Campus Internet of Things terminal scheduling method based on digital twinning

Abstract

The invention relates to a campus Internet of Things terminal scheduling method based on digital twinning. The method comprises the following steps: defining a general twinning state representation structure; introducing a state mapping relationship, and establishing a one-to-many dynamic mapping relationship between a physical state and a virtual twinning state by using a finite state machine; when the state of the physical terminal changes, triggering multi-level response mapping of the twin in the virtual space; semantic fusion is carried out on twinborn real-time data, a scheduling sensitivity index is introduced, disturbance simulation is carried out based on past scheduling behaviors, and the potential influence of a certain scheduling behavior on the whole system is analyzed; carrying out local division on the situation map; using Bayesian estimation and Monte Carlo combination to sample and evaluate the feasibility and income of each solution set in a period of time in the future; simulating the expected influence of each scheduling strategy through a twin body, and selecting a scheduling path; unified modeling and semantic fusion of multiple types of terminals are realized, the intelligent level and predictive ability of scheduling decision are improved, the collaborative scheduling ability of the terminals is enhanced, and coupling unit level regulation is realized.

Description

Technical Field

[0001] The present invention relates to a campus Internet of Things terminal scheduling method, specifically a campus Internet of Things terminal scheduling method based on digital twins. Background Art

[0002] Although several technical approaches have attempted to integrate digital twin technology with the Internet of Things (IoT) to optimize intelligent management scenarios such as campus safety, teaching, and environmental control, such as the School Safety Smart Management System proposed in Publication No. CN202211266391.2, which integrates multiple modules, including servers, data collection, early warning analysis, and emergency response, to build a foundational system for campus safety management, numerous deficiencies and potential drawbacks exist in both practical and technical architecture, limiting the true intelligent and in-depth application of digital twins in campus IoT scheduling. First, existing approaches to digital twin applications focus on static data visualization and event monitoring, lacking dynamic scheduling and control capabilities. For example, while CN202211266391.2 incorporates multi-source data fusion mechanisms such as AI video analysis, public opinion perception, and an electronic sandbox, these are primarily used for post-event early warning and risk response, rather than for behavior prediction, coordinated simulation, and dynamic resource allocation based on the real-time operational status of IoT terminals. Consequently, its digital twin functionality remains largely at the early stages of information mirroring and multi-source data integration, making it difficult to achieve proactive, closed-loop scheduling optimization at the device level.

[0003] Secondly, current methods are still crude in terms of data structure design and model-driven capabilities. Although the invention involves the aggregation and management of multiple data types, including behavioral data, learning data, and course data, it lacks a unified terminal state vector abstract model and a behavioral graph structure for expressing terminal collaborative relationships, making it difficult to support unified modeling and collaborative scheduling of large-scale heterogeneous terminals. Furthermore, it lacks the dependency reasoning capabilities of graph neural networks (such as subgraph convolutional neural networks), and is unable to intelligently identify and generate linkage scheduling strategies for different devices, such as air conditioning and lighting, access control and monitoring, and environmental sensing and ventilation. Therefore, the system struggles to implement flexible policy division and resource allocation for different floors of the teaching building and for different scenarios (such as exam weeks, energy-saving weeks, and emergency drills). Furthermore, existing systems lack pre-execution prediction mechanisms and disturbance simulation capabilities for scheduling policies. This makes it easy for misjudgments or execution errors in scheduling policies to lead to resource mismatches, a degraded user experience, and even security risks. This is especially true in high-use areas (such as large lecture halls or dormitories), where resource fluctuations, device conflicts, and network bottlenecks caused by policies cannot be assessed. Scheduling control becomes a blind operation, and system stability is difficult to ensure.

[0004] Furthermore, while the invention incorporates a comprehensive structure for safety education and emergency response, these response processes still rely heavily on manual triggering, lacking a closed-loop mechanism for system self-perception, autonomous judgment, and dispatch control. For example, emergency response relies heavily on passive responses via an electronic sandbox and server commands, without incorporating self-evolutionary capabilities such as digital twin prediction, simulation, path deduction, and strategy drill scoring. Another key issue is the lack of state deviation verification and dynamic twin correction mechanisms. Summary of the Invention

[0005] The purpose of the present invention is to provide a campus Internet of Things terminal scheduling method based on digital twins, so as to solve some of the drawbacks and shortcomings pointed out in the background technology.

[0006] The present invention solves the above-mentioned technical problems by adopting the following technical solutions: a campus IoT terminal scheduling method based on digital twins, comprising: defining a universal twin state representation structure, abstracting the operating status of different types of terminals into a unified data vector form including operating status code, power status, occupancy status, and functional parameters; introducing a state mapping relationship and using a finite state machine to construct a one-to-many dynamic mapping relationship between physical state and virtual twin state; when the physical terminal state changes, triggering the multi-level response mapping of the twin in the virtual space; The real-time twin data is semantically fused to construct a situational map based on region, time, and device type. A scheduling sensitivity index is introduced to simulate disturbances based on past scheduling behaviors, generating potential impact vectors corresponding to each scheduling behavior. The situational map is then locally partitioned based on these potential impact vectors, and multiple sets of locally feasible scheduling solutions are dynamically constructed based on the current environmental state. For each locally feasible scheduling solution, a combination of Bayesian estimation and Monte Carlo methods is used to jointly evaluate the feasibility probability and expected benefit within the future target period. Based on these evaluation results, a scoring function is constructed, and the digital twin is used to simulate each scheduling strategy and select a scheduling path.

[0007] Furthermore, the selection of the scheduling path includes the following steps: S1. Generate twin dependency relationships using graph structure mining algorithms based on historical behavior data and real-time communication; S2. Use the subgraph convolutional neural network to infer the dependency strength, identify the coupled scheduling units, and obtain the coupled scheduling unit set; S3. Based on the set of coupling scheduling units obtained in S2, perform coordinated scheduling according to the relationship between the master terminal and the sub-terminal, so that the master terminal drives the corresponding sub-terminal; S4. During the collaborative scheduling process, the deviation between the physical terminal behavior and the expected behavior of the twin is compared in real time; if the deviation exceeds the threshold, the deviation identification model is activated to determine the cause of the drift.

[0008] Furthermore, the potential impact analysis method on the entire system includes: Establish a digital twin of each terminal in the campus Internet of Things, which is used to reflect the operating status, environmental parameters and functional behavior of the physical terminal; Build a unified twin state expression model to abstract the operating status of different types of terminals into standard state vectors; collect and integrate the historical scheduling behavior data of the twins to generate a system situation map; Introducing a scheduling sensitivity index to simulate disturbances in various scheduling behaviors and predict the potential impact of scheduling behaviors on the overall system operation status; screening scheduling priorities based on sensitivity analysis results and dynamically adjusting scheduling strategies; The scheduling plan is synchronized to the physical terminal through the twin for execution.

[0009] Furthermore, the scheduling sensitivity indicators include: resource fluctuation sensitivity, functional coupling sensitivity, behavior chain impact and system stability risk value; the disturbance simulation includes: simulating the execution of specific scheduling behaviors in the digital twin space, tracking the state changes of multiple terminal twins, and evaluating the impact trends on regional load, terminal response delay, and device coupling behavior.

[0010] Furthermore, the system situation map is constructed based on three dimensions: time, space and device type, reflecting the current system resource occupancy, event density and device activity.

[0011] Furthermore, in the digital twin space, multiple candidate scheduling strategies are constructed by combining the current state of the system, historical load patterns and resource constraints; a disturbance simulation mechanism is introduced to virtually execute each candidate solution on the twin; the system response impact of each scheduling strategy is calculated, and the solution set member with the least interference to the system is selected and executed on the physical device system.

[0012] After scheduling, the feedback state of the actual physical terminal may deviate from the expected state of the twin. This deviation may be caused by hardware response delay, environmental interference, and network instability. To ensure system accuracy and stability, a state deviation comparison mechanism is introduced: the twin simulation output state and the physical terminal feedback state are continuously monitored; when the deviation exceeds the set threshold, the deviation identification and compensation module is triggered to dynamically calibrate the twin model, scheduling parameters, or terminal instruction set. When evaluating the system impact of the scheduling strategy, the following disturbance response integral function is designed to measure the response strength of a certain scheduling behavior to the overall system:

[0013] in: :Indicates a scheduling strategy from time arrive The response value of the cumulative disturbance intensity caused by the system; : The current resource occupancy rate of the scheduling unit (such as energy consumption and bandwidth), reflecting the operational weight of the scheduling behavior; : System multi-terminal coupling function, reflecting the intensity of dependent interaction between devices; : The change range of the current state of a single terminal; : Indicates the impact rate of device state changes on the overall coupling structure and is a core sensitivity indicator; : The current stability coefficient of the system, derived from the system scheduling load curve; : The instantaneous deviation between the twin state and the physical state, i.e., the twin-physical difference; : The dynamic weight term that constitutes the state deviation. When the system stability is low and the deviation is large, the value of this term is rapidly amplified; the integral interval : Represents the time window of the scheduling behavior during simulation or actual execution.

[0014] Furthermore, the method for identifying a coupling scheduling unit includes: S1. Build digital twins of each terminal in the campus IoT system and generate a terminal interaction graph based on historical behavior data, with each digital twin as a graph node and the behavioral interactions between devices as graph edges. S2. Divide the terminal interaction graph into multiple subgraphs according to space, function or task frequency; S3. Perform dependency strength inference on the terminal nodes in each subgraph based on the subgraph convolutional neural network to generate a dependency matrix between devices; based on the dependency matrix, identify terminal combinations with high correlation and define them as coupling scheduling units; S4. The coupling scheduling unit is used as a minimum scheduling unit for unified scheduling, so that the terminals are controlled in a linked manner.

[0015] Furthermore, the terminal interaction graph is generated based on the following data: terminal operation logs, communication events, task collaboration history, spatial location relationships and user usage behavior sequences; the subgraph division basis includes: terminal location area, function type similarity, and behavior interaction frequency.

[0016] Furthermore, the digital twins are used as graph nodes to construct an edge structure based on behavioral temporal sequence, spatial adjacency, and task collaboration to form an initial device interaction graph. The global graph is divided into subgraphs by region or function to extract behavioral association patterns of local device clusters.

[0017] Subgraph convolution is performed on the subgraph to extract the embedded representation of each node in the local structure, which is used to express its behavioral characteristics. Based on the graph embedding results, a custom function is used to quantify the potential dependencies between any two terminals, forming a coupling strength matrix that can be used for scheduling judgment. To express the strength of behavioral collaboration between nodes within the subgraph, the following function is used to quantify the behavioral dependency between any two terminals:

[0018] in: : Indicates a digital twin terminal and The value of the strength of the synergistic dependence in the behavior map; the integration interval : represents the terminal and terminal The boundaries of active behavior periods are automatically learned from historical data; :terminal At the moment The main feature components of the embedding vector reflect its behavioral characteristics; :terminal The behavior influence function of The degree of signal impact that may be generated on neighboring nodes at any given moment; : Node pair In time The graph structure density factor is , where the larger the value, the more common neighbors the two have in the graph and the more complex the collaborative background. : Indicates the matching factor between terminals in terms of functional semantics, which is a static adjustment constant; : Indicates the terminal and The rate of change of state correlations in time evolution, i.e., the dynamics of behavioral isotropy.

[0019] Furthermore, the combination of the coupling scheduling unit includes: lighting system and teaching equipment, access control system and monitoring equipment, environmental sensing equipment and ventilation system; the coupling scheduling unit executes linkage strategies for the units, including unified startup, joint energy-saving control or interlocking safety response.

[0020] This invention achieves intelligent, refined, and adaptive control of campus IoT terminal scheduling and management by introducing key technologies such as digital twin modeling, graph neural network dependency reasoning, situational map perception, disturbance simulation prediction, and feedback closed-loop control. Compared with traditional scheduling methods based on rules, static strategies, or distributed single-point control, this invention has the following beneficial effects: The present invention constructs a universal twin state expression model, breaking the integration difficulties caused by the non-uniform types and incompatible protocols of traditional equipment, so that different equipment such as air conditioning, lighting, sensors, access control, and monitoring can be connected to the scheduling platform with a unified data structure, significantly improving the scalability and maintenance efficiency of the system; by introducing scheduling sensitivity indicators and disturbance response simulation mechanisms, the system can predict the comprehensive impact of a certain strategy on overall resource occupancy, equipment coordination, behavioral chain reaction and system stability before scheduling, avoiding chain failures and energy waste caused by blind scheduling, and realizing the transition from responsive control to predictive scheduling; through subgraph convolutional neural network The network extracts the dependencies between terminals and constructs coupled scheduling units (such as air conditioning + lighting, access control + cameras, sensors + ventilation systems, etc.) to achieve coordinated control of multiple devices, so that the devices no longer work independently, but form an overall collaborative response based on tasks, scenarios and behaviors, thereby improving control efficiency and scenario adaptability; through the state deviation identification and compensation module, the system can compare the actual scheduling execution results with the expected results of the twin simulation in real time. Once the deviation is found, it can automatically analyze the cause (such as equipment failure, network delay, environmental interference) and trigger the compensation strategy, effectively ensuring the reliability of the scheduling closed loop and the robustness of the system operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a diagram showing the unified scheduling function relationship of campus Internet of Things terminals based on digital twins in the present invention.

[0022] Figure 2 This is the multi-dimensional twin scheduling flow chart of the present invention.

[0023] Figure 3 This is a structural diagram of the dynamic collaborative scheduling system of the present invention.

[0024] Figure 4 This is a simplified structural diagram of the digital twin intelligent scheduling system of Teaching Building A according to an embodiment of the present invention.

[0025] Figure 5 The figure is a schematic diagram of the intelligent scheduling process based on interaction graph and subgraph reasoning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0027] Combined with attachment Figure 1 In this paper, in order to realize the efficient and unified scheduling and control of various terminals in the campus Internet of Things (such as smart lighting, access control, air conditioning, monitoring, environmental sensors, etc.), a scheduling method based on digital twins is proposed. The core lies in constructing a universal twin state representation structure and establishing a one-to-many dynamic mapping relationship between physical states and virtual twins to support the unified modeling, identification, response and optimization of complex heterogeneous equipment by the scheduling system.

[0028] First, to build a unified scheduling framework, different types of IoT terminals must be abstracted into a unified data representation. To this end, the system defines a universal twin state representation structure. This structure is a vector-based state descriptor consisting of four core fields: operating status code (representing the device's current functional state, such as active, standby, or faulty), power status (such as normal power supply or low battery), occupancy status (for human-related devices, such as whether the classroom is occupied or the access control is open), and functional parameters (such as current brightness, set temperature, monitoring frequency, and switching frequency). This structure ensures that data from all terminal devices, regardless of the communication protocol or physical characteristics of the source device, is ultimately converted into a unified data format, enabling centralized analysis, comparison, and computation. Data acquisition is accomplished through a unified IoT access platform. The system periodically pulls or receives data packets from different terminals, aligns the fields, and converts them into a unified state vector. After constructing the state representation structure, the system further introduces a state mapping mechanism. This mechanism uses a finite state machine (FSM) modeling approach to define one-to-many mapping rules from the physical terminal state space to the virtual twin state space. This means that a physical state may correspond to multiple twin state levels in the digital space, depending on its context, operational history, or environmental influences. For example, when a classroom's lighting is detected as on, this state is not only reflected in the twin system as the light being on, but may also simultaneously trigger complex states in the twin environment, such as classroom occupancy, air conditioning linkage preparation, and security perception switching to active mode. To support this multi-layered response, the system binds the state vector to the state transition diagram and drives the active linkage evolution of the virtual twin state through a mapping function.

[0029] In terms of model training, to enhance mapping accuracy, the system collects a large amount of historical physical device behavior data and its corresponding twin response behavior as training samples. The sample data undergoes preprocessing, including the removal of abnormal data (such as signal loss and sensor fault values), time synchronization processing (time alignment of devices with inconsistent acquisition cycles), and discrete state standardization (uniform encoding of multiple operating states). The processed data is used to train a lightweight graph state transition model, which is essentially a converter between the state graph structure and the state vector space. Its state response strategy is optimized using reinforcement learning: the model performs simulated scheduling in the twin space and adjusts its mapping strategy based on feedback on its impact on the actual physical equipment. For example, when the twin system predicts that the air conditioner can be turned off but actually causes the room temperature to get out of control, the system will adjust the mapping path through a feedback mechanism, update the confidence coefficient of the path in the original state transition graph, and reduce the probability of triggering similar behaviors in the future.

[0030] During system operation, whenever a physical terminal reports a state change, such as an access control device switching to a card-swipe state, the system immediately resolves this state into a unified state vector and uses the state mapper to find its response path in the twin space. Subsequently, the multiple associated state layers (functional layer, behavioral layer, and security layer) in the twin are activated, triggering a series of virtual behavioral changes, such as simulating a student entering the room, preheating the projector, adjusting lighting brightness, and adjusting ambient noise sensitivity. This process exhibits a multi-level response characteristic, meaning that a single physical state can trigger feedback actions at multiple virtual levels. This allows the scheduling system to move beyond single-point command control and instead possess the intelligent response capability of perception, linkage, and evolution.

[0031] Combined with attachment Figure 2 The system collects real-time status data from each twin across multiple dimensions, including each terminal's operating parameters (such as device temperature, current, voltage, and data flow), spatial information (device location and area usage), temporal information (current time, class schedule, sleep schedule, etc.), and device type (lighting, air conditioning, cameras, access control, etc.). This data is initially aggregated by edge nodes and uploaded to the main control platform. During data preprocessing, the system performs time alignment (to address device reporting delays or different sampling frequencies), unit standardization (to unify measurement units), gap filling (using nearest neighbor or regression interpolation to fill data gaps), and noise filtering (for example, removing small fluctuations in frequently changing values). Subsequently, based on graph modeling, each twin is treated as a graph node, and connecting edges are weighted based on spatial proximity, functional coupling, and historical behavioral similarity. Ultimately, a three-dimensional situational map is constructed, structured as region, time, and device type. This map not only reflects the current device density and resource consumption in each region, but also records the evolving behavioral patterns of the device population over a specific time period.

[0032] To effectively assess potential impacts before scheduling, the system incorporates a scheduling sensitivity metric. This metric simulates historical scheduling behaviors to quantify the chain reaction of different devices or policies on the overall system. This process is based on historical data showing whether a device scheduling operation significantly changes the operating status of other devices within the system. For example, whether air conditioning scheduling causes peak power consumption, whether lighting activation interferes with security systems, or whether device concurrency causes network congestion. The system quantifies these causal perturbation relationships into a sensitivity weight matrix, which is dynamically updated and serves as a key input for pre-scheduling analysis.

[0033] The situation map is then further divided into multiple local area maps, with each sub-map representing a collection of devices within a functional area, building, or logical partition. Based on the current environmental conditions (such as temperature, crowd density, and teaching schedule), the system automatically generates multiple local feasible scheduling solution sets in each sub-map. These solution sets combine candidate scheduling strategies for terminal devices in different ways while meeting basic functional requirements. To determine whether these strategies are worth implementing in the future, the system uses a combination of Bayesian estimation and Monte Carlo methods: First, a Bayesian model is used to probabilistically model the feasibility of each strategy in a specific scenario (such as stability, failure probability, and resource consumption). Then, a large number of set scenario simulations are performed in combination with the Monte Carlo sampling method to evaluate the performance changes of each strategy under different disturbance conditions, thereby obtaining the expected return and risk value of a strategy in the future time window.

[0034] After obtaining multiple candidate scheduling solutions, the system does not directly issue scheduling instructions. Instead, it relies on digital twins to conduct strategy simulation rehearsals. Specifically, without interfering with the actual physical devices, each solution is simulated in the twin space, tracking the changes in key indicators (such as system load, device response latency, changes in user comfort, energy consumption, network fluctuations, etc.). The simulation results generate a multi-objective evaluation matrix. Based on the comprehensive score, strategy stability, and expected benefits, the system selects the optimal scheduling path, which is then issued to the actual device system for execution.

[0035] Combined with attachment Figure 3 Through the above mechanism, the present invention completes a complete set of dynamic decision-making processes in the digital twin space, including data semantic modeling → situation map construction → scheduling sensitivity analysis → solution set generation and prediction → twin simulation preferential execution.

[0036] The above Bayesian + Monte Carlo combined evaluation method includes: Random variable modeling: set up is the uncertain parameter vector describing the campus scene, which represents the real-time flow of people, indoor temperature, electricity price and network load. Establishing prior distribution , which comes from historical statistical data and environmental sensor sampling.

[0037] Bayesian Update: When the current observation is collected , including real-time CO2, temperature and humidity, and people counting, calculate the posterior:

[0038] in Use Gaussian likelihood or Poisson likelihood, and flexibly set according to data type.

[0039] Strategy return-feasibility function: Candidate strategies definition:

[0040] Where ΔE is the energy saving benefit, ΔQ is the comfort improvement, and Risk is the fault / complaint penalty. Set by the management's policy weight.

[0041] Monte Carlo sampling and expected returns: 1. From N groups of independent sampling ; 2. Calculation and constraint decision function =I{ exist The resource and security constraints are met under the conditions}; 3. Get:

[0042]

[0043] 4. Overall rating:

[0044] 5. Select and is sent to the physical terminal after verification through digital twin simulation.

[0045] Parameter settings: Sampling times , dynamically adjusted according to real-time computing power; Prior distribution: , P and Lognormal; Weight The default ratio is 0.5:0.3:0.2 and can be adjusted by the operation and maintenance strategy.

[0046] To further enhance the accuracy and stability of scheduling strategies and the efficiency of inter-device collaboration, the system designed a dependency-driven coupled scheduling mechanism. This mechanism, centered around scheduling path selection, incorporates a multi-layered technical logic encompassing graph-structured dependency modeling, neural network reasoning, master-slave linkage control, and bias feedback compensation. The key approach is to organize the large number of heterogeneous terminal devices on campus into logically coupled units through behavioral associations. These devices are then modeled, analyzed, and dynamically controlled through a twin system to achieve system-level coupled scheduling and robust response to anomalies.

[0047] In actual operation, the system first needs to build a behavioral dependency graph between devices. To this end, it collects data from two main sources: one is historical behavior data, including collaborative startup records, task sequence, response sequence, and other information of various terminals in different time periods and under different event conditions; the other is real-time communication data, including the current status linkage between terminals, communication delay, task feedback, etc. The system merges and processes this information by timestamp and forms a continuous device behavior sequence through a sliding window method. After data cleaning, outlier removal, and vectorization processing, these sequences are input into the graph mining module. By calculating characteristic indicators such as collaborative frequency, shared control source, and time correlation between node pairs, a preliminary behavioral dependency graph is established. Each node in the graph represents a digital twin terminal, and the weight of each edge represents the initial value of the potential dependency strength between the two terminals.

[0048] To improve the accuracy of identifying complex dependencies, the system further introduces a subgraph convolutional neural network (Sub-GCN) to infer the dependency strength between devices within a local area. Specifically, the system divides subgraphs into building units, functional areas, or high-frequency interaction clusters. The adjacency matrix and node feature matrix of each subgraph are then encoded and input into the Sub-GCN model for graph convolution. The model architecture consists of an input layer, two convolutional propagation layers, and an output dependency strength scoring layer. Node features include information such as device type, task frequency, energy consumption level, and recent behavior vectors. Model training uses supervised learning with labeled data, with the training objective being to fit the device collaborative response relationships observed in real-world scheduling history. The loss function is designed to integrate structural reconstruction error with prediction accuracy differences, and iterative training is performed using the Adam optimizer. The training data consists of multiple historical task execution cycles, covering typical usage scenarios such as daytime, nighttime, holidays, and peak activity, to enhance the model's generalization in real-world scheduling.

[0049] After model training is complete, the system can infer the dependency strength between terminals in real time based on the input current state subgraph and automatically identify coupled scheduling units—a group of terminals with closely coordinated behavior and highly coupled structure. These scheduling units typically have a unified operational purpose or response logic, such as classroom lighting and air conditioning, surveillance and access control, or projection equipment and curtains. During the scheduling execution phase, the system adopts a master-slave control structure, with the master terminal serving as the scheduling control hub, issuing commands to its subordinate subterminals to ensure synchronized responses according to coordination rules. The master terminal is dynamically selected based on its centrality score in the dependency graph and the stability of its past task responses, thereby improving the timeliness and coordination of the overall scheduling. After the scheduling is completed, the system compares the expected behavior of each twin with the actual feedback from the physical terminal in real time. This comparison mechanism is based on multi-dimensional state matching, including switch status, sensor readings, latency, response consistency, and other indicators. If the deviation exceeds a set threshold, the system immediately triggers the deviation identification model. The model uses multimodal inputs, including device response time, network packet transmission delays, and fluctuations in ambient environmental parameters (such as temperature, voltage, and noise). It uses a trained multi-classifier to determine the cause of the deviation. The model's output is categorized into three types: ① Terminal failure, such as device hardware anomalies or unexecuted control commands; ② Communication delays, such as delayed response due to network fluctuations; and ③ Environmental changes, such as state changes caused by non-scheduling factors, such as sudden power outages, human error, and automatic adjustments to the building's temperature control system. Based on the judgment results, the system will implement different compensation strategies, such as reissuing commands, switching to a backup terminal, adjusting the task schedule, or providing feedback to operations and maintenance personnel to issue an exception notification.

[0050] The entire mechanism forms a closed scheduling loop consisting of structural reasoning, policy issuance, deviation verification, and adaptive adjustment to anomalies, demonstrating significant intelligence, stability, and adaptability. This approach not only improves the collaborative efficiency of IoT terminal scheduling but also enhances resilience to uncertainty and sudden anomalies. It is particularly suitable for smart campus environments with diverse equipment, complex task logic, and frequent state changes. Compared to traditional timing or single-point control logic, this method boasts significant technological advancements and adaptability to different scenarios, demonstrating the predictive power and collaborative intelligence of digital twin systems in the management and control of physical entities.

[0051] Example 1: Building A of a university has six floors, each equipped with approximately 80 IoT terminals. These devices include smart lighting (LED dimming lights), central air conditioning vent control modules, infrared motion sensors, smart access control systems, environmental monitoring nodes (temperature, humidity, and CO2 concentration), electric curtains, projectors, wireless access points, and other equipment, totaling nearly 500 terminals. To optimize energy consumption and ensure a balanced teaching experience, the building is integrated into a digital twin scheduling platform for intelligent management.

[0052] The first step is to establish a digital twin. The platform generates a digital twin for each of the aforementioned physical terminals. Each twin synchronizes in real time its operating status (e.g., on / off, operating mode, load factor), functional parameters (illuminance settings, temperature settings, etc.), and environmental information (spatial location, classroom usage, and current curriculum). For example, the main light in Room 302 on the third floor, numbered L3-R302-LIGHT-01, has a twin status of: status code 1 (on), power 45.6W, preset brightness 80%, and is associated with the Data Structures class in the curriculum from 8:00 AM to 9:30 AM.

[0053] The platform then uses unified state modeling to construct a standard state vector: [operating status code, power status, occupancy status, functional parameters 1-n]. For example, the main light above is converted to [1, 1, 1, 0.8], where status code 1 indicates on, normal power, occupied, and a normalized brightness of 0.8. Status data for all terminals is collected every 5 seconds and integrated into a system status map. The map is indexed by floor, time period, and terminal type. During the current morning peak (8:00–9:00), the system detected high load rates on the third and fourth floors, with peak air conditioning power reaching 220kW, 64% of network APs approaching their upper limit, an average ambient CO2 level of 750ppm, and a total lighting power consumption of 96.4kW.

[0054] At this time, the property management platform plans to automatically turn off all air conditioners and some lighting during the class break at 9:30 to save energy. The system first performs a disturbance simulation on this pre-planned scheduling behavior, that is, analyzes the impact of this behavior on the overall operation of the system. The platform looks back at the scheduling history of the past month and extracts events caused by the unified shutdown of air conditioners under high temperature conditions (≥28°C), including an increase in complaint tickets (3.4 times), excessively fast temperature rise in the classroom (average temperature rise rate of 0.45°C / min), etc. At the same time, the simulation model predicts: under the current room temperature of 27.5°C and dense crowds (estimated to be 1.5m per person), the temperature in the classroom will rise rapidly. 2), shutting off the air conditioners uniformly would cause the comfort index to drop to a critical value of 0.62 (threshold 0.6) within 10 minutes, triggering the automatic complaint mechanism in at least four classrooms. Based on this, the platform assigned a scheduling sensitivity score of 0.88 (out of a maximum of 1, with higher values indicating greater risk) to the uniform AC shutdown action. Instead, the system selected a localized scheduling solution with a sensitivity score of 0.34: shutting off AC in empty classrooms, delaying shutdown in overheated classrooms by 2 minutes, and automatically delaying control for excessive CO2 concentrations. Based on these differences in scores, the platform automatically adjusted scheduling strategy priorities, abandoning a unified building-wide strategy in favor of a regional sensitivity-based approach, prioritizing comfort in teaching areas. All scheduling paths were simulated using a digital twin. Simulations demonstrated that this solution reduced the probability of complaints by 70% while reducing energy savings by 7%, resulting in the final implementation path.

[0055] At 9:25, the system issued a command to all master control devices through the twin. For example, the R308 air conditioning controller on the third floor switched to a pre-shutdown state and issued a 5-second countdown, waiting for the students to leave before automatically shutting off the cooling system. The device's actual behavior feedback indicated a 2-second delayed response, with the status written back as a shutdown failure. At this point, the system immediately compared the physical terminal's expected state with the twin's and discovered a deviation with a drift code of 1, triggering the deviation identification model. The model analyzed and determined that the terminal's local thermostat battery voltage was low, and the communication signal was interrupted, resulting in the command not being executed. The platform automatically initiated a compensation mechanism: switching to the backup temperature control loop and simultaneously pushing a maintenance request to the operation and maintenance system to prevent scheduling failures from further amplifying system deviations.

[0056] In the aforementioned teaching building A, after the class ends at 9:30, the system plans to implement the next round of empty classroom energy saving strategy. On this basis, the property management plan adds a new scheduling behavior: shut down the wireless AP equipment and air conditioning modules of all unreserved classrooms on the third floor to further reduce the peak energy consumption. The platform first initiates a disturbance simulation evaluation of the scheduling behavior. According to the method of the present invention, the system calculates four indicators of scheduling sensitivity in sequence. The first is resource fluctuation sensitivity, which measures the impact of scheduling behavior on the instantaneous fluctuation of key resources such as the overall energy consumption, bandwidth, and power load of the system. The platform queries the resource operation baseline data of the scheduling target area for the past 10 days and finds that the average air conditioning power consumption during normal periods is 48.7kW, and the average peak bandwidth load of AP is 36.2MB / s. After the scheduling plan is executed, the air conditioning power consumption is expected to drop by about 32.4%. However, due to the shutdown of the AP, some students may switch to the neighboring AP, resulting in a sudden increase in their bandwidth. Using simulated twin networks, the platform found that the load on four adjacent APs could increase by 25% within 5 minutes, with a total power consumption change of 17.1%. The resource fluctuation sensitivity score was set at 0.73 (ranging from 0 to 1, with higher values representing greater fluctuations, and a threshold of 0.65 serving as a warning line).

[0057] The second factor is functional coupling sensitivity. Based on the device coupling matrix learned by the previous Sub-GCN graph network model, the platform analyzes the dependencies between air conditioning shutdown and environmental monitoring devices and human motion sensing modules. The probability of linkage between the air conditioning systems and CO2 sensors in classrooms R310 and R312 is 0.86, indicating a high co-occurrence of air conditioning activation and CO2 over-limit warnings in the past. If the air conditioning is turned off, the CO2 sensor may provide abnormal feedback, triggering automatic window opening, which in turn affects the overall pressure differential stability of the central air conditioning system. Based on this, the platform sets a functional coupling sensitivity score of 0.81 and marks the presence of a behavioral linkage chain warning after the scheduling trigger.

[0058] The third item is the behavioral chain impact, which measures the extent to which a scheduling action propagates across multiple levels of states within the twin. The system uses data from twin linkage events to record data. In a test simulation of shutting down an air conditioner and access point (AP), it detected a 21% increase in response latency for human sensors triggered within 5 seconds, a 33% increase in the frequency of CO2 sensor state fluctuations, and a brief signal loss in the access control system. The system assessed that the propagation path spanned six types of devices, with an event transmission chain depth of four levels and a behavioral chain impact score of 0.79 (a threshold of 0.7 indicates a complex chain warning).

[0059] The fourth item is the system stability risk value. This value integrates information such as the state deviation trends of multiple twins after the execution of the scheduling plan, as well as feedback response delays and execution failure rates to construct a predictive risk factor. Historical platform data shows that in this type of scheduling, the instruction implementation failure rate is 3.1%, with an average response delay of 2.8 seconds. Combined with simulation results, three air conditioning units in this scenario were detected to have potential state writeback failures, with an estimated system stability risk value of 0.65. The platform sets this threshold at 0.6, indicating a moderate risk.

[0060] After combining and analyzing the four sensitivity results, the system used a weighted linear combination (setting resource fluctuation weights at 0.3, functional coupling at 0.25, behavioral chaining at 0.25, and stability at 0.2) to determine an overall sensitivity score of 0.76 for this scheduling behavior, placing it at a high risk. To further validate the impact, the system performed a complete perturbation simulation within the digital twin. After the shutdown command took effect in the digital twin, the system monitored the state change paths of the twin nodes. The system observed a rapid increase in CO2 concentration within 10 minutes in the simulations of rooms R310 and R312, a rise in the probability of projector wake-up failure to 7.2%, and a two-fold increase in the number of delay alarms due to bandwidth transfer in the nearby teaching network. The simulation predicted that this scheduling behavior would result in a 0.41 decrease in the user experience satisfaction index in actual execution. The platform determined that batch shutdowns during class breaks were inappropriate and recommended batch shutdowns based on the dual conditions of rooms idle for at least 5 minutes and the absence of respiratory signals. The system automatically reconfigured the scheduling strategy to a three-level progressive control strategy, prioritizing shutdown of terminals in low-risk rooms such as R307 and R309.

[0061] To illustrate the operational process of this solution in a real-world scenario, the following complete simulation and case analysis of the operational status of Teaching Building A is conducted during the peak teaching period from 8:00 AM to 10:00 AM on Monday, June 10, 2024. The system first collects operational data from all twin terminals, covering a total of 483 active terminals, at a 5-second interval. This data includes operational status codes, power consumption, device type identification, location information, terminal latency, and environmental perception values (such as temperature, humidity, light intensity, and CO2 concentration). After preprocessing, standardization, time alignment, and outlier removal, the system projects the data into a multidimensional situation map construction model.

[0062] The first dimension is timeline modeling. The system divides data into 120 time segments at a minute granularity, recording dynamic indicators such as the number of operations, fault alarms, average power consumption, and average response delay for each device category within each time segment. For example, between 8:45 and 9:00, the system recorded that 93 of the 122 terminals on the third floor were active (76.2% activity), the average power consumption of the air conditioning modules was 62.8kW, the average response delay of the lighting system was 0.9 seconds, and the access control card swipe success rate was 98.7%, a 3.2 percentage point increase from 8:00 to 8:15. The system marks this as a high-activity, high-load period.

[0063] The second dimension is spatial axis modeling. The system divides the teaching building into six floors, which are further subdivided into three types of spatial modules: left and right corridors + classroom units + public areas, with a total of 32 sub-areas. The resource usage of each area is visualized in the form of a heat map. The numerical values are derived from indicators such as the total power load of the area, the LAN bandwidth utilization rate, and the estimated crowd density (based on the number of infrared + access control card swipes). For example, at 9:00, the power load in the classroom area on the east side of the fourth floor reached 112.3kW, the highest in the entire building. The local environmental CO2 average was 815ppm, exceeding the recommended upper limit (800ppm). The crowd density was estimated to be 1.2 people / ㎡. The system marked the area as an overloaded high-density area and prompted that the high-power scheduling strategy could not be executed.

[0064] The third dimension is modeling the device type axis. The system categorizes all devices and calculates their type activity, response volatility, communication stability, and task frequency. For example, during the aforementioned time window, there were 91 air conditioning devices, 68 of which were active, with an average response delay of 1.3 seconds and three reported faults (control anomalies of R311, R408, and R503). The system marked the air conditioning system as operating in a medium-risk state. There were 183 lighting devices, with an activity rate of 91.2%, but low power consumption fluctuation, indicating strong stability and suitability as a candidate for load peak shaving scheduling.

[0065] Combining these three dimensions, the system constructs a three-dimensional system status map: time (T) × space (S) × device type (D). Each node is annotated with three core metrics: resource utilization, event density, and device activity. A clustering algorithm (based on DBSCAN) is then used to automatically identify highly sensitive sub-blocks and low-load scheduling candidates. For example, the system detected high activity but stable resource utilization (low power consumption and low latency) in the lighting system in the west corridor on the second floor between 8:50 and 9:10 AM, marking it as a priority scheduling zone. However, the air conditioning system in the R310-R312 area on the third floor was designated a scheduling restriction zone due to frequent fluctuations and abnormal ambient CO2 levels.

[0066] To further validate the graph's decision-making capabilities, the platform implemented a lighting energy-saving strategy, temporarily shutting off all unused corridor lights for 10 minutes at 9:05 AM. By querying the situation graph, the system quickly identified the west side of the second floor and the east side of the fifth floor as low-risk areas. The corresponding state graph node scores were resource usage ≤ 0.25, device activity < 0.4, and event density < 1 (with zero alarms in the past 10 minutes). The system immediately generated scheduling instructions and completed a simulation exercise using the digital twin. Simulation results showed that shutting off the lights in these areas reduced system power consumption by approximately 8.6%, with no significant impact on personnel flow or classroom equipment. The event risk score remained at 0.17, well below the intervention threshold of 0.45.

[0067] The background of the embodiment is set as June 10, 2024. The fifth floor of Building A of the teaching building is scheduled to perform energy peak shaving scheduling tasks between 10:30 and 11:30 to alleviate the pressure of the total power of the floor approaching the threshold (monitored at 128.5kW, close to the limit of 130kW). The scheduling goal is to shut down the redundant lighting and non-essential air conditioners in the six classrooms (R501~R506) on this floor. The system first generates candidate scheduling strategies in the digital twin, taking into account factors including the current system resource status (real-time power, grid load), historical load change patterns (energy consumption records during the peak teaching period in the past three weeks), and equipment constraints (such as some classrooms need to retain projection and air conditioning for special teaching tasks). A total of 4 groups of scheduling strategies are generated: Strategy A: Turn off the lights and air conditioning in all classrooms; Strategy B: Turn off lighting only in unused classrooms; Strategy C: Turn off all air conditioning but keep lighting on; Strategy D: Turn off the lights and some air conditioning vents in each classroom as needed.

[0068] The system substitutes these four strategies into the twin space for disturbance simulation. The duration window of each simulation is 10 minutes, that is, the integration interval During the simulation, the platform collects dynamic variables and inputs disturbance response functions based on the following parameters: resource utilization :For example, in strategy A, the total instantaneous power of air conditioning + lighting is 67.2kW, accounting for 0.52 of the total load; the coupling function :Dynamic values are calculated based on the lighting-air conditioning and lighting-human body induction linkage diagrams, with a maximum of 1 (complete coupling). The instantaneous value of strategy A in the simulation is 0.76; state changes : Under policy triggering, the average amplitude of terminal power / response state fluctuation is ±0.3 (normalized value); stability coefficient : Extracted through regression fitting of the system stable state, the mean value in the scheduling window fluctuates between 0.45 and 0.9; state deviation : It is determined by the degree of deviation between the twin simulation state and the historical expected state, ranging from 0.01 to 0.18.

[0069] Substituting these data into the disturbance response integration function:

[0070] in, The state evolution graph in the twin system shows that the value is approximately 0.67 in strategy A, indicating that a small state change causes a large coupled structural disturbance. The system is calculated using the discrete integration method, and the following results are obtained: Strategy A disturbance response value ; Strategy B disturbance response value ; Strategy C disturbance response value ; Strategy D disturbance response value ; The results show that Strategy D has the smallest disturbance response value, indicating that it has the least impact on system operation and achieves the best balance between resource release efficiency and system stability. Based on this, the platform selects Strategy D for execution and sends it to the physical terminal for execution. After it is actually sent, the twins synchronously receive feedback data and monitor the twin-physical state deviation in real time: The response of the classroom R503 air conditioner controller was delayed by 6 seconds, and the state feedback value was partially invalid; the system compared the twin's expected state (off) with the actual feedback (still running), resulting in a deviation. =0.12, exceeding the preset threshold of 0.08; the deviation recognition model is triggered, and combined with the anomaly detection results, it is determined that the communication delay is caused by network packet loss; the platform immediately activates the compensation mechanism: it issues instructions through the backup control loop and adds the terminal to the next round of strategy exclusion queue to prevent system-level uncertainty from being triggered again.

[0071] The final feedback data showed that after the actual implementation of Strategy D, the power consumption dropped by 16.2%, the average response delay of the system was controlled within 1.3 seconds, and there were no other failed responses except R503. The cumulative value of the whole round of disturbance deviation was , which is lower than the system set intervention threshold (0.8), the error between the simulation prediction and the actual result is less than 6.5%, and the system stability score increases from 0.81 before scheduling to 0.91, proving that the strategy selection and compensation mechanism are effective.

[0072] Example 2: Focusing on the actual operation and maintenance scenario of Teaching Building A in Example 1, the four steps (S1-S4) of this solution are explained in detail through examples, and the feasibility and effectiveness of the steps are verified by introducing actual data and calculation results.

[0073] On the morning of June 12, 2024, the campus operations team planned to implement a multi-device coordinated energy-saving strategy scheduling task for Teaching Building A to reduce peak energy consumption, minimize communication conflicts, and avoid scheduling failures. The system first executed S1: Build a Terminal Interaction Graph. The platform retrieved the operation logs, linkage records, and physical communication data of various terminals over the past 60 days. For example, on the third floor, 135 devices were involved, including lights, air conditioners, access control systems, environmental sensors, projectors, and curtain controllers. By performing behavioral sequence mining on event log sequences, the system identified high-frequency co-occurrence relationships among multiple devices. For example, after the air conditioner was turned on, approximately 78% of classrooms automatically adjusted their lighting brightness within three minutes, indicating implicit collaborative logic. The system treats each twin as a graph node and the behavioral interactions between devices (including action simultaneity, triggering dependencies, and adjacent communication paths) as edges. This constructs an undirected weighted graph, where edge weights represent the frequency of collaboration or the weight of historical linkages.

[0074] Next, S2: Subgraph Division is executed. The platform divides subgraphs based on functional areas and behavioral clustering characteristics. Using the floor + function combination as the division criteria, the third floor is divided into four subgraphs: ① Classroom Area Subgraph (including lighting, air conditioning, and projector for classrooms R301-R306), ② Corridor Subgraph (including lighting and access control), ③ West Multimedia Lab Area Subgraph (projector, audio, and curtains), and ④ Environmental Sensing Subgraph (distributed CO2, temperature, and humidity sensors). Each subgraph has a high internal interaction frequency and functional consistency, facilitating subsequent subgraph neural network processing.

[0075] In the S3: Sub-GCN Inference Dependency Strength step, the system encodes the adjacency matrix and node attribute matrix of each subgraph. The node feature vector includes device type (one-hot), usage frequency (normalized), time interval since last linkage, unit power consumption, communication response rate, and more. The model structure is a two-layer graph convolutional network, with each layer containing 128 neurons, using the ReLU activation function and Dropout = 0.3 for regularization. The training data consists of manually annotated collaborative behavior samples (such as lighting + air conditioning linkage events and curtain + projector synchronization triggering events). The training objective is to minimize the mean squared error between the predicted collaborative probability and the actual linkage relationship between nodes. After 120 rounds of iterative training, the model achieved an accuracy of 91.2% on the validation set.

[0076] The inference results generate a dependency matrix R(n×n), where R(i,j) represents the degree of collaborative dependency between device i and device j (with a value range of [0,1]). For example, in the classroom subgraph, the system identifies the following highly dependent device combinations: R301 classroom air conditioning and lighting: dependence 0.87; R303 projector and curtain control: dependency 0.92; R304 air conditioning, lighting and access control: the average dependency strength of the triplet is 0.79; The platform set a recognition threshold of 0.75 and constructed coupled scheduling units based on the device combinations in the matrix that exceeded the threshold. Ultimately, five coupled scheduling units were identified on the third floor, including R301 air conditioning + lighting, R303 projector + curtains + CO2 sensor, and so on. Each unit was independently defined as the minimum scheduling entity.

[0077] S4: After the scheduling and control linkage execution phase begins, the system prioritizes energy-saving strategies for less sensitive areas based on floor environmental conditions and power load forecasts. For example, in classroom R301, if the number of students is detected as zero (via access control and infrared sensors), the system activates the corresponding coupling units and sends a unified shutdown signal to the lighting and air conditioning systems. Synchronous simulation feedback within the twin shows a device response time of approximately 1.2 seconds, resulting in an expected 8.3% reduction in energy consumption, which is consistent with historical data.

[0078] In classroom R303, because the next class was about to begin in 10 minutes, the platform postponed the scheduling operation, but retained the curtains in the half-open state and the projector preheating process. Using twin feedback, the curtain rail status feedback was verified to be offset by +2°. This deviation exceeded the threshold (±1.5°). The system immediately incorporated the deviation status into the Sub-GCN feedback channel and updated the curtain + projector combination stability score in the next round of training. The entire identification and scheduling process took approximately 6.3 seconds. The system logged a total of 12 scheduling operations with a 100% success rate. The coupled scheduling unit response time averaged 1.1 seconds, and energy consumption decreased by 12.6%, exceeding the 7.3% decrease in the independent scheduling mode during the same period. This demonstrates that coupled scheduling based on Sub-GCN identification of coupled units is more efficient and accurate, and also enables the twin model to possess a complete closed-loop capability of structure learning, behavior evolution, and coupled prediction.

[0079] Fast forward to the morning of June 15, 2024. To optimize the morning class scheduling process on the sixth floor of Building A, the school hopes to predict and automatically generate classroom device linkage strategies in advance. Based on the method described in this invention, the system first generates a terminal interaction graph and, relying on multiple data sources, constructs a diagram of the linkage relationships between devices. A total of 162 terminal devices are deployed on the sixth floor, distributed across four spaces: classrooms, corridors, lab areas, and public discussion areas. To form the graph structure, the platform extracts the following five types of data: 1. Terminal operation log: For example, the R605 lighting device is in high-brightness working state between 8:00 and 8:45, and the state change frequency is 0.12 times / minute; 2. Communication event records: The average communication frequency between the R602 air conditioner and the main control system is 1.5Hz, and the most recent interruption lasted 4.2 seconds; 3. Task collaboration history: Data from the past three weeks shows that the probability of the air conditioner and lights in classroom R603 being turned on simultaneously is 82%, and the probability of the curtains and projector being turned on simultaneously is 76.5%. 4. Spatial location relationship: GIS spatial data shows that R601 and R602 are only 5 meters apart, with no solid wall between them, and their device signal coverage areas overlap. 5. User behavior sequence: By integrating access control with teaching plans, it was found that R604 used the combination of air conditioning, projection, and lighting for three consecutive classes. The usage sequence was access control → air conditioning → lighting → projection.

[0080] After uniformly encoding the above data, the system constructs an undirected weighted terminal interaction graph, with nodes representing each twin terminal and edge weights representing the strength of behavioral collaboration, ranging from [0 to 1]. For example, the edge weight between R603 air conditioner and R603 light is 0.82, and the edge weight between R605 projection and curtains is 0.77. However, the edge weight between two access control devices in different classrooms (e.g., R602 access control and R606 access control) is only 0.09, indicating a weak interaction. The system sets an edge weight threshold of 0.25, and edges below this value are pruned to increase graph sparsity and focus on key behavioral pathways.

[0081] After the graph structure is completed, the subgraph division phase begins. The subgraph division in the method of the present invention is based on three types of criteria: ① Terminal location area: For example, based on the physical location of the classroom, R601 to R606 are the six classroom sub-maps, R607 to R609 are the public area sub-maps, and R610 to R612 are the corridor and staircase sub-maps; ② Functional type similarity: For example, energy-consuming devices such as air conditioners, lights, and curtains are grouped into one category, security devices such as monitoring, access control, and infrared sensors are grouped into another category, and communications and media devices (APs, projectors) form a third category. ③ Frequency of behavioral interactions: We used a co-occurrence matrix to count the number of times devices simultaneously participated in the same task and clustered them using hierarchical clustering. We found that the combination of R602 air conditioner + lights + curtains worked together 52 times in the past 12 days, with an average frequency of 4.3 times per day, which is higher than other combinations.

[0082] The system combines these three criteria, using a label propagation-based community discovery algorithm to partition the subgraph and perform a structural assessment on each subgraph. Ultimately, seven core subgraphs were obtained. Classrooms R603 and R604 each formed two highly coupled subgraphs, with average internal edge weights of 0.71 and 0.66, respectively. The corridor subgraph (containing four lighting and sensor devices) had an average internal edge weight of only 0.31, indicating weak functional connectivity, and was therefore labeled a low-coupling region by the system.

[0083] Next, the platform continues the subsequent subgraph convolutional neural network training and strategy identification process based on these subgraphs. However, at this stage, the interaction graph and subgraph partitioning system alone can quickly identify multiple areas with scheduling potential. For example: son Figure 1 (R603 classroom) The air conditioning, lighting and curtains are identified to have high coordination, and the platform marks them as coupling scheduling priority units; Figure 4 (Corridor equipment) is downgraded to a non-core subgraph due to sparse interactions, and the overall shutdown strategy is only considered in extreme energy-saving scenarios. In subgraph 6 (R610-R612), the access control + camera combination has an edge weight greater than 0.8 and is marked as a security linkage area. It does not participate in energy-saving scheduling but is used as a security automatic warning response area.

[0084] Finally, to verify the effectiveness of the subgraph partitioning strategy, the system randomly samples 5 subgraphs to perform a local scheduling strategy evaluation based on simulation preview. Figure 2 For example, in classroom R602, a simulation strategy was used to automatically shut off the air conditioner within five minutes of class and delay the curtains' retraction by five seconds. Simulation results showed that this strategy responded in 1.6 seconds, reduced power consumption by 8.2%, and reduced user experience expectations by less than 2%. Compared to the average failure rate of 4.5% in the past when scheduling was not grouped, the failure rate of this coupled unit scheduling was only 0.8%. The system concluded that this subgraph partitioning strategy significantly enhanced scheduling.

[0085] In this example, a behavioral interaction graph is constructed using digital twins as graph nodes. Data sources include three dimensions: ① Behavior time series logs: for example, the projector and main light in classroom R601 have been simultaneously activated 43 times in the past 20 days; ② Spatial adjacency: The distance between the air conditioner and lighting in R603 is less than 3 meters, and the control system logical ports share the same area ID; and ③ Task collaboration frequency: In the past two weeks' course plan, there were 19 pre-class curtain, projector, and air conditioner collaboration events, with a collaboration probability greater than 0.83. The initial edge weights are set by normalizing the collaboration frequency multiplied by spatial similarity (location proximity), with a range of [0.0, 1.0]. This results in an initial graph structure consisting of 148 nodes and approximately 2,300 edges.

[0086] The system then divides the graph into multiple subgraphs based on rules such as regional location (such as floor and functional area) and device function labels (such as air conditioning, lighting, and sensors). For example, the equipment cluster of classrooms R601-R605 is subgraph A, the access control and monitoring area of the third-floor corridor is subgraph B, and the environmental sensing and air conditioning of the experimental area is subgraph C. Each subgraph undergoes an independent graph convolution operation, with the input data being the node embedding vector (including operation frequency, recent linkage behavior score, energy consumption characteristics, and mean response delay). A two-layer Sub-GCN structure is used, with the output dimension of the first layer being 64 and the second layer being 32. During training, Dropout = 0.2, the learning rate is 0.001, and 80 rounds of iterations are used. The loss function is to minimize the dependency strength prediction error.

[0087] After the convolution is completed, the system obtains the principal component of the embedding vector of each node in the local structure , and cooperate with the behavior influence function of other nodes in the same subgraph , graph structure density factor , state-dependent derivative , substitute into the core function of the present invention:

[0088] The value ranges of each coefficient are as follows: : Node embedding principal components, range ; : Behavior affects function value, range ; : Graph density factor, value , the larger it is, the more common neighbors it has; : Function matching factor, which is a fixed value and is classified as follows: Similar equipment (e.g., light-light) = 0.2; Cross-category highly collaborative devices (such as projectors-curtains, access control-cameras) = 0.6; Weak synergy (such as access control-air conditioning) = 0.1; : The rate of change of the same direction, the empirical valuation range is ; Take the R602 projector (node i) and the R602 curtain controller (node j) as an example: , , , , , Seconds (corresponding to the past 5-minute linkage interval); Calculation results:

[0089] The system calculates the value of the preset threshold value ( Considered strongly coupled), confirming that the projector and curtains form a highly coupled scheduling unit. Similar calculations yielded a 105.3% access control and camera combination, and a 99.6% sensor and exhaust vent combination, both meeting the linkage strategy requirements.

[0090] Entering the strategy execution phase, the platform sets different strategies for these three types of coupling units: Teaching equipment (such as projectors and curtains) is set to start up uniformly, with one-click activation 5 minutes before class, and window closing, projector activation, and lighting are completed in sequence; security equipment (access control and monitoring) implements interlocking safety responses. For example, if the access control card is swiped abnormally, the camera will be immediately activated and switched to recording mode; environmental equipment (CO2 sensor and ventilation system) is configured for joint energy-saving control, and the exhaust fan will be automatically shut down when the sensing value is lower than 500ppm.

[0091] In actual implementation, the average response time of the platform recording strategy was 1.4 seconds, the abnormal response rate was 0, the power saving rate reached 11.2% in the CO2 sensing + ventilation linkage, and an abnormal access control event was captured in the security response and the camera was linked to record the video successfully, proving that the dependency graph function and scheduling unit identification method of the present invention have extremely high engineering applicability and effectiveness.

[0092] Through the above complete example, it is demonstrated how to realize intelligent linkage scheduling between campus equipment based on behavioral graph construction, subgraph convolution and coupling function integral analysis, thereby opening up the collaborative intelligent control channel between digital twins in teaching equipment, security control and environmental energy consumption systems, reflecting the systematicity, logical closure and high dynamic scheduling capabilities of the method of the present invention.

Claims

1. Campus IoT terminal scheduling method based on digital twins, characterized by The following steps are involved: Define a universal twin state representation structure, abstracting the operating status of different types of terminals into a unified data vector format including operating status code, power status, occupancy status, and functional parameters. Introduce a state mapping relationship and use a finite state machine to construct a one-to-many dynamic mapping relationship between physical state and virtual twin state. When the physical terminal state changes, trigger the multi-level response mapping of the twin in the virtual space. The real-time data of the twin is semantically fused to construct a situation map with dimensions of region, time, and equipment type; a scheduling sensitivity index is introduced, and disturbance simulation is performed based on past scheduling behaviors to obtain the potential impact vector corresponding to each scheduling behavior; the situation map is locally divided according to the potential impact vector, and multiple local feasible scheduling solution sets are dynamically constructed based on the current environmental state; for each local feasible scheduling solution set, a combination of Bayesian estimation and Monte Carlo method is used to jointly evaluate the feasibility probability and benefit expectation within the future target period; a scoring function is constructed based on the evaluation results, and the digital twin is called to simulate each scheduling strategy to select a scheduling path.

2. The campus Internet of Things terminal scheduling method based on digital twins according to claim 1 is characterized in that The selection of the scheduling path includes the following steps: S1. Generate twin dependency relationships using graph structure mining algorithms based on historical behavior data and real-time communication; S2. Use the subgraph convolutional neural network to infer the dependency strength, identify the coupled scheduling units, and obtain the coupled scheduling unit set; S3. Based on the set of coupling scheduling units obtained in S2, perform coordinated scheduling according to the relationship between the master terminal and the sub-terminal, so that the master terminal drives the corresponding sub-terminal; S4. During the collaborative scheduling process, the deviation between the physical terminal behavior and the expected behavior of the twin is compared in real time; if the deviation exceeds the threshold, the deviation identification model is activated to determine the cause of the drift.

3. The campus Internet of Things terminal scheduling method based on digital twins according to claim 2 is characterized in that The potential impact analysis method on the entire system includes: Establish a digital twin of each terminal in the campus Internet of Things, which is used to reflect the operating status, environmental parameters and functional behavior of the physical terminal; Build a unified twin state expression model to abstract the operating status of different types of terminals into standard state vectors; collect and integrate the historical scheduling behavior data of the twins to generate a system situation map; Introducing a scheduling sensitivity index to simulate disturbances in various scheduling behaviors and predict the potential impact of scheduling behaviors on the overall system operation status; screening scheduling priorities based on sensitivity analysis results and dynamically adjusting scheduling strategies; The scheduling plan is synchronized to the physical terminal through the twin for execution.

4. The campus Internet of Things terminal scheduling method based on digital twins according to claim 3 is characterized in that The scheduling sensitivity indicators include: resource fluctuation sensitivity, functional coupling sensitivity, behavior chain impact and system stability risk value; the disturbance simulation includes: simulating the execution of specific scheduling behaviors in the digital twin space, tracking the state changes of multiple terminal twins, and evaluating the impact trends on regional load, terminal response delay, and device coupling behavior.

5. The campus Internet of Things terminal scheduling method based on digital twins according to claim 4 is characterized in that The system situation map is constructed based on three dimensions: time, space and device type, reflecting the current system resource usage, event density and device activity.

6. The campus Internet of Things terminal scheduling method based on digital twins according to claim 5 is characterized in that The scheduling strategy selects a set of locally feasible scheduling solutions based on the disturbance simulation results, and executes the best one after a preview simulation through the digital twin. The digital twin and the physical terminal use a state deviation comparison mechanism. If the deviation exceeds a threshold, the deviation identification and compensation module is triggered.

7. The campus Internet of Things terminal scheduling method based on digital twins according to claim 2 is characterized in that The method for identifying a coupling scheduling unit includes: S1. Build digital twins of each terminal in the campus IoT system and generate a terminal interaction graph based on historical behavior data, with each digital twin as a graph node and the behavioral interactions between devices as graph edges. S2. Divide the terminal interaction graph into multiple subgraphs according to space, function or task frequency; S3. Perform dependency strength inference on the terminal nodes in each subgraph based on the subgraph convolutional neural network to generate a dependency matrix between devices; based on the dependency matrix, identify terminal combinations with high correlation and define them as coupling scheduling units; S4. The coupling scheduling unit is used as a minimum scheduling unit for unified scheduling, so that the terminals are controlled in a linked manner.

8. The campus Internet of Things terminal scheduling method based on digital twins according to claim 7 is characterized in that The terminal interaction graph is generated based on the following data: terminal operation logs, communication events, task collaboration history, spatial location relationships, and user usage behavior sequences; The subgraph division is based on the following criteria: the area where the terminal is located, the similarity of functional types, and the frequency of behavioral interactions.

9. The campus Internet of Things terminal scheduling method based on digital twins according to claim 8 is characterized in that The subgraph convolutional neural network is used to extract the behavioral collaboration features between devices in the subgraph, and infer the dependency strength between the terminals through graph embedding representation.

10. The campus Internet of Things terminal scheduling method based on digital twins according to claim 9 is characterized in that The combination of the coupling scheduling unit includes: lighting system and teaching equipment, access control system and monitoring equipment, environmental sensing equipment and ventilation system; the coupling scheduling unit executes linkage strategies for the units, including unified startup, joint energy-saving control or interlocking safety response.

Images