling modes; wherein, the spatiotemporal attributes in the task feature vector are encoded as the spatial coupling components of the primary scheduling modes, and the resource requirement attributes in the task feature vector are encoded as the resource requirement components of the primary scheduling modes.

[0008] The coupling weights are iteratively updated between the multiple primary scheduling modes and the higher-level scheduling modes, and convergence is determined when a preset convergence condition is met. The current resource allocation mode of the system is determined based on the converged higher-level scheduling mode, and the correlation strength of each task relative to the current resource allocation mode is determined based on the converged coupling weights.

[0009] Based on the association strength and the spatial location identifier, a set of scheduling instructions is generated, and the set of scheduling instructions is converted into executable control instructions for the corresponding operation and maintenance resources; the set of scheduling instructions includes joint scheduling paths for shared operation and maintenance resources and independent scheduling instructions for independent operation and maintenance resources.

[0010] Specifically, high-level scheduling modes refer to typical scheduling patterns that represent the current globally optimal or most urgent resource coordination methods, abstracted from multi-task characteristics through a mode recognition network; a task feature vector set is constructed based on the execution feature data, including:

[0011] Get multiple pending operation and maintenance work orders, where each pending operation and maintenance work order is associated with at least one original timestamp, a static space identifier and a set of static resource requirement attributes;

[0012] The original timestamp associated with each pending maintenance work order is calibrated to obtain the calibration timestamp of each pending maintenance work order;

[0013] Based on the static spatial identifier associated with each pending maintenance work order, real-time location awareness data is obtained, and the real-time location awareness data is compared with the static spatial identifier to obtain the approved spatial coordinates and location confidence of each pending maintenance work order.

[0014] Based on the static resource requirement attributes of each pending maintenance work order, the real-time availability status of the corresponding resource entity is obtained, and a dynamic resource availability code is generated.

[0015] Specifically, constructing a task feature vector set based on the execution feature data further includes:

[0016] Based on the calibration timestamp and approved spatial coordinates of each pending maintenance work order, the environmental impact coefficient of each pending maintenance work order is calculated using a pre-trained environmental impact assessment model.

[0017] Based on the theoretical estimated working hours in the static resource requirement attributes of each pending maintenance work order, and combined with the environmental impact coefficient, the weighted estimated working hours of each pending maintenance work order are calculated.

[0018] Based on the calibration timestamp, fault level, weighted estimated working hours, and dynamic resource availability code of each pending maintenance work order, a recommended priority score is calculated for each pending maintenance work order according to the preset priority calculation rules.

[0019] The calibration timestamp, approved spatial coordinates, location confidence, dynamic resource availability coding, weighted estimated working hours, and recommended priority score of each pending maintenance work order are integrated to form the maintenance feature vector of the pending maintenance work order, and a task feature vector set is generated.

[0020] Specifically, the mapping yields multiple primary scheduling modes, including:

[0021] Extract the spatiotemporal attribute sub-vector set and the resource requirement attribute sub-vector set from the task feature vector set;

[0022] A spatiotemporal attribute matrix is ​​constructed based on the spatiotemporal attribute sub-vector set. The spatiotemporal attribute matrix is ​​then input into a multi-head self-attention layer, and a weighted spatiotemporal feature matrix is ​​obtained by calculating the spatial correlation between tasks.

[0023] Based on the weighted spatiotemporal feature matrix, a linear transformation is performed through the first fully connected layer to output the spatial coupling component matrix;

[0024] A resource demand attribute matrix is ​​constructed based on the resource demand attribute sub-vector set. The resource demand attribute matrix is ​​then input into a resource demand encoding network consisting of at least one fully connected layer. The matrix is ​​transformed by a nonlinear activation function to output a resource demand component matrix.

[0025] The row vector corresponding to each task in the spatial coupling component matrix is ​​concatenated with the row vector corresponding to the task in the resource demand component matrix to obtain the primary scheduling mode vector of each task.

[0026] The primary scheduling mode vectors of all tasks are aggregated to obtain the multiple primary scheduling modes.

[0027] Specifically, determining the correlation strength of each task relative to the current resource allocation mode includes:

[0028] Initialize a higher-level scheduling mode vector and the coupling weights between each primary scheduling mode and the higher-level scheduling mode vector;

[0029] Obtain the latest real-time dynamic status data for each task, and calculate the dynamic status adjustment factor based on the real-time dynamic status data;

[0030] The corresponding primary scheduling modes are modified element-wise using the dynamic state adjustment factor to obtain the set of dynamically modified primary scheduling modes.

[0031] Based on the dynamically corrected primary scheduling mode set and the higher-level scheduling mode vector, the coupling weights are updated using the Softmax function;

[0032] Based on the updated coupling weights, the dynamically corrected primary scheduling mode set is weighted and summed to update the higher-level scheduling mode vector.

[0033] Specifically, determining the correlation strength of each task relative to the current resource allocation mode further includes:

[0034] Calculate the change between the high-level scheduling modal vector after the current iteration and the high-level scheduling modal vector after the previous iteration. If the change is less than a preset convergence threshold, it is determined to be converged.

[0035] The high-level scheduling mode vector at the time of convergence is determined as the current resource allocation mode of the system.

[0036] Obtain the dependencies between tasks, and construct a task dependency graph based on the dependencies;

[0037] Using the coupling weights at convergence as the initial association scores, and combining them with the task dependency graph, the association strength of each task relative to the current resource allocation mode is calculated using a graph attention network model.

[0038] Specifically, a set of scheduling instructions is generated, including:

[0039] Based on the preset association strength threshold and the association strength of each task relative to the current resource allocation mode, all tasks are divided into a high association task group and a low association task group.

[0040] Based on the approved spatial coordinates of each task in the highly correlated task group, at least one spatial task cluster is obtained through a density-based spatial clustering algorithm.

[0041] Based on the dynamic resource availability coding of tasks within each space task cluster, each space task cluster is divided into at least one shared resource task cluster, wherein tasks within each shared resource task cluster are executed by the same shared operation and maintenance resource.

[0042] Specifically, generating the set of scheduling instructions also includes:

[0043] For each shared resource task cluster, a joint scheduling path for the shared operation and maintenance resource is generated based on the approved spatial coordinates of all tasks within the cluster using a traveling salesman problem algorithm.

[0044] For each task in the low-association task group, an independent scheduling instruction for the independent operation and maintenance resource is generated based on the approved spatial coordinates and dynamic resource availability code of the task.

[0045] All the joint scheduling paths and all the independent scheduling instructions are aggregated to generate the scheduling instruction set.

[0046] Specifically, a joint scheduling path for this shared operational resource is generated using a traveling salesman problem solving algorithm, including:

[0047] A building path network is obtained based on a building information model. The building path network consists of nodes and edges, where nodes correspond to traffic hubs and edges correspond to traffic paths between nodes.

[0048] Map the approved spatial coordinates of each task in the shared resource task cluster to the nearest node in the building path network to obtain the target node corresponding to each task;

[0049] Based on the building path network, the shortest travel distance between every two target nodes is calculated using Dijkstra's algorithm, and a target node distance matrix is ​​constructed.

[0050] Specifically, generating a joint scheduling path for this shared operational resource using a traveling salesman problem solving algorithm also includes:

[0051] Obtain real-time traffic status data for each edge in the building path network, and adjust the traffic weight of the corresponding edge based on the real-time traffic status data;

[0052] Based on the adjusted passage weights, the shortest passage distance between every two target nodes is recalculated, and the target node distance matrix is ​​updated.

[0053] The updated target node distance matrix is ​​input into the genetic algorithm, which solves the traveling salesman problem to obtain the shortest path sequence that visits all target nodes and returns to the starting node.

[0054] Each target node in the shortest path sequence is mapped back to its corresponding task to generate the joint scheduling path.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] This invention addresses the shortcomings of existing technologies by accurately collecting task execution characteristic data and spatial location identifiers, and constructing a task feature vector set by combining multi-dimensional information such as environmental impact coefficients and dynamic resource availability, ensuring comprehensive and accurate scheduling data. It utilizes a modality recognition network model to extract spatiotemporal and resource demand attributes and map them to a primary scheduling modality. Through iterative updates of coupling weights and higher-level scheduling modalities, the resource allocation pattern is made to fit real-time operation and maintenance scenarios. Based on association strength and task dependency graphs, task groups are divided, and clustering algorithms and the Traveling Salesman Problem are used to generate joint scheduling paths and independent scheduling instructions, significantly improving resource utilization and task execution efficiency. Ultimately, this achieves optimized operation and maintenance personnel configuration, reduced resource waste, and improved task response time, while simultaneously reducing manual scheduling errors and operation and maintenance costs, ensuring refined, efficient, and intelligent park operation and maintenance services. Attached Figure Description

[0057] Figure 1 This is a flowchart of the spatial resource dynamic allocation method based on AI model and BIM operation and maintenance in Embodiment 1 of the present invention.

[0058] Figure 2 Here is a flowchart of the priority score acquisition process for Embodiment 1 of the present invention;

[0059] Figure 3 This is a flowchart illustrating how multiple primary scheduling modes are mapped in Embodiment 1 of the present invention. Detailed Implementation

[0060] Example 1

[0061] Please see Figure 1The present invention provides an embodiment of a method for dynamic allocation of spatial resources based on AI models and BIM operation and maintenance, comprising the following steps:

[0062] S1. Obtain the execution feature data of multiple tasks to be processed within the target area and the spatial location identifiers corresponding to each task, and construct a task feature vector set based on the execution feature data;

[0063] In one implementation, the target area is a physical park with a unified management boundary defined based on a Building Information Model (BIM). The tasks to be processed are specific BIM operation instances that require response and completion by BIM resources, including at least fault repair work orders, preventative inspection work orders, and planned maintenance work orders. To facilitate unified modeling and subsequent collaborative scheduling of the parallel execution of multiple tasks, each task to be processed is configured with a spatial location identifier. This spatial location identifier characterizes the physical location where the task execution occurs and its relationship within the management topology, including at least a spatial unit code and a region identifier. The spatial unit code is used to associate with the smallest management unit in the BIM, such as a room, equipment room, or public area. The region identifier characterizes that multiple tasks physically or logically belong to the same scheduling partition, such as the same floor, the same fire compartment, or the same property management responsibility area. This implies that tasks within the partition have a small spatial distance and high collaborative scheduling potential, thus providing a spatial semantic basis for subsequent task clustering, path planning, and resource allocation.

[0064] In this implementation, acquiring execution characteristic data includes collecting attribute data from business systems, IoT devices, or manual input channels associated with each pending task to characterize the task status and requirements. Execution characteristic data includes at least multi-dimensional attribute data reflecting task urgency, resource requirements, estimated time consumption, and environmental constraints, such as work order creation time, fault level, required job types, tool list, theoretical working hours, and associated spatial environmental indicators. This data can be acquired through data interface integration with work order management systems, asset management systems, and IoT platforms. To ensure support for refined dynamic scheduling decisions, the collected data attributes should cover key dimensions affecting scheduling optimization. To guarantee unified processing and feature extraction from heterogeneous data sources, the execution characteristic data can undergo necessary cleaning and standardization after collection, including timestamp synchronization, unit unification, enumeration value mapping, and null value handling, making task data from different sources comparable and computable.

[0065] Constructing a task feature vector set based on execution feature data involves: using a scheduling engine or data preprocessing module to generate corresponding feature records for each task to be processed within each scheduling decision cycle or preset time window, and binding the feature records with spatial location identifiers. The feature records include at least task identifier information, spatial location identifier information, and standardized multidimensional execution feature data, enabling the task feature vector to be represented as a set of attributes incorporating spatial semantics and task dimensions.

[0066] In one optional implementation, time-related features are calibrated using the time of the upper-level scheduling system as a unified benchmark. When there are deviations in the update time of different data sources, the feature data can be aligned to the unified scheduling decision time according to their actual effective time or through interpolation. When there is contradictory or abnormal data, it can be verified and fused according to preset data quality rules to ensure the accuracy and reliability of the task feature vector set.

[0067] Further explanation is needed; please refer to [link / reference]. Figure 2 In this embodiment, the recommendation priority score is obtained, including:

[0068] S101. Obtain multiple pending maintenance work orders, wherein each pending maintenance work order is associated with at least one original timestamp, a static spatial identifier, and a set of static resource requirement attributes. It should be further explained that the static spatial identifier in this embodiment includes a spatial unit code uniquely associated with a specific physical space (such as a room or equipment room) in the building information model and an area identifier representing the management affiliation (such as a floor or responsibility zone), which is used to accurately locate the location where the task occurs and provide a spatial topology basis for task clustering and path planning based on geographical proximity. The static resource requirement attributes include the types of work necessary to complete the task, a list of special tools, and a rated working time estimate, which are used to clearly define the core requirements of the task for maintenance manpower, materials, and time, and provide key constraints for subsequent resource matching, conflict detection, and scheduling optimization.

[0069] S102. Based on a unified time benchmark, calibrate the original timestamp associated with each pending maintenance work order and obtain the calibration timestamp of each pending maintenance work order.

[0070] S103. Based on the static spatial identifier associated with each pending maintenance work order, query the associated IoT sensing devices, obtain real-time location sensing data, compare the real-time location sensing data with the static spatial identifier, and obtain the approved spatial coordinates and location confidence of each pending maintenance work order.

[0071] S104. Based on the static resource requirement attributes of each pending maintenance work order, query the resource management system in real time to obtain the real-time availability status of the corresponding resource entity and generate a dynamic resource availability code.

[0072] Optionally, in scenarios where maintenance work orders from multiple heterogeneous systems within a park converge, work order data is often generated and reported independently by different business systems or mobile terminals. Due to differences in system clocks, network transmission, and data processing, the original timestamps of different pending maintenance work orders may have time discrepancies. This time discrepancy is equivalent to priority misalignment in the scheduling sequence, causing inconsistencies in the time ordering of work orders with the same physical event or similar urgency. If this original timestamp is directly used as the input for the calibration timestamp, the scheduling engine may misjudge the sorting error introduced by the time synchronization difference as a difference in the urgency of the work orders themselves, thereby affecting the reliability of subsequent clustering, association strength calculation, and scheduling timing.

[0073] Based on this, this application introduces time deviation detection and alignment calibration to ensure that the timestamps entering the scheduling decision have a consistent time caliber across work orders. In this optional implementation, the scheduling engine collects the original timestamps of all pending maintenance work orders. Timestamp parsing, time zone conversion, and necessary format standardization can be implemented using existing engineering methods to ensure the availability of time series data; the above processing is not a limiting focus of this application.

[0074] The distinguishing feature of this application is that the scheduling engine first determines the transmission and processing delay of the reported data stream relative to a unified time base for work orders from different sources, and generates a time compensation amount accordingly to calibrate and align the original timestamps. The unified time base can be provided by the scheduling engine's master clock, a campus time server, or the master timestamp of a designated high-priority business system. Time deviation can be obtained based on network time synchronization protocols, data packet arrival timestamps, and internal processing delay estimates within the business system. In the absence of reliable delay data, deviation estimates can also be obtained through alignment analysis between the event logic relationships and timestamp sequences of batch work orders, and consistency checks are performed on the deviation estimates to avoid miscalibration caused by occasional anomalies. The time compensation amount can be generated according to the transmission and processing delays and fine-tuned by combining typical delay experience values ​​preset for the work order type. To avoid timestamp jumps introduced by calibration, the scheduling engine can use batch-wise gradual updates to the time compensation amount, ensuring logical continuity of the calibrated timestamp sequence.

[0075] For example, in a scenario where the work order collection cycle is 1 minute and the maximum allowable scheduling response delay is 5 minutes, the time deviation tolerance can be configured to ±30 seconds, and the transition window for calibration gradient can be configured to 2-3 collection cycles. When the time deviation of a certain work order source is detected to continuously exceed the allowable tolerance, or the deviation exhibits unstable jumps within multiple consecutive cycles, the scheduling engine can mark the work order source as a time anomaly source and set a low-confidence flag for the work order timestamps from that source. In subsequent priority calculations and time window planning, a conservative strategy or the estimated value of the valid timestamp from the previous cycle can be used instead to reduce the interference of time anomaly data on the stability of the scheduling sequence.

[0076] Optionally, to suppress the instability of dynamic resource availability coding caused by frequent fluctuations in resource status or non-real-time query responses, which propagates to subsequent task cluster partitioning and path planning, the scheduling engine performs sparsification and state preservation processing on the original resource status set obtained from the query, so that the resource availability representation focuses on key constraint states. Specifically, the scheduling engine can select a subset of key states that affect scheduling decisions from multiple state attributes of the same resource entity as the coding basis, while other state attributes are ignored or used only as background references. In implementations that require maintaining state continuity, fluctuations in non-critical states can also be smoothed by a preset filtering factor, and threshold judgment and state preservation can be performed on key states, so that the generated dynamic resource availability coding has a consistent scheduling interpretation across different decision cycles. For example, in a configuration with 5-8 resource status attributes, the key state subset can be preset to include 2-3 items such as whether it is occupied, the expected release time, and the health status. The filtering factor can be 0.2-0.5, and the state preservation time can be set to 1-2 decision cycles, so that the coding output presents the effect of stable main constraints and filtered out secondary fluctuations.

[0077] Furthermore, to avoid frequent critical state flips in the sparsity-preserving resource availability code under critical resource contention conditions, which would lead to repeated recalculation of task cluster partitioning and path schemes, the scheduling engine introduces an update suppression strategy: when a critical state of the same resource entity is detected to have changed within two consecutive decision cycles, and the resource is depended upon by an currently executing or planned path, the availability code of the previous decision cycle is maintained until the current scheduling instruction dependent on the resource is completed or a preset safe release window is reached. Here, critical state changes can be determined by comparing state codes; the preset safe release window is used to limit the maximum time the state is maintained, and its length can be configured based on the typical execution time of tasks on the resource. For example, with a decision cycle of 5 minutes and a typical task execution time of 15-60 minutes, the safe release window can be configured to 30-90 minutes, and a criterion of no new high-priority dependent tasks can be added within the window to avoid the resource being invalidally locked.

[0078] In one optional implementation, after the update suppression policy is triggered, the scheduling engine can continue to receive the real-time status of the resource for monitoring and early warning. However, before the release condition is met, the status is only used for internal monitoring and not for regenerating the dynamic resource availability code. Alternatively, during the duration of the update suppression policy, the scheduling engine uses a resource reservation identifier for task clusters that depend on the resource, and the path planning module processes them according to the resource reservation availability. Through the above-mentioned sparse coding and update suppression policy for resource status, it is possible to suppress the frequent flipping of resource availability identifiers to propagate to downstream task clustering and path planning stages under the conditions where resource status fluctuates over time and scheduling dependencies are complex due to multi-task competition. This reduces the risk of system oscillation and decision instability caused by frequent reconfiguration of scheduling schemes and improves the continuity and feasibility of dynamic allocation.

[0079] S105. Based on the calibration timestamp and approved spatial coordinates of each pending maintenance work order, query the environmental data service and internal system status data, and calculate the environmental impact coefficient of each pending maintenance work order through the pre-trained environmental impact assessment model.

[0080] In one implementation, after the scheduling engine obtains the calibration timestamp and approved spatial coordinates of the maintenance work order to be processed, it queries environmental and situational data based on the calibration timestamp and approved spatial coordinates, and calculates the environmental impact coefficient through a pre-trained environmental impact assessment model. In this embodiment, to ensure that the environmental impact assessment model has a definite scope that is feasible for engineering, the scheduling engine pre-establishes a spatiotemporal feature-environmental impact factor mapping table or an environmental scenario prototype library, so that the environmental scenario identifiers obtained in the offline training phase are stored in correspondence with their typical impact factor ranges. When a certain environmental scenario is identified online, the main impact factors corresponding to that environmental scenario are directly read as the basis for calculation. When calculating the environmental impact coefficient of the maintenance work order, the scheduling engine first obtains the basic environmental impact factors through the predefined scenario library, and then introduces specific environmental indicators collected in real time, including at least temperature, humidity, and wind speed, and dynamically fine-tunes the weights of these factors based on the correlation between environmental indicators and work order execution efficiency in historical data.

[0081] For example, the daily scheduling cycle is 24 hours, the data update frequency is every 15 minutes, the model inference window is the data of the most recent 4 hours, and the main environmental impact factors are configured according to the environmental scenario prototype library as one of the following types: extreme weather scenarios correspond to an outdoor operation efficiency reduction coefficient of 0.5 to 0.8; high-density population scenarios correspond to a public area operation delay coefficient of 1.2 to 1.5; and important event support scenarios correspond to a specific area resource priority coefficient of 0.9 to 1.1. For a specific work order, its environmental impact coefficient is calculated by weighting multiple factors, and the weights are adaptively adjusted according to the work order type and location.

[0082] In one implementation, the scheduling engine generates a set of scheduling strategy adjustment parameters that match the environmental impact coefficient, emphasizing robustness in harsh environments and efficiency in normal environments. The scheduling strategy adjustment parameter set includes at least two types of fields: a work hour flexibility coefficient and a resource redundancy coefficient. The work hour flexibility coefficient is used to adjust the estimated work hours of work orders upwards or downwards based on environmental impact, while the resource redundancy coefficient is used to allocate spare resources or extend resource occupancy time when the environment is uncertain. Strategy adjustment is achieved by configuring constraints associated with the environmental impact coefficient in the scheduling optimization model. The scheduling optimization model in this embodiment refers to the computational framework used to generate the final work order resource allocation scheme. Its core optimization objective is to minimize the total execution cost or maximize the completion efficiency, and it can dynamically adjust its objective function and constraints based on the environmental impact coefficient. This model automatically adjusts the scheduling strategy adaptively during optimization by using the time flexibility coefficient as an adjustment multiplier for task duration constraints and the resource redundancy coefficient as an adjustment parameter for resource availability constraints: when the environmental impact coefficient is high, the model relaxes time constraints and increases resource budget through the above parameters to generate a robust solution; when the coefficient is low, the constraints are tightened to generate an efficiency-oriented solution.

[0083] In this implementation, to ensure the feasibility of work plans under harsh environments, the scheduling engine configures the time flexibility coefficient corresponding to extreme weather scenarios to be no less than a preset baseline flexibility threshold. The baseline flexibility threshold is configured according to the time adjustment ratio, increasing the theoretically estimated time by 30% to 50%. Simultaneously, to prevent environmental impacts from causing critical task failures, the scheduling engine configures a resource redundancy coefficient for high-importance work orders with a level no less than a preset lower limit. This coefficient is reflected in the allocation of an additional collaborator or the reservation of a longer tool usage time window.

[0084] In one implementation, the scheduling engine uses the calculated environmental impact coefficient for subsequent weighted estimated work hours calculations, and applies boundary constraints and smoothing to the coefficient to ensure the stability of the scheduling plan. The scheduling replanning cycle is 1 hour, limiting the variation of the environmental impact coefficient of a single work order between adjacent cycles to within ±20% of the coefficient of the previous cycle, and setting upper and lower limits of 0.5 to 2.0 for the coefficient value to keep it within an empirically verified safe range.

[0085] Optionally, when there is a significant discrepancy between the environmental impact coefficient calculated by the model and the real-time sensor data, or when the input data is obviously abnormal, the scheduling engine triggers a data verification and fallback strategy. Specifically, before outputting the environmental impact coefficient, the scheduling engine compares it with the empirical coefficients directly calculated based on real-time sensor data. When the deviation exceeds a preset deviation threshold, a verification process is initiated, and the model coefficients are corrected according to rules such as data source reliability and historical accuracy, or the system is switched to empirical coefficients based on real-time data. The preset deviation threshold is set to 0.3 per unit value.

[0086] The empirical coefficient is calculated in real time using readings from temperature, humidity, light, and pedestrian flow sensors deployed in the target area, through a lightweight linear regression model. A fallback strategy ensures that no work order will fail to calculate the coefficient due to abnormal environmental data. When the model output is abnormal for three consecutive periods, it automatically switches to the average value based on recent historical data for the same period, and marks the environmental impact coefficient as having a backup value enabled in the work order status.

[0087] S106. Based on the theoretical estimated working hours in the static resource requirement attributes of each pending maintenance work order, and combined with the environmental impact coefficient, calculate the weighted estimated working hours of each pending maintenance work order.

[0088] S107. Based on the calibration timestamp, fault level, weighted estimated working hours, and dynamic resource availability code of each pending maintenance work order, calculate the recommended priority score for each pending maintenance work order through the preset priority calculation rules.

[0089] In one implementation, when a pending maintenance work order enters the priority calculation stage, the scheduling engine further determines its workload level based on the weighted estimated working hours output by S106, adapting to differentiated scheduling strategies. To ensure this division has a clear and executable caliber, the scheduling engine pre-establishes a working hour-load level mapping table, dividing the weighted estimated working hours into three levels—short, medium, and long—according to preset intervals, and corresponding to different scheduling resource matching rules and path planning flexibility. The scheduling engine can also fine-tune the level boundaries near the center value of each level, combining the actual time distribution of similar historical work orders. The fine-tuning can be based on the quantile interval obtained from the time statistics of the completed work order database, ensuring that the level division reflects the expected load of the current work order while avoiding misjudgment of the level due to single estimation deviations.

[0090] For example, in typical operation and maintenance scenarios, the workload levels can be configured according to the mapping table as follows: short-term corresponds to a weighted estimated workload of less than 30 minutes; medium-term corresponds to 30 minutes to 2 hours; and long-term corresponds to more than 2 hours. For work orders that are temporary or highly uncertain, the level boundaries can be set to a floating range, for example, medium-term can be defined as 25 minutes to 130 minutes, in order to reserve a certain buffer margin.

[0091] In one implementation, the scheduling engine generates priority calculation rules that match the workload level and fault level. This ensures that during resource allocation, the recommended priority score prioritizes rapid response for high-fault-level work orders and resource assurance for high-load-level work orders. The priority calculation rules include at least two adjustment dimensions: a time urgency coefficient and a resource assurance coefficient. The time urgency coefficient increases the ranking weight of high-fault-level work orders in the scheduling sequence, while the resource assurance coefficient ensures that high-load-level work orders receive necessary resource reservations or concurrent support during resource matching. This adjustment is achieved by configuring differentiated weighting factors in the recommended priority score calculation formula. These weighting factors can be obtained by combining the fault level and workload level of the work order.

[0092] In this implementation, to ensure timely response to high-fault-level work orders, the scheduling engine configures the lower limit of the time flexibility coefficient for urgent work orders to be no less than a preset baseline urgency coefficient. This coefficient can directly constrain the time window in the optimization model, for example, by increasing the weight of time urgency in the objective function from the base value of 0.35 to 0.50. Simultaneously, to avoid excessive delays in high-load or long-duration work orders due to resource contention, the scheduling engine configures a resource redundancy coefficient no less than a preset lower limit for them. This coefficient can be reflected in the query priority of dynamic resource encoding, or directly in the time window and reserve capacity for pre-occupying critical resources.

[0093] For example, the resource security factor lower limit can be configured to ensure that the required critical tasks or tools are locked in at least 60 minutes before their expected start time.

[0094] In one implementation, the scheduling engine uses the final calculated recommended priority score for task sorting and scheduling instruction generation, and applies smoothing filtering and boundary protection to the score to ensure the stability of the scheduling plan. For example, in a scenario with a scheduling cycle of 5 minutes and task queue updates, the change in the priority score of a single work order between adjacent cycles can be limited to ±10% of the score of the previous cycle, and a minimum score threshold (e.g., 0.1) can be set for work orders entering the scheduling sequence to filter out minor or invalid scheduling requests.

[0095] Optionally, when the recommended priority score calculated using preset priority calculation rules shows significant contradictions (e.g., high fault level but extreme resource conflict) or when the input data is obviously abnormal, the scheduling engine can trigger conflict arbitration and fallback strategies. Specifically, before outputting the priority score, the scheduling engine verifies its compatibility with the current global resource occupancy rate; when the scheduling sequence indicated by the score will cause resource conflicts exceeding the preset conflict threshold, conflict arbitration is initiated, and the score is adjusted a second time according to rules such as work order type and committed service level agreement, or it is included in the arbitration queue for manual confirmation. The preset conflict threshold can be set according to the concurrent utilization rate of critical resources (such as special vehicles, senior technicians), for example, 85%.

[0096] For example, conflict arbitration can postpone the scheduling time of conflicting work orders to a limited extent, or allocate alternative resources to them, and mark the work order status as arbitrated and adjusted. A fallback strategy ensures that no work order is indefinitely shelved due to system logic errors; for example, it automatically assigns a progressively increasing waiting time bonus to work orders that have been waiting in the queue for more than 2 hours, forcing them into a high-priority sequence.

[0097] S108. Integrate the calibration timestamp, approved spatial coordinates, location confidence, dynamic resource availability coding, weighted estimated working hours, and recommended priority score of each pending maintenance work order to form the maintenance feature vector of the pending maintenance work order. Summarize the maintenance feature vectors of all pending maintenance work orders to generate a task feature vector set.

[0098] S2. Input the task feature vector set into the modality recognition network model to map multiple primary scheduling modes; wherein, the spatiotemporal attributes in the task feature vector are encoded as the spatial coupling components of the primary scheduling modes, and the resource requirement attributes in the task feature vector are encoded as the resource requirement components of the primary scheduling modes.

[0099] In one example, such as Figure 3 As shown, based on the task feature vector set of pending maintenance work orders within the physical park defined by the Building Information Model, multiple primary scheduling modes are mapped. The specific process includes:

[0100] S201. Extract the spatiotemporal attribute sub-vector set and the resource requirement attribute sub-vector set from the task feature vector set. Specifically, the task feature vector set contains multi-dimensional attribute information of the task, and directly using it for scheduling mode generation would lead to low efficiency due to feature redundancy. By separating the spatiotemporal attribute and resource requirement attribute sub-vector sets, we can focus on the spatial and temporal distribution characteristics and resource requirement characteristics of the task, respectively, laying the foundation for subsequent targeted feature processing.

[0101] In this embodiment, the task feature vector set is traversed, and attributes such as location coordinates, execution time window, time priority, and spatial association range of each task are extracted to form a spatiotemporal attribute sub-vector set. For example, the spatiotemporal attribute sub-vector of task 1 is (X=100, Y=200, execution time window=9:00-11:00, time priority=0.8, spatial association range=30 meters), and the spatiotemporal attribute sub-vector of task 2 is (X=120, Y=210, execution time window=9:30-11:30, time priority=0.7, spatial association range=25 meters). Simultaneously, attributes such as the skill level of maintenance personnel, type and quantity of equipment and tools, material specifications and usage, and resource occupation duration of each task are extracted to form a resource requirement attribute sub-vector set. For example, the resource requirement attribute sub-vector for Task 1 is (Operation and Maintenance Personnel Skill Level = Intermediate Electrician, Equipment and Tool Type and Quantity = 1 Multimeter + 1 Set of Insulation Tools, Material Specifications and Usage = 5 Meters of Wire, Resource Occupancy Time = 60 Minutes), and the resource requirement attribute sub-vector for Task 2 is (Operation and Maintenance Personnel Skill Level = Basic HVAC, Equipment and Tool Type and Quantity = 1 Pressure Gauge + 1 Wrench, Material Specifications and Usage = 1 Tube of Sealant, Resource Occupancy Time = 45 Minutes).

[0102] S202. Construct a spatiotemporal attribute matrix based on the spatiotemporal attribute sub-vector set, input the spatiotemporal attribute matrix into the multi-head self-attention layer, and obtain a weighted spatiotemporal feature matrix by calculating the spatial correlation between tasks. It should be further noted that the spatiotemporal attribute matrix in this embodiment represents the distribution state of all tasks in the position and time dimensions, with rows corresponding to tasks and columns corresponding to spatiotemporal features, which are used to quantify the proximity and temporal correlation between tasks in the subsequent process. It should also be noted that the weighted spatiotemporal feature matrix in this embodiment represents the strength of the spatial relationship after adjustment by the mutual attention mechanism between tasks, and its element values ​​reflect the degree of spatial coupling between any two tasks at the scheduling level, which is used to support the accurate calculation of the subsequent spatial coupling components to identify task clusters that can be merged and scheduled.

[0103] Specifically, the spatiotemporal attribute sub-vector set is in the form of a dispersed vector. Constructing a spatiotemporal attribute matrix enables the structured organization of task spatiotemporal features, facilitating the batch calculation of inter-task relationships. The multi-head self-attention layer can effectively capture the spatial correlation between tasks, highlighting key correlations through attention weight adjustment. The generated weighted spatiotemporal feature matrix can accurately reflect the degree of spatial coupling between tasks.

[0104] In this embodiment, the spatiotemporal attribute subvectors of all tasks are arranged in rows to construct a spatiotemporal attribute matrix. The rows of the matrix correspond to a single task, and the columns correspond to spatiotemporal features such as location coordinates and execution time windows. For example, the spatiotemporal attribute matrix containing two tasks is: [[100, 200, 9:00-11:00, 0.8, 30], [120, 210, 9:30-11:30, 0.7, 25]].

[0105] The matrix is ​​input into a multi-head attention layer with 4 attention heads. Initial correlations are obtained by calculating the Euclidean distance between spatial coordinates of tasks and the overlap of execution time windows. Then, the matrix is ​​fused and weighted through a multi-head attention mechanism to generate a weighted spatiotemporal feature matrix. For example, the spatial coupling weight between task 1 and task 2 is 0.75, the attention weight of task 1 itself is 0.25, and the attention weight of task 2 itself is 0.30. The resulting weighted spatiotemporal feature matrix is: [[75, 150, 0.6, 0.6, 22.5], [90, 157.5, 0.525, 0.525, 18.75]].

[0106] S203. Based on the weighted spatiotemporal feature matrix, a linear transformation is performed through the first fully connected layer to output the spatial coupling component matrix. It should be further explained that the spatial coupling component matrix in this embodiment represents the coupling strength vector calculated by each task based on its spatial interaction relationship with all other tasks. The rows of the matrix correspond to the tasks, and each element in the row represents the coupling strength value of the task in a specific spatial association dimension. This is used to quantify the tendency of each task to integrate into the spatial task cluster, providing core spatial feature input for subsequent task clustering and joint scheduling path planning.

[0107] Specifically, the weighted spatiotemporal feature matrix contains information on the spatial correlation strength between tasks, but it needs to be mapped to a coupling strength vector of a unified dimension through a linear transformation in order to quantify the tendency of each task to integrate into the spatial task cluster. The first fully connected layer can achieve the unification of feature dimensions and the condensation of feature information, and the output spatial coupling component matrix provides core spatial feature support for subsequent task clustering.

[0108] In this embodiment, the input dimension of the first fully connected layer is 5 columns of the spatiotemporal attribute matrix, and the output dimension is set to 4. This output dimension is determined based on the historical clustering effect and feature expression requirements. The first fully connected layer performs a linear transformation on the weighted spatiotemporal feature matrix, mapping the weighted spatiotemporal features of each task to a 4-dimensional spatial coupling strength vector. The spatial coupling strength vectors of all tasks are arranged in rows to form a spatial coupling component matrix. For example, the spatial coupling strength vector of task 1 is (0.82, 0.76, 0.68, 0.73), and the spatial coupling strength vector of task 2 is (0.71, 0.65, 0.79, 0.62). The spatial coupling component matrix is: [[0.82, 0.76, 0.68, 0.73], [0.71, 0.65, 0.79, 0.62]].

[0109] S204. Construct a resource demand attribute matrix based on the resource demand attribute sub-vector set. Input the resource demand attribute matrix into a resource demand encoding network consisting of at least one fully connected layer. Transform it using a nonlinear activation function to output a resource demand component matrix. Specifically, the resource demand attribute sub-vector set contains the task's demand information for various resources. Constructing the resource demand attribute matrix enables structured management of resource demand features. The resource demand encoding network maps the original resource demand features into normalized intensity vectors through nonlinear transformation and dimensionality compression, facilitating the quantification of the urgency and exclusivity of the task's resource demands.

[0110] In this embodiment, the resource requirement attribute sub-vectors of all tasks are arranged by row to construct a resource requirement attribute matrix. The rows of the matrix correspond to a single task, and the columns correspond to resource requirement characteristics such as the skill level of maintenance personnel and the quantity of equipment and tools. For example, the resource requirement attribute matrix containing two tasks is: [[Intermediate Electrician, 1 multimeter + 1 set of insulating tools, 5 meters of wire, 60 minutes], [Basic HVAC, 1 pressure gauge + 1 wrench, 1 tube of sealant, 45 minutes]].

[0111] The matrix is ​​input into a resource requirement encoding network, which consists of two fully connected layers and two activation function layers. The first fully connected layer has an input dimension of 4 and an output dimension of 8. It performs a linear transformation on the resource requirement attribute matrix, mapping the resource requirement attribute sub-vectors of each task to an 8-dimensional intermediate feature space. A modified linear unit function is applied to the output of the first fully connected layer for non-linear activation, introducing a non-linear feature transformation to enhance feature representation. The second fully connected layer has an input dimension of 8 and an output dimension of 3. It performs a dimensionality reduction linear transformation on the activated features, compressing the feature dimension to the preset output dimension. A sigmoid activation function is applied to the output of the second fully connected layer for normalization mapping, restricting the final output vector of each task to the range of 0 to 1, forming the resource requirement component matrix. For example, the normalized resource demand intensity vector for Task 1 is (0.85, 0.78, 0.69), and the normalized resource demand intensity vector for Task 2 is (0.72, 0.81, 0.75). The resource demand component matrix is: [[0.85, 0.78, 0.69], [0.72, 0.81, 0.75]].

[0112] S205. Concatenate the row vectors corresponding to each task in the spatial coupling component matrix with the row vectors corresponding to the task in the resource demand component matrix to obtain the primary scheduling mode vector for each task. Specifically, in this embodiment, the spatial coupling component matrix focuses on the spatial features of the task, and the resource demand component matrix focuses on the resource demand features of the task. Concatenating the corresponding row vectors of the two can form a joint low-dimensional embedding representation that combines spatial coupling characteristics and resource demand characteristics, providing a unified feature input for subsequent modal clustering and collaborative optimization decision-making.

[0113] In this embodiment, the row vector corresponding to each task in the spatial coupling component matrix is ​​concatenated with the row vector corresponding to the task in the resource demand component matrix in sequence. For example, the spatial coupling strength vector (0.82, 0.76, 0.68, 0.73) of task 1 is concatenated with the resource demand strength vector (0.85, 0.78, 0.69) to obtain the primary scheduling mode vector of task 1 (0.82, 0.76, 0.68, 0.73, 0.85, 0.78, 0.69); the spatial coupling strength vector (0.71, 0.65, 0.79, 0.62) of task 2 is concatenated with the resource demand strength vector (0.72, 0.81, 0.75) to obtain the primary scheduling mode vector of task 2 (0.71, 0.65, 0.79, 0.62, 0.72, 0.81, 0.75). The concatenated vector has a spatial coupling component in the first half and a resource requirement component in the second half, thus achieving the fusion of the two types of features.

[0114] S206. Aggregate the primary scheduling mode vectors of all tasks to obtain the multiple primary scheduling modes. Specifically, the primary scheduling mode vector of a single task only reflects the feature fusion result of that task. Aggregating the vectors of all tasks can form a set of primary scheduling modes covering all tasks to be processed, providing a complete feature basis for subsequent high-level scheduling mode iteration updates and resource allocation mode determination.

[0115] In this embodiment, the primary scheduling mode vectors of all tasks to be processed are collected and arranged in order of task number to form multiple primary scheduling modes. For example, in addition to tasks 1 and 2, if there are tasks 3 to n, whose corresponding primary scheduling mode vectors are V3 to Vn respectively, then the multiple primary scheduling modes obtained after aggregation are {V1, V2, V3, ..., Vn}, and each vector is a 7-dimensional joint low-dimensional embedding representation, which fully covers the spatial coupling characteristics and resource requirement characteristics of each task.

[0116] S3. Iteratively update the coupling weights between the multiple primary scheduling modes and the higher-level scheduling modes, and determine convergence when the preset convergence condition is met; determine the current resource allocation mode of the system based on the converged higher-level scheduling mode, and determine the correlation strength of each task relative to the current resource allocation mode based on the converged coupling weights.

[0117] In one example, the correlation strength of each task relative to the current resource allocation mode is determined based on pending maintenance work orders within the physical park defined by the Building Information Model (BIM). These pending maintenance work orders include fault repair work orders, preventative inspection work orders, and planned maintenance work orders. The specific process includes:

[0118] S301. Initialize a high-level scheduling mode vector and the coupling weights between each primary scheduling mode and the high-level scheduling mode vector;

[0119] It should be further explained that the process of initializing the high-level scheduling mode vector and coupling weights in this embodiment is as follows:

[0120] Based on the dimensions of the spatiotemporal attribute sub-vectors and resource requirement attribute sub-vectors in the task feature vector set, the dimension of the high-level scheduling modal vector is determined to be the sum of the dimensions of the two types of sub-vectors. The spatiotemporal attribute sub-vectors include location coordinates, execution time windows, etc., while the resource requirement attribute sub-vectors include the skill level of maintenance personnel, the quantity of equipment and tools, etc. A random initialization method is used to generate the high-level scheduling modal vector, with each element value distributed between 0 and 1. The similarity between each primary scheduling modal vector and the high-level scheduling modal vector is calculated using a cosine similarity algorithm, and this similarity value is used as the initial coupling weight. The primary scheduling modal vector is formed by concatenating the corresponding row vectors of the spatial coupling component matrix and the resource requirement component matrix.

[0121] For example, if the spatiotemporal attribute sub-vector has a dimension of 8 and the resource demand attribute sub-vector has a dimension of 6, then the high-level scheduling modality vector has a dimension of 14. The generated vector element values ​​can be randomly distributed between 0.2 and 0.8. When the cosine similarity calculation result of the two vectors is 0.65, the initial coupling weight corresponding to the primary scheduling modality is 0.65.

[0122] S302. Obtain the latest real-time dynamic status data for each task, and calculate the dynamic status adjustment factor based on the real-time dynamic status data. It should be further explained that the process of calculating the dynamic status adjustment factor in this embodiment is as follows: Real-time dynamic status data for each task is collected in real-time by IoT sensing devices. This includes real-time environmental data such as temperature, humidity, and air quality at the task execution location; related equipment operating status data such as equipment operating parameters and fault alarm information; real-time availability status data of required resources such as resource occupancy status and estimated release time; and task execution progress data such as completed workload and remaining working hours. The collected data is processed using the Z-Score standardization method. The weight coefficients of each dimension of data are determined based on the analytic hierarchy process (AHP), where the weight of real-time environmental data is 0.3, the weight of equipment operating status data is 0.35, the weight of real-time resource availability status data is 0.2, and the weight of task execution progress data is 0.15. A weighted summation is used to obtain a comprehensive dynamic status value. Based on this comprehensive value and a preset baseline status value, the dynamic status adjustment factor is obtained. For example, the standardized environmental data score of a certain fault repair work order is 0.7, the equipment operation status data score is 0.8, the resource real-time availability status data score is 0.6, and the task execution progress data score is 0.5. After weighted summation, the dynamic status comprehensive value is 0.68, the preset baseline status value is 0.5, and the calculated dynamic status adjustment factor is 1.36.

[0123] S303. The corresponding primary scheduling mode is corrected element-by-element by weighting using the dynamic state adjustment factor to obtain the set of primary scheduling modes after dynamic correction. It should be further explained that the process of correcting the corresponding primary scheduling mode by weighting using the dynamic state adjustment factor in this embodiment is as follows: weighting correction is performed by element-by-element multiplication based on the dynamic state adjustment factor of each task and the corresponding primary scheduling mode vector. For example, a primary scheduling mode vector is [0.5, 0.7, 0.6, 0.8, 0.4, 0.9, 0.5, 0.6, 0.7, 0.8, 0.5, 0.6, 0.7, 0.8], with a corresponding dynamic state adjustment factor of 1.2. After multiplying each element, the dynamically corrected primary scheduling mode vector is obtained as [0.6, 0.84, 0.72, 0.96, 0.48, 1.08, 0.6, 0.72, 0.84, 0.96, 0.6, 0.72, 0.84, 0.96]. The corrected vectors of all tasks are then combined to form a dynamically corrected primary scheduling mode set.

[0124] S304. Based on the dynamically corrected primary scheduling mode set and the higher-level scheduling mode vector, update the coupling weights using the Softmax function;

[0125] S305. Based on the updated coupling weights, the dynamically corrected primary scheduling mode set is weighted and summed to update the higher-level scheduling mode vector. It should be further explained that the process of updating the higher-level scheduling mode vector in this embodiment is as follows: based on the dynamically corrected primary scheduling mode set and the higher-level scheduling mode vector, the coupling weights are updated using the Softmax function. For example, the correlation results between the three dynamically corrected primary scheduling mode vectors and the higher-level scheduling mode vector are 0.3, 0.4, and 0.3, respectively. After normalization using the Softmax function, the updated coupling weights are 0.24, 0.52, and 0.24, respectively. Based on the updated coupling weights, the dynamically corrected primary scheduling mode set is weighted and summed to update the higher-level scheduling mode vector. For example, the three corrected vectors are [0.6, 0.84, ...], [0.55, 0.78, ...], and [0.62, 0.86, ...]. The first two elements of the updated high-level scheduling modality vector after weighted summation are 0.6×0.24+0.55×0.52+0.62×0.24≈0.58 and 0.84×0.24+0.78×0.52+0.86×0.24≈0.81.

[0126] S306. Calculate the change between the high-level scheduling modal vector after the current iteration and the high-level scheduling modal vector after the previous iteration. If the change is less than the preset convergence threshold of 0.001, it is determined to be converged. For example, the first two elements of the high-level scheduling modal vector in the previous iteration are 0.57 and 0.80, and in the current iteration they are 0.58 and 0.81, with a change of 0.01 in both cases, which is greater than the convergence threshold, so the iteration continues. When iterating to the 8th iteration, the change between the vectors in the previous and next iterations is 0.0008, which is less than 0.001, so it is determined to be converged, and the high-level scheduling modal vector of this iteration is determined as the current resource allocation mode.

[0127] S307. The high-level scheduling mode vector at convergence is determined as the current resource allocation mode of the system. The system in this embodiment is a multi-task collaborative scheduling system for park operation and maintenance. The system is based on multiple pending operation and maintenance work orders in a physical park with a unified management boundary defined by the Building Information Model. By acquiring the execution feature data of the pending operation and maintenance work orders, a task feature vector set is constructed. The task feature vector set is input into the modality recognition network model to map multiple primary scheduling modes. The high-level scheduling mode vector and the coupling weight between each primary scheduling mode and the high-level scheduling mode vector are initialized. The real-time dynamic status data of each task is acquired to calculate the dynamic status adjustment factor. The primary scheduling modes are weighted and corrected element by element using the dynamic status adjustment factor to obtain a dynamically corrected set of primary scheduling modes. The coupling weight is updated by the Softmax function, and the high-level scheduling mode vector is updated by weighted summation based on the updated coupling weight. Finally, the high-level scheduling mode vector at convergence is determined as the current resource allocation mode of the system, realizing the collaborative scheduling of multiple tasks for park operation and maintenance.

[0128] S308. Obtain the dependencies between tasks and construct a task dependency graph based on the dependencies. It should be further explained that the method for constructing the task dependency graph in this embodiment is as follows: based on the calibration timestamp, approved spatial coordinates, static resource requirement attributes, and dynamic resource availability codes of each task in the task feature vector set, the Apriori algorithm is used to obtain three types of dependencies: time, resource, and space. For example, by comparing the calibration timestamps, it is determined that task A must be executed after task B is completed (time dependency); by analyzing the resource requirement and availability codes, it is determined that task C and task D share the same set of detection equipment (resource dependency); and by calculating the approved spatial coordinate distance, it is determined that the distance between the execution locations of task E and task F is less than 50 meters (space dependency). Treat each task as a node, add directed edges to nodes with dependencies, and use the analytic hierarchy process (AHP) to determine the weight coefficients for time, resource, and space dependencies as 0.4, 0.35, and 0.25, respectively. The weighted sum is then used to obtain the weight value of each directed edge. For example, the edge weight between task A and task B is 0.8, the edge weight between task C and task D is 0.75, and the edge weight between task E and task F is 0.6. This constructs a task dependency graph.

[0129] S309. Using the coupling weights at convergence as the initial association score, and combining them with the task dependency graph, the association strength of each task relative to the current resource allocation mode is calculated using a graph attention network model. It should be further explained that the process of calculating the association strength of each task relative to the current resource allocation mode using the graph attention network model in this embodiment is as follows: A graph attention network model containing an input layer, a graph attention layer, and an output layer is constructed. The input layer receives a node feature matrix constructed based on the dynamically corrected primary scheduling mode vector and the coupling weights at convergence, as well as an adjacency matrix generated based on the task dependency graph. For example, the feature vector of a certain task in the node feature matrix is ​​[0.6, 0.84, ..., 0.24], and the edge weight value between the corresponding task and the dependent task in the adjacency matrix is ​​0.8. The node feature matrix and the adjacency matrix are input into the graph attention layer. The similarity is calculated based on the task node feature vector and the adjacent node feature vector. The attention weight is obtained by normalization using the Softmax function. After edge weight correction, the feature vectors are weighted and aggregated to generate an aggregated feature vector. The aggregated feature vector is input to the output layer, and then processed by a fully connected layer through linear transformation and sigmoid function normalization to obtain the correlation strength. For example, after processing, the aggregated feature vector of a certain task has a correlation strength value of 0.78, indicating that the task is highly adapted to the current resource allocation mode.

[0130] S4. Based on the association strength and the spatial location identifier, generate a set of scheduling instructions, which includes joint scheduling paths for shared operation and maintenance resources and independent scheduling instructions for independent operation and maintenance resources; and convert the set of scheduling instructions into executable control instructions for the corresponding operation and maintenance resources.

[0131] In one example, based on the correlation strength of the task relative to the current resource allocation mode and combined with the spatial location identifier, a set of scheduling instructions is generated and converted into executable control instructions. The specific process includes:

[0132] S401. Based on a preset association strength threshold and the association strength of each task relative to the current resource allocation mode, all tasks are divided into a high-association task group and a low-association task group.

[0133] Specifically, different tasks differ in their resource requirements, execution locations, and time windows compared to the current resource allocation model. Direct, unified scheduling can lead to resource waste or low task execution efficiency. Tasks in highly correlated task groups are well-suited to the current resource allocation model and are suitable for joint scheduling to improve resource utilization; tasks in low-correlation task groups are poorly adapted and require independent resource allocation to ensure execution effectiveness.

[0134] In this embodiment, the preset association strength threshold is 0.6. This threshold is determined based on historical scheduling data statistics, taking into account both resource utilization and task execution priority. All tasks are iterated through, and tasks with an association strength greater than or equal to 0.6 are assigned to the high-association task group, while tasks with an association strength less than 0.6 are assigned to the low-association task group. For example, if task 1 has an association strength of 0.75, task 2 has an association strength of 0.63, task 3 has an association strength of 0.48, and task 4 has an association strength of 0.55, then tasks 1 and 2 are assigned to the high-association task group, and tasks 3 and 4 are assigned to the low-association task group.

[0135] S402. Based on the approved spatial coordinates of each task in the highly correlated task group, at least one spatial task cluster is obtained through a density-based spatial clustering algorithm.

[0136] Specifically, tasks in a highly correlated task group may be located in dispersed or concentrated areas. Concentrated tasks can be clustered into spatial task clusters, which can reduce the time spent moving maintenance resources back and forth and improve scheduling efficiency. Density-based spatial clustering algorithms can effectively identify spatially clustered tasks and avoid the clustering effect being affected by isolated tasks.

[0137] In this embodiment, the DBSCAN algorithm is used to cluster all tasks in the highly correlated task group; the neighborhood radius is set to 50 meters, which is determined based on the park's building layout and operation and maintenance radius; the minimum number of points is set to 2, meaning that at least two tasks can form a cluster within the neighborhood radius. For example, if the approved spatial coordinates of task 1 in the highly correlated task group are (100, 200), task 2 is (120, 210), task 5 is (300, 400), and task 6 is (310, 420), then task 1 and task 2 will cluster into one spatial task cluster, and task 5 and task 6 will cluster into another spatial task cluster.

[0138] S403. Based on the dynamic resource availability coding of tasks within each space task cluster, each space task cluster is divided into at least one shared resource task cluster, wherein tasks within each shared resource task cluster are executed by the same shared operation and maintenance resource.

[0139] Specifically, tasks within the same spatial task cluster may require different types of shared operation and maintenance resources. Dividing shared resource task clusters according to dynamic resource availability coding can ensure that tasks within the same cluster can be executed by the same shared operation and maintenance resource, thus avoiding resource conflicts.

[0140] In this embodiment, the dynamic resource availability code includes resource requirement identifiers such as the skill level of maintenance personnel and the type of equipment and tools. The tasks within each spatial task cluster are traversed, and the dynamic resource availability code for each task is extracted. Tasks with the same code are grouped into the same shared resource task cluster. For example, if the dynamic resource availability codes for tasks 1 and 2 within a spatial task cluster both correspond to electrician skills + multimeter + insulating tools, then they are grouped into one shared resource task cluster; the codes for tasks 7 and 8 correspond to HVAC technician + pressure gauge + wrench, then they are grouped into another shared resource task cluster.

[0141] S404. For each of the shared resource task clusters, based on the approved spatial coordinates of all tasks within the cluster, a joint scheduling path for the shared operation and maintenance resource is generated using a traveling salesman problem solving algorithm.

[0142] It should be further explained that this embodiment generates a joint scheduling path for the shared operation and maintenance resource using a traveling salesman problem solving algorithm, including:

[0143] S4041. Obtain a building path network based on a building information model. The building path network consists of nodes and edges, where nodes correspond to traffic hubs and edges correspond to traffic paths between nodes.

[0144] Specifically, the building path network is the basis for generating joint scheduling paths. It needs to accurately reflect the relationship between traffic nodes and paths within the park to ensure that the path planning conforms to actual traffic conditions.

[0145] In this embodiment, a building path network is extracted from the park based on the Building Information Model (BIM). Nodes include access hubs such as floor entrances / exits, elevator entrances, stairwells, and corridor intersections. Edges are the walkable paths between nodes, such as corridors, elevator shafts, and staircases. Basic attributes such as the length and directional restrictions of each edge are also labeled. For example, the elevator shaft edge is labeled as bidirectional with a length of 20 meters, while the corridor edge is labeled as unidirectional with a length of 50 meters.

[0146] S4042. Map the approved spatial coordinates of each task in the shared resource task cluster to the nearest node in the building path network to obtain the target node corresponding to each task;

[0147] Specifically, the approved spatial coordinates of the task are the location of the work point, which need to be mapped to the nearest accessible node in the building path network in order to perform path planning based on the path network.

[0148] In this embodiment, the Euclidean distance formula is used to calculate the distance between the approved spatial coordinates of each task and the coordinates of all nodes in the building path network, and the node with the smallest distance is selected as the target node of the task. For example, the approved spatial coordinates of task 1 are (100, 200), the coordinates of node A in the building path network are (95, 198), and the coordinates of node B are (110, 205). The calculated distance between task 1 and node A is 5.39 meters, and the distance between task 1 and node B is 11.18 meters. Therefore, task 1 is mapped to node A.

[0149] S4043. Based on the building path network, calculate the shortest travel distance between every two target nodes using the Dijkstra algorithm, and construct a target node distance matrix;

[0150] Specifically, the shortest travel distance between target nodes is the core data for solving the Traveling Salesman Problem. Dijkstra's algorithm can efficiently calculate the shortest path between nodes in a weighted graph.

[0151] In this embodiment, based on the building path network, the edge length is used as the initial weight, and the shortest travel distance between each pair of target nodes is calculated using Dijkstra's algorithm. For example, the target nodes include A, B, C, and D. The calculated shortest distances are: A to B 30 meters, A to C 80 meters, A to D 60 meters, B to C 50 meters, B to D 35 meters, and C to D 25 meters. The target node distance matrix is ​​shown in Table 1.

[0152] Table 1. Distance between nodes in the building path network

[0153]

[0154] S4044. Obtain the real-time traffic status data of each edge in the building path network, and adjust the traffic weight of the corresponding edge based on the real-time traffic status data.

[0155] Specifically, the real-time traffic status of paths within the park affects actual traffic efficiency. For example, temporary construction in corridors or elevator maintenance can cause traffic obstruction, requiring adjustments to the traffic weights of the corresponding sides.

[0156] In this embodiment, real-time traffic status data for each edge is obtained through IoT sensors and the park management system, including whether it is congested, closed, and the speed of passage. The weighting coefficient is set to 1.0 for normal traffic conditions; if the path is congested, the weighting coefficient is adjusted to 2.0; if the path is closed, the weighting coefficient is adjusted to 1000 (considered impassable). For example, if the path from node B to C is congested due to temporary construction, its weighting coefficient is adjusted from 1.0 to 2.0, and the corresponding shortest passage distance is corrected from 50 meters to 100 meters.

[0157] S4045. Based on the adjusted passage weights, recalculate the shortest passage distance between every two target nodes and update the target node distance matrix.

[0158] Specifically, the shortest travel distance is recalculated based on the adjusted travel weights, and the target node distance matrix is ​​updated to ensure that the path planning matches the real-time traffic situation.

[0159] In this embodiment, the same Dijkstra algorithm as S4043 is used, with the adjusted edge weights as input, to recalculate the shortest travel distance between every two target nodes. For example, the updated shortest distance from node B to C is 100 meters, and the shortest distance from node A to C becomes A→B→C (30 meters + 100 meters = 130 meters) due to path adjustment. The distances between the remaining nodes are recalculated according to the weight changes, forming the updated target node distance matrix.

[0160] S4046. Input the updated target node distance matrix into the genetic algorithm, and solve the traveling salesman problem through the genetic algorithm to obtain the shortest path sequence that visits all target nodes and returns to the starting node.

[0161] Specifically, the core of the Traveling Salesman Problem is to find the shortest path that visits all nodes and returns to the starting node. Genetic algorithms have global search capabilities and can efficiently solve this combinatorial optimization problem.

[0162] In this embodiment, the updated target node distance matrix is ​​used as the input to the genetic algorithm. The population size is set to 50, the number of iterations to 100, the crossover probability to 0.8, and the mutation probability to 0.1. Through selection, crossover, and mutation operations of the genetic algorithm, iterative optimization is performed step by step to finally obtain the optimal path sequence. For example, if the target nodes are A, B, C, and D, the shortest path sequence is A→B→D→C→A, with a total travel distance of 30 meters + 35 meters + 25 meters + 130 meters = 220 meters.

[0163] S4047. Map each target node in the shortest path sequence back to its corresponding task to generate the joint scheduling path.

[0164] Specifically, by mapping the target nodes in the shortest path sequence back to the corresponding tasks, a joint scheduling path for shared operation and maintenance resources can be formed, and the task execution order can be clearly defined.

[0165] In this embodiment, the nodes in the shortest path sequence are sequentially associated with their corresponding tasks. For example, if node A corresponds to task 1, node B corresponds to task 2, node D corresponds to task 8, and node C corresponds to task 7, then the joint scheduling path is task 1 → task 2 → task 8 → task 7 → return to the starting point. At the same time, the execution position of each task and the travel path between the nodes are marked.

[0166] S405. For each task in the low-association task group, based on the approved spatial coordinates and dynamic resource availability code of the task, generate an independent scheduling instruction for the independent operation and maintenance resource.

[0167] Specifically, tasks in low-association task groups are poorly adapted to the current resource allocation mode or require dedicated resources for execution. Generating independent scheduling instructions can ensure that tasks receive dedicated resource support and guarantee execution quality.

[0168] In this embodiment, for each task in a low-association task group, its approved spatial coordinates and dynamic resource availability code are extracted. Based on the approved spatial coordinates, the travel path and work location of the maintenance resources are determined. Based on the dynamic resource availability code, the required independent maintenance resource type is determined, generating an independent scheduling instruction containing resource allocation information, work location, and execution time window. For example, if the approved spatial coordinates of Task 3 are (200, 300), and the dynamic resource availability code corresponds to a low-voltage technician and a network tester, then the independent scheduling instruction would be to allocate one low-voltage technician and one network tester to perform a network maintenance task at location (200, 300) between 10:00 and 11:00.

[0169] S406. Aggregate all the joint scheduling paths and all the independent scheduling instructions to generate the scheduling instruction set.

[0170] Specifically, all joint scheduling paths and independent scheduling instructions are aggregated to form a complete set of scheduling instructions, providing a unified basis for subsequent conversion into executable control instructions.

[0171] In this embodiment, the joint scheduling paths corresponding to all shared resource task clusters and the independent scheduling instructions corresponding to all tasks in low-association task groups are collected and integrated into a scheduling instruction set after being sorted by task priority. For example, the scheduling instruction set contains 2 joint scheduling paths and 3 independent scheduling instructions, specifying the operation and maintenance resources, task sequence, execution path and time requirements corresponding to each instruction.

[0172] Specifically, the set of scheduling instructions is a scheduling scheme at the logical level, which needs to be converted into control instructions that can be directly executed by operation and maintenance resources to ensure that the scheduling scheme is implemented.

[0173] In this embodiment, for shared maintenance resources, the joint scheduling path is converted into control commands containing navigation information, task execution order, and resource usage specifications, and pushed to the maintenance personnel's terminals via the park maintenance APP. For independent maintenance resources, the independent scheduling commands are converted into device start commands, resource unlock commands, etc., and sent to the corresponding devices via the IoT platform. For example, the joint scheduling path is converted into executable commands that navigate to task 1 location (100, 200), perform an electrical maintenance task, and then proceed to task 2 location (120, 210) according to the path after completion, keeping the tools powered on during resource usage.

[0174] This application, based on Building Information Modeling (BIM) and technologies such as the Internet of Things (IoT) and Artificial Intelligence (AI), constructs a fully intelligent multi-task collaborative scheduling solution for park operation and maintenance. In particular, it achieves a comprehensive improvement in operation and maintenance efficiency, resource utilization, and service quality through multi-dimensional technological innovation. The specific derivation is as follows:

[0175] From the perspective of data collection and feature construction, this application collects execution feature data and spatial location identifiers of pending maintenance work orders through multiple channels. After standardized processing such as timestamp calibration, spatial coordinate verification, and dynamic resource availability coding generation, a task feature vector set is constructed by combining environmental impact coefficient and priority score. This effectively solves the problems of heterogeneous data sources, inaccurate spatiotemporal information, and lagging resource status in traditional maintenance. The time deviation detection and alignment calibration mechanism ensures the accuracy of task timing. The sparsity and update suppression strategy of dynamic resource availability coding avoids scheduling instability caused by frequent fluctuations in resource status. The introduction of environmental impact coefficient enables accurate prediction of task time consumption, providing a comprehensive and reliable data foundation for subsequent scheduling decisions, and enabling scheduling instructions to fit the actual maintenance scenario of the park.

[0176] In the modality mapping and association strength calculation stage, spatiotemporal attributes and resource demand attribute sub-vectors are extracted through a modality recognition network model. These sub-vectors are then used to generate a primary scheduling modality via a multi-head self-attention layer and a resource demand encoding network. The coupling weights and higher-level scheduling modality are iteratively updated, and the association strength is calculated using a task dependency graph and a graph attention network model. This achieves precise quantification of the adaptability of task and resource allocation patterns. The multi-head self-attention layer effectively captures the spatial coupling relationship between tasks, while the resource demand encoding network normalizes and compresses the dimensionality of resource demand features. The iterative update mechanism enables the resource allocation pattern to dynamically adapt to real-time task states. The task dependency graph and graph attention network fully consider the temporal, resource, and spatial dependencies between tasks, avoiding resource conflicts and path redundancy caused by isolated scheduling. This ensures that highly correlated tasks can be accurately clustered, providing a scientific basis for joint scheduling.

[0177] In the scheduling instruction generation and execution phase, high- and low-association task groups are divided based on their correlation strength. Spatial task clusters are clustered using the DBSCAN algorithm, and shared resource task clusters are further divided using dynamic resource availability coding. A joint scheduling path is then generated using a traveling salesman problem algorithm, while independent scheduling instructions are generated for low-association tasks, ultimately converting them into executable control instructions. The generation of joint scheduling paths, through building path network construction, real-time traffic weight adjustment, and genetic algorithm optimization, achieves optimal path planning for shared maintenance resources, significantly reducing round-trip time. Independent scheduling instructions ensure dedicated resource support for low-adaptability tasks, avoiding execution delays caused by resource contention. This differentiated scheduling mode improves the utilization rate of shared resources and ensures the execution quality of individual tasks. Furthermore, instruction push notifications via mobile apps and IoT platforms enable rapid implementation of the scheduling solution, addressing the pain points of traditional manual scheduling, such as low efficiency, unreasonable paths, and uneven resource allocation.

[0178] Furthermore, this application breaks down the information silos of traditional operations and maintenance through digital and intelligent means, achieving comprehensive collaboration among personnel, equipment, space, and tasks. The combination of joint scheduling and independent scheduling reduces the cost of configuring operations and maintenance personnel, the efficient use of shared resources reduces the idle waste of equipment and tools, and accurate path planning and time estimation improve the speed of operations and maintenance response. At the same time, the full traceability and data accumulation of the scheduling process provide data support for the continuous optimization of park operations and maintenance, helping the park achieve its core goals of cost reduction and efficiency improvement, energy conservation and consumption reduction, improving the level of refinement of operations and maintenance management and service quality, and providing the park with a safe, comfortable, and efficient operations and maintenance environment.

[0179] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A method for dynamic allocation of spatial resources based on AI models and BIM operation and maintenance, characterized in that, include: Obtain execution feature data of multiple tasks to be processed within the target area and spatial location identifiers corresponding to each task, and construct a task feature vector set based on the execution feature data; The task feature vector set is input into the modality recognition network model to map multiple primary scheduling modes; wherein, the spatiotemporal attributes in the task feature vector are encoded as the spatial coupling components of the primary scheduling modes, and the resource requirement attributes in the task feature vector are encoded as the resource requirement components of the primary scheduling modes. The coupling weights are iteratively updated between the multiple primary scheduling modes and the higher-level scheduling modes, and convergence is determined when a preset convergence condition is met. The current resource allocation mode of the system is determined based on the converged higher-level scheduling mode, and the correlation strength of each task relative to the current resource allocation mode is determined based on the converged coupling weights. Based on the association strength and the spatial location identifier, a set of scheduling instructions is generated, and the set of scheduling instructions is converted into executable control instructions for the corresponding operation and maintenance resources; the set of scheduling instructions includes joint scheduling paths for shared operation and maintenance resources and independent scheduling instructions for independent operation and maintenance resources; The mapping yields multiple primary scheduling modes, including: Extract the spatiotemporal attribute sub-vector set and the resource requirement attribute sub-vector set from the task feature vector set; A spatiotemporal attribute matrix is ​​constructed based on the spatiotemporal attribute sub-vector set. The spatiotemporal attribute matrix is ​​then input into a multi-head self-attention layer, and a weighted spatiotemporal feature matrix is ​​obtained by calculating the spatial correlation between tasks. Based on the weighted spatiotemporal feature matrix, a linear transformation is performed through the first fully connected layer to output the spatial coupling component matrix; A resource demand attribute matrix is ​​constructed based on the resource demand attribute sub-vector set. The resource demand attribute matrix is ​​then input into a resource demand encoding network consisting of at least one fully connected layer. The matrix is ​​transformed by a nonlinear activation function to output a resource demand component matrix. The row vector corresponding to each task in the spatial coupling component matrix is ​​concatenated with the row vector corresponding to the task in the resource demand component matrix to obtain the primary scheduling mode vector of each task. The primary scheduling mode vectors of all tasks are aggregated to obtain the multiple primary scheduling modes.

2. The method for dynamic allocation of spatial resources based on AI models and BIM operation and maintenance as described in claim 1, characterized in that, The high-level scheduling mode refers to a typical scheduling mode that is abstracted from the characteristics of multiple tasks through a mode recognition network, representing the current global optimal or most urgent resource coordination method. Based on the execution feature data, a task feature vector set is constructed, including: Get multiple pending operation and maintenance work orders, where each pending operation and maintenance work order is associated with at least one original timestamp, a static space identifier and a set of static resource requirement attributes; The original timestamp associated with each pending maintenance work order is calibrated to obtain the calibration timestamp of each pending maintenance work order; Based on the static spatial identifier associated with each pending maintenance work order, real-time location awareness data is obtained, and the real-time location awareness data is compared with the static spatial identifier to obtain the approved spatial coordinates and location confidence of each pending maintenance work order. Based on the static resource requirement attributes of each pending maintenance work order, the real-time availability status of the corresponding resource entity is obtained, and a dynamic resource availability code is generated.

3. The method for dynamic allocation of spatial resources based on AI models and BIM operation and maintenance as described in claim 2, characterized in that, Constructing a task feature vector set based on the aforementioned execution feature data further includes: Based on the calibration timestamp and approved spatial coordinates of each pending maintenance work order, the environmental impact coefficient of each pending maintenance work order is calculated using a pre-trained environmental impact assessment model. Based on the theoretical estimated working hours in the static resource requirement attributes of each pending maintenance work order, and combined with the environmental impact coefficient, the weighted estimated working hours of each pending maintenance work order are calculated. Based on the calibration timestamp, fault level, weighted estimated working hours, and dynamic resource availability code of each pending maintenance work order, a recommended priority score is calculated for each pending maintenance work order according to the preset priority calculation rules. The calibration timestamp, approved spatial coordinates, location confidence, dynamic resource availability coding, weighted estimated working hours, and recommended priority score of each pending maintenance work order are integrated to form the maintenance feature vector of the pending maintenance work order, and a task feature vector set is generated.

4. The method for dynamic allocation of spatial resources based on AI models and BIM operation and maintenance as described in claim 3, characterized in that, Determining the correlation strength of each task relative to the current resource allocation mode includes: Initialize a higher-level scheduling mode vector and the coupling weights between each primary scheduling mode and the higher-level scheduling mode vector; Obtain the latest real-time dynamic status data for each task, and calculate the dynamic status adjustment factor based on the real-time dynamic status data; The corresponding primary scheduling modes are modified element-wise using the dynamic state adjustment factor to obtain the set of dynamically modified primary scheduling modes. Based on the dynamically corrected primary scheduling mode set and the higher-level scheduling mode vector, the coupling weights are updated using the Softmax function; Based on the updated coupling weights, the dynamically corrected primary scheduling mode set is weighted and summed to update the higher-level scheduling mode vector.

5. The method for dynamic allocation of spatial resources based on AI models and BIM operation and maintenance as described in claim 4, characterized in that, Determining the correlation strength of each task relative to the current resource allocation mode further includes: Calculate the change between the high-level scheduling modal vector after the current iteration and the high-level scheduling modal vector after the previous iteration. If the change is less than a preset convergence threshold, it is determined to be converged. The high-level scheduling mode vector at the time of convergence is determined as the current resource allocation mode of the system. Obtain the dependencies between tasks, and construct a task dependency graph based on the dependencies; Using the coupling weights at convergence as the initial association scores, and combining them with the task dependency graph, the association strength of each task relative to the current resource allocation mode is calculated using a graph attention network model.

6. The method for dynamic allocation of spatial resources based on AI models and BIM operation and maintenance as described in claim 5, characterized in that, The set of generated scheduling instructions includes: Based on the preset association strength threshold and the association strength of each task relative to the current resource allocation mode, all tasks are divided into a high association task group and a low association task group. Based on the approved spatial coordinates of each task in the highly correlated task group, at least one spatial task cluster is obtained through a density-based spatial clustering algorithm. Based on the dynamic resource availability coding of tasks within each space task cluster, each space task cluster is divided into at least one shared resource task cluster, wherein tasks within each shared resource task cluster are executed by the same shared operation and maintenance resource.

7. The method for dynamic allocation of spatial resources based on AI models and BIM operation and maintenance as described in claim 6, characterized in that, The generated scheduling instruction set also includes: For each shared resource task cluster, a joint scheduling path for the shared operation and maintenance resource is generated based on the approved spatial coordinates of all tasks within the cluster using a traveling salesman problem algorithm. For each task in the low-association task group, an independent scheduling instruction for the independent operation and maintenance resource is generated based on the approved spatial coordinates and dynamic resource availability code of the task. All the joint scheduling paths and all the independent scheduling instructions are aggregated to generate the scheduling instruction set.

8. The method for dynamic allocation of spatial resources based on AI models and BIM operation and maintenance as described in claim 7, characterized in that, The step of generating a joint scheduling path for the shared operation and maintenance resource using the Traveling Salesman Problem algorithm includes: A building path network is obtained based on a building information model. The building path network consists of nodes and edges, where nodes correspond to traffic hubs and edges correspond to traffic paths between nodes. Map the approved spatial coordinates of each task in the shared resource task cluster to the nearest node in the building path network to obtain the target node corresponding to each task; Based on the building path network, the shortest travel distance between every two target nodes is calculated using Dijkstra's algorithm, and a target node distance matrix is ​​constructed.

9. The method for dynamic allocation of spatial resources based on AI models and BIM operation and maintenance as described in claim 8, characterized in that, The step of generating a joint scheduling path for the shared operation and maintenance resource using the Traveling Salesman Problem algorithm also includes: Obtain real-time traffic status data for each edge in the building path network, and adjust the traffic weight of the corresponding edge based on the real-time traffic status data; Based on the adjusted passage weights, the shortest passage distance between every two target nodes is recalculated, and the target node distance matrix is ​​updated. The updated target node distance matrix is ​​input into the genetic algorithm, which solves the traveling salesman problem to obtain the shortest path sequence that visits all target nodes and returns to the starting node. Each target node in the shortest path sequence is mapped back to its corresponding task to generate the joint scheduling path.

Images