Property operation decision support system based on digital twinning and generative AI

Abstract

The invention discloses a property operation decision support system based on digital twinning and generative AI, which relates to the technical field of property operation management and comprises a sensing layer, a data layer, a digital twinning layer, a generative AI layer and a decision execution layer which are connected in sequence. In the invention, the generative AI layer is based on a special Transform architecture model, fuses historical operation data and digital twinning real-time data, can automatically generate an executable optimization strategy in multiple scenes such as facility maintenance, resource allocation, personalized service, emergency disposal and the like, verifies feasibility and effects in a virtual environment, and improves the implementation efficiency. The decision execution layer efficiently implements a strategy subjected to simulation verification into a physical space by using an intelligent scheduling and equipment linkage mechanism, performs quantitative evaluation and strategy iterative optimization based on an execution effect, deeply fuses a high-fidelity digital twin with a generative AI, constructs an AI decision closed-loop system based on simulation verification, and performs decision making on the AI decision closed-loop system. And a high-feasibility and feasible optimization scheme can be generated for a complex and dynamic property operation environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of property operation management, and in particular to a property operation decision support system based on digital twinning and generative AI. BACKGROUND

[0002] With the acceleration of urbanization and the increasing demand for living environment, property management is facing increasingly complex challenges. Traditional property management mode often relies on manual inspection, experience judgment and passive response, which has problems such as low efficiency, slow response, resource waste and high management cost. For example, it is difficult to give early warning of equipment failure, which leads to untimely maintenance and affects the experience of residents; energy consumption management lacks fine means, resulting in energy waste; the response of resident services is slow, and personalized needs are difficult to meet; and in the event of an emergency, there is a lack of scientific guidance and rapid linkage mechanism for emergency disposal.

[0003] Published patent: A property digital housekeeper system (CN120107026A) relates to the technical field of property management, which includes the following modules: intelligent interaction experience module, personalized service customization module, intelligent work order management module and data analysis and decision support module. This property digital housekeeper system uses AI technology and large model capability to bring significant efficiency to the customer service scene of property managers. Through intelligent interaction, the needs of property owners are responded to at any time, and service efficiency is improved; accurate property owner portraits realize personalized service push, and enhance property owner satisfaction and loyalty; intelligent work order management automatically generates and distributes work orders, and real-time tracking feedback progress ensures service quality. In addition, comprehensive data collection and analysis provide support for property decision-making, optimize operation processes and reduce costs. The digital housekeeper effectively solves the pain points of the property industry, improves service quality and value-added capacity, and lays a solid foundation for property enterprises to reduce costs and increase efficiency.

[0004] The above patent has defects: it lacks high-precision modeling and virtual simulation capability based on digital twinning, cannot fuse BIM, GIS and Internet of Things real-time data to build a panoramic property model, lacks visual and predictable dynamic simulation of equipment operation, energy consumption changes and emergency scenarios; it cannot uniformly standardize and centrally store multi-source heterogeneous data to form real-time data collaboration across systems and scenarios; it cannot rely on generative AI models to generate executable optimal strategies for facility maintenance, resource allocation, personalized services and emergency disposal, and verify the effect in a virtual environment. In the face of large-scale device connection, second-level response and dynamic change of operation environment, the existing system is difficult to realize active prediction, intelligent optimization and closed-loop execution, resulting in that property management still stays in passive response and single function support, rather than global optimization and continuous improvement. SUMMARY

[0005] The purpose of the present application is to solve the problems existing in the prior art and to provide a property operation decision support system based on digital twinning and generative AI.

[0006] In order to achieve the above-mentioned purpose, the present application adopts the following technical scheme: the property operation decision support system based on digital twinning and generative AI comprises a perception layer, a data layer, a digital twinning layer, a generative AI layer and a decision execution layer connected in turn, the perception layer is deployed in the property physical space, used for collecting equipment state, space structure information and resident behavior data, and providing original input support for digital twinning body construction, the data layer is used for converging and storing the multi-source heterogeneous data collected by the perception layer, constructing a property data lake, the digital twinning layer fuses BIM, GIS, Internet of Things sensor data and resident behavior data, constructs a multi-dimensional high-fidelity digital twinning body of building structure + equipment state + resident activity, and has a virtual scene simulation function, the generative AI layer is provided with an inference end, trains a property scene special generative AI model based on digital twinning full-scene data and property historical data, generates multi-scene decision suggestions and performs virtual simulation verification, and the decision execution layer is used for executing the decision suggestions verified by the digital twinning simulation in the property physical space, and feeding back the execution results to the generative AI layer.

[0007] As a further description of the above technical scheme: The perception layer comprises multi-type Internet of Things sensors installed on elevators, air conditioners, water pumps, illuminations and electric meters, used for collecting temperature, pressure, current, voltage, energy consumption data and fault diagnosis signals, a LoRaWAN wireless protocol network adopts a star topology structure, and is configured with channel frequency hopping and conflict avoidance technology, an edge server supports data redundant storage and UPS power supply, runs a containerized data processing service, adopts an outlier rejection and data compression algorithm, transmits data through an MQTT protocol, video data adopts a compression encoding format, voice data adopts an efficient encoding format, and in privacy protection, resident ID and location information are desensitized through a hash algorithm.

[0008] As a further description of the above technical scheme: The data cleaning process of the data layer comprises field standardization, missing value processing and outlier detection, unstructured data processing comprises OCR optical character recognition and ASR speech recognition, image, video and voice data are converted into structured information, all data are labeled with multi-dimensional tags according to equipment number, resident ID and geographic location dimension, privacy is protected by using encrypted storage and key exchange, and redundant storage adopts a distributed file system.

[0009] As a further description of the above technical scheme: The digital twin layer establishes a building structure model through BIM software, including walls, columns, stairs, elevator shafts, and spatial topological relationships, maps the real-time state of air conditioners, elevators, water meters, monitoring equipment to the digital twin body based on a unique ID mapping table, collects residents' travel time, service habits, and payment cycle data through access control, mobile phones, and smart home appliance terminals, generates a resident behavior portrait, and a simulation engine combines physical simulation, discrete event simulation, and finite element analysis to simulate equipment, energy consumption, and emergency scenarios.

[0010] As a further description of the above technical solution: The generative AI layer includes a property scene-specific generative AI model based on a Transformer architecture. The model is trained using property work orders, equipment data, and resident interaction logs, and feature engineering is performed. The inference end uses quantization acceleration, calls data lake through API, obtains energy consumption and location data, generates decision recommendations for facility maintenance, resource allocation, personalized services, and emergency response, and performs virtual testing in the digital twin body.

[0011] As a further description of the above technical solution: The decision execution layer automatically classifies tasks by type, priority, and geographic location, assigns tasks through a scheduling algorithm, issues task details through mobile terminals, wearable devices, or dispatch consoles, has emergency response device linkage logic, and returns execution state information to the digital twin body through an edge gateway, and provides acceptance feedback based on execution effect indicators.

[0012] As a further description of the above technical solution: The sensor density of the perception layer is determined by the type of equipment, and the installation location is optimized through building BIM models and thermal distribution simulation to ensure that the sampling location covers the key parts of the equipment. The edge server of the perception layer has multiple network interfaces and wireless communication modules, supporting data redundancy storage and dual-time calibration.

[0013] As a further description of the above technical solution: The digital twin layer realizes real-time synchronization between the digital twin body and the physical space through edge computing and message queue protocols, supports large-scale device connection, establishes association rules between equipment state, resident demand, and physical space, realizes demand triggering based on state changes, and the emergency scenario simulation of the digital twin layer includes virtual simulation of fire and water pipe explosion events.

[0014] As a further description of the above technical solution: The generative AI layer labels the equipment data when generating decision suggestions, and converts the resident behavior data into a demand preference vector, and when generating an emergency response strategy, the generative AI layer calls the emergency scene simulation module of the digital twin to predict the safety and loss control effect of the disposal scheme.

[0015] Further description of the above technical solution: The execution effect evaluation of the decision execution layer includes equipment operation stability, energy consumption change ratio, resident feedback score and safety risk elimination rate, and if the execution effect is not up to standard, secondary execution tasks or manual intervention are triggered.

[0016] The present application has the following beneficial effects: 1、In the present application, through the organic combination of the perception layer, the data layer, the digital twin layer, the generative AI layer and the decision execution layer, the whole-process intelligent management from multi-source data collection, real-time state perception, intelligent decision generation to execution feedback closed loop in property operation is realized, covering equipment management, energy consumption optimization, resident service and emergency response and other core links, using the multi-type Internet of Things sensors and edge computing technology deployed by the perception layer, high-precision and low-delay data collection and preprocessing of equipment operation, space structure and resident behavior of the property physical space can be realized, a unified data lake and multi-dimensional label system are constructed through the data layer to realize standardized, structured and secure storage of multi-source heterogeneous data, the digital twin layer fuses BIM, GIS, Internet of Things and resident data to generate a multi-dimensional high-fidelity virtual model of building structure, equipment state and resident activity, and has virtual simulation capability of equipment failure, energy consumption optimization and emergency scene, realizes dynamic visualization and state prediction of physical space, significantly improves the automation, predictability and refinement level of property management, changes the passive response of property management to active prediction and continuous optimization, reduces the decision risk and improves the execution efficiency, in the aspects of personalized service and emergency response, the resident satisfaction and safety are significantly improved through accurate recommendation and rapid linkage.

[0017] 2、In the application, the generative AI layer is based on a special Transformer architecture model, which integrates historical operation data and digital twin real-time data. It can automatically generate executable optimization strategies in multiple scenarios such as facility maintenance, resource allocation, personalized service, and emergency disposal, and verify the feasibility and effectiveness in a virtual environment. The decision execution layer uses intelligent scheduling and device linkage mechanism to efficiently implement the simulation-verified strategies in the physical space, and quantitatively evaluates and optimizes the strategies based on the execution results. The deep integration of high-fidelity digital twin and generative AI not only realizes panoramic modeling and dynamic simulation of multi-source heterogeneous data, but also builds an AI decision-making closed-loop system based on simulation verification. It can generate high feasibility and implementable optimization solutions for complex and dynamic property operation environment, and continuously optimize itself during the execution process, showing high innovation and practical value. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The system architecture diagram of the application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0020] Referring to Figure 1 An embodiment provided by the application: a property operation decision support system based on digital twin and generative AI, comprising a perception layer, a data layer, a digital twin layer, a generative AI layer and a decision execution layer connected in sequence. The perception layer is deployed in the physical space of the property, used to collect device state, spatial structure information and resident behavior data, and provide original input support for digital twin construction. The data layer is used to converge and store multi-source heterogeneous data collected by the perception layer, and construct a property data lake. The digital twin layer integrates BIM, GIS, Internet of Things sensor data and resident behavior data, constructs a multi-dimensional high-fidelity digital twin of building structure + device state + resident activity, and has a virtual scene simulation function. The generative AI layer is provided with an inference end, based on digital twin full-scene data and property historical data, trains a property scene special generative AI model, generates multi-scenario decision suggestions and performs virtual simulation verification. The decision execution layer is used to execute the decision suggestions verified by the digital twin simulation in the physical space of the property, and feedback the execution results to the generative AI layer.

[0021] The perception layer includes multiple types of Internet of Things sensors installed in elevators, air conditioners, water pumps, lighting, and electricity meters for collecting temperature, pressure, current, voltage, energy consumption data, and fault diagnosis signals. The LoRaWAN wireless protocol networking adopts a star topology structure and is configured with channel hopping and collision avoidance technology. The edge server supports data redundancy storage and UPS power supply, runs containerized data processing services, uses outlier rejection and data compression algorithms, transmits data through the MQTT protocol, video data uses compressed encoding format, voice data uses efficient encoding format, and in privacy protection, the resident ID and location information are desensitized through a hash algorithm. The sensor layout density of the perception layer is determined according to the type of equipment, and the installation location is optimized through building BIM models and thermal distribution simulation to ensure that the sampling location covers the key parts of the equipment. The edge server of the perception layer has multiple network interfaces and wireless communication modules, supporting data redundancy storage and dual-time calibration.

[0022] The data cleaning process of the data layer includes field standardization, missing value processing, and outlier detection. Unstructured data processing includes OCR optical character recognition and ASR speech recognition, which converts image, video, and voice data into structured information. All data is labeled with multi-dimensional tags according to equipment number, resident ID, and geographic location. Privacy is protected through encryption storage and key exchange. Redundant storage uses a distributed file system.

[0023] The digital twin layer establishes a building structure model through BIM software, including walls, columns, stairs, elevator shafts, and spatial topology relationships. Based on a unique ID mapping table, it maps the real-time status of air conditioners, elevators, water meters, and monitoring equipment to digital twins. It relies on access control, mobile phones, and smart home appliance terminals to collect resident travel time, service habits, and payment cycle data to generate resident behavior portraits. The simulation engine combines physical simulation, discrete event simulation, and finite element analysis for equipment, energy consumption, and emergency scenario simulation. The digital twin layer achieves real-time synchronization of digital twins and physical space through edge computing and message queue protocols, and supports large-scale device connections. The digital twin layer establishes association rules between equipment status, resident demand, and physical space, enabling demand triggering based on state changes. The emergency scenario simulation of the digital twin layer includes virtual simulation of fire and water pipe rupture events.

[0024] The generative AI layer includes a property scene-specific generative AI model based on a Transformer architecture. The model is trained using property work orders, equipment data, and resident interaction logs, and feature engineering is performed. The inference end uses quantization acceleration, calls data lake through API, and obtains energy consumption and location data. Decision suggestions are generated for four scenarios: facility maintenance, resource allocation, personalized services, and emergency response. The virtual test is performed in the digital twin. When generating decision suggestions, the generative AI layer labels the equipment data and converts the resident behavior data into demand preference vectors. When generating emergency response strategies, the generative AI layer calls the emergency scenario simulation module of the digital twin to predict the safety and loss control effectiveness of the disposal scheme.

[0025] The decision execution layer automatically classifies tasks by type, priority, and geographic location, and allocates tasks through a scheduling algorithm. Task details are issued through mobile terminals, wearable devices, or dispatch consoles. There is an emergency disposal equipment linkage logic. The execution state information is returned to the digital twin through the edge gateway, and the acceptance feedback is performed according to the execution effect index. The execution effect evaluation of the decision execution layer includes equipment operation stability, energy consumption change ratio, resident feedback score, and safety risk elimination rate. If the execution effect is not up to standard, a secondary execution task or manual intervention is triggered.

[0026] Embodiment one: This embodiment is aimed at a high-rise comprehensive office building, which uses a property operation decision support system based on digital twinning and generative AI to realize equipment operation optimization and energy consumption management. Multi-type Internet of Things sensors are installed on key equipment such as elevators, air conditioning units, water pumps, lighting circuits, and electricity meters in the perception layer to collect temperature, pressure, current, voltage, energy consumption data, and fault diagnosis signals. The sensor layout position is optimized through building BIM model and heat distribution simulation to ensure that the data sampling covers the key parts of the equipment. All sensors are connected through LoRaWAN wireless protocol networking, using a star topology to access the centralized gateway. The edge server runs Docker containerized data processing services, which can real-time remove outliers and use DeltaEncoding+LZ4 compression algorithm to reduce data packet size. The data layer standardizes the fields, handles missing values, and detects outliers of the real-time multi-source heterogeneous data from the perception layer, and stores them in the property data lake. The digital twinning layer uses Revit software to establish a building structure model, and maps the equipment state data such as air conditioning, elevators, and lighting circuits to the digital twin, and generates behavior portraits by combining the time and service habits of residents. The generative AI layer analyzes energy consumption and equipment status based on the PropertyGPT model, generates decision suggestions for air conditioning temperature control optimization, lighting brightness adjustment, and equipment operation schedule optimization, and performs virtual simulation verification in the digital twin. The decision execution layer is dispatched to energy management personnel according to priority, and the optimization strategy is executed through the building control system. The execution effect is evaluated by energy consumption optimization ratio and resident travel impact index, and secondary adjustment is triggered when it does not meet the standard.

[0027] The perception layer is deployed in the property physical space to collect equipment state, space structure information and resident behavior data, providing original input support for digital twin construction. In terms of equipment perception, multiple types of Internet of Things sensors are installed on key equipment such as elevators, air conditioners, water pumps, lighting and electricity meters to collect temperature, pressure, current, voltage, energy consumption data and fault diagnosis signals. The temperature sensor selects an industrial PT100 type with an accuracy of ±0.2°C, and the sampling frequency is 1Hz; the pressure sensor selects a MEMS type with an accuracy of ±0.1%FS, and the sampling frequency is 0.5Hz; the current / voltage sensor adopts a Hall effect type with an accuracy of ±0.5% and a sampling frequency of 2Hz; the energy consumption acquisition module meets the IEC62053-21 standard, with a minimum resolution of 0.1kWh; the sensor layout density is determined according to the equipment type: one set of temperature, current and voltage sensors are installed on each elevator; one set of temperature, pressure and energy consumption sensors are installed on each air conditioner; one energy consumption monitoring node is set for each lighting circuit; one set of temperature and pressure sensors are installed on each water pump unit, and the installation position is optimized through building BIM model and heat distribution simulation to ensure that the sampling position covers the key parts of the equipment. These sensors form a network through LoRaWAN wireless protocol and send data to the edge server to realize real-time, low-power and wide-coverage data acquisition. The LoRaWAN wireless protocol network adopts a star topology, with all nodes connected to a centralized gateway through LoRaWAN. A single gateway can support up to 2000 nodes, and each 2 buildings in the park are equipped with one gateway to ensure that the maximum distance from the node to the gateway does not exceed 500m. Anti-interference measures include LoRaWAN Class C mode and channel hopping to reduce packet loss rate, and LBT technology is used at the gateway port to avoid channel conflict. The edge server uses an ARM architecture 8-core processor, 16GB of memory and 1TB of SSD storage, has dual gigabit Ethernet ports and dual wireless communication modules, supports RAID1 data redundancy storage, and is equipped with a UPS power supply to ensure 30 minutes of power supply. The operating system is Ubuntu Server 22.04LTS, which runs Docker containerized data processing services. The real-time data processing algorithm uses the three-sigma method to eliminate instantaneous values that deviate from the mean by more than 3 times the standard deviation. The formula is: When the sampling value deviates from the mean by more than , the value is determined as an abnormal value and is eliminated, is the th sampling data, is the data mean, is the data standard deviation, and the key parameters use the DeltaEncoding+LZ4 compression algorithm to reduce the data packet size to 30% of the original size. The DeltaEncoding compression algorithm formula is: only store the difference of adjacent sample values Instead of the original value, cooperate with the compression algorithm to reduce the data packet size, the first The original data of the second sampling, The original data of the second sampling, The original data of the second sampling, The difference of adjacent sample values, using MQTTv5.0 protocol, QoS=2 to ensure that the data is delivered once, the delay is controlled between 30\~45ms, through NTP network time protocol and local GPS time module double time correction, the time stamp accuracy is not less than ±1ms, in the aspect of space perception, using laser mapping instrument and Revit software to build building information model, combined with GIS data to build the overall space map of the park, forming an integrated indoor and outdoor space model, providing accurate spatial basic data for subsequent three-dimensional modeling of digital twin, in the aspect of resident behavior perception, through the access control system, camera and smart home terminal to collect the activity track and behavior preference information of residents, including entry and exit time and frequency, common service type, common elevator number and floor stop number, all collected perception data need to be processed through edge computing gateway, including local filtering, format unification operation, to reduce the load of cloud transmission and processing, the perception layer should ensure low delay and high reliability of data transmission, provide timely and accurate multi-source data input for digital twin, support subsequent data analysis, simulation and intelligent decision-making.

[0028] The data layer gathers and stores the multi-source heterogeneous data collected by the perception layer, and constructs the property data lake. The data aggregation module is responsible for receiving real-time data from the perception layer and performing preliminary cleaning, formatting and standardization processing. The processed data is stored in the property data lake. The property data lake is a large-scale, multi-format data storage and management platform that can accommodate structured data, semi-structured data and unstructured data, providing rich and high-quality data foundation for subsequent digital twin modeling and AI model training. The data layer unifies the device status data, resident terminal behavior data and video image data uploaded from the Internet of Things sensors into the data lake architecture, and performs field standardization processing on different sources of data to ensure consistency of data format and naming, realizing centralized management of multi-source heterogeneous data. The data cleaning process includes: ① Field standardization, using ApacheNiFi to convert field names and units according to a unified field dictionary; ② Missing value processing, time series data is filled by linear interpolation, the linear interpolation formula is: , , , , , , , , The known value of is the time point corresponding to the missing value, It is the time point before and after the missing value, and the categorical data is completed by the nearest neighbor; ③ Outlier detection, using the IQR and Z-Score double detection method to eliminate outliers, IQR formula: , ,like or , then it is an outlier, is the first quartile, is the third quartile, is the interquartile range, is the detected data point, Z-Score formula: ,like ,but is an outlier, It is data, is the mean, is the standard deviation, It is a standardized score. For structured data including equipment operation logs, energy consumption records, and work order records, time series modeling is required to extract the key indicators of equipment operation time and failure frequency. For unstructured data such as failure photos, surveillance videos, and resident voice, optical character recognition (OCR) and speech recognition (ASR) are required to extract key semantic fields and convert them into computable structured information. Video data is encoded using H.265, with a resolution compressed to 720p and a key frame interval of 2 seconds. Voice data is encoded using Opus with a sampling rate of 16kHz. All data is multi-dimensionally tagged according to device number, resident ID, and geographic location, supporting precise association and retrieval. Privacy protection uses AES-256 encrypted storage + RSA-2048 key exchange. Resident ID and location information are desensitized using the SHA-256 + random salt hash algorithm. Kafka message queues are used for real-time processing of high-concurrency data streams. Sensor and resident data are written with low latency. Redundant storage uses the Ceph distributed file system with 3 replicas. A single node failure automatically triggers data reconstruction, and data durability reaches 99.999%.

[0029] The digital twin layer fuses BIM, GIS, Internet of Things sensor data, and resident behavior data to construct a multi-dimensional and high-fidelity digital twin of the building structure, equipment state, and resident activities. Multi-source data is modeled. Building structure modeling establishes the structural information and spatial topological relationship of walls, columns, stairs, and elevator shafts through Revit software. Equipment state modeling maps the real-time state of air conditioners, elevators, water meters, and monitoring equipment, including temperature, pressure, energy consumption, and running time, and maps it to the digital twin based on a unique ID mapping table. The mapping delay is less than 30 ms. Resident behavior modeling relies on mobile phones, access control, and smart home appliance terminals to collect travel time, service usage habits, and payment cycle data to construct a resident behavior profile. The digital twin layer needs to undertake virtual scene simulation functions to verify the feasibility and impact assessment of strategies before decision-making, including equipment failure, energy optimization, and emergency scenario simulation. The simulation results will serve as an important basis for the implementation of generative AI layer strategies, verifying the feasibility of generative AI decisions. Equipment failure diffusion simulation simulates the impact of elevator failure on other equipment in the building. Resident impact simulation evaluates the actual impact of decisions on residents. Emergency scenario simulation pre-rehearses the disposal effect of fire and water pipe explosion events. The digital twin interface displays equipment failure anomalies or resident entry and exit states in real time, establishes association rules between equipment state, resident demand, and physical space, triggers demand based on state changes, and achieves intelligent linkage effects. When equipment is abnormal, it automatically associates with residents who frequently use the equipment and triggers travel reminders. When residents enter a specific area, it automatically pushes related service guidelines. The simulation engine uses Unity3D combined with the open-source physics simulation library BulletPhysics to simulate equipment movement and physical characteristics. Emergency scenarios use discrete event simulation combined with finite element analysis to improve structural safety simulation accuracy. Real-time synchronization of the digital twin and the physical space is achieved through edge computing and the MQTT protocol, ensuring that data transmission delay does not exceed 50 milliseconds. Edge servers need to be deployed in the property community to locally process sensor data, filter outliers in Internet of Things sensor data, aggregate key parameter changes to reduce cloud pressure, and push sensor data to the digital twin in real time through the lightweight message queue protocol, supporting million-level device connections. Equipment state mapping uses a data dictionary mapping table to ensure that each object in the digital twin has a unique ID, with an average mapping delay of less than 30 ms.

[0030] The generative AI layer is based on a 12-layer Transformer architecture with 1024 hidden layers and approximately 340 million parameters, which integrates digital twin full-scenario data and property historical data. The full-scenario data includes device status, resident behavior, and energy consumption. The property operation historical data includes work order records, maintenance logs, and complaint data. The PropertyGPT model is trained using these data. The PropertyGPT model can understand complex property operation contexts and is trained using nearly 5 years of 12 million work orders, 50 TB of device data, and 30 million resident interaction logs.The feature engineering adopts Min-Max normalization, One-Hot encoding, and WordPiece segmentation. The training is performed on 8 NVIDIA A100 GPUs with a batch size of 256, a learning rate of 1e-4, an optimizer of AdamW, and an inference end with INT8 quantization. The delay control is within 200 ms. The digital twin data is structured, such as converting device temperature into abnormal / normal labels and converting resident behavior into demand preference vectors. The historical data is cleaned, including removing duplicate work orders and filling in missing fields. The pre-trained model is fine-tuned using property scene data to improve its semantic understanding of the property scene, identify fault types, and understand resident needs. External tool calling functions are integrated to enable the model to access energy consumption data from the data lake or device locations in the digital twin in real time during decision generation, thereby improving the executability of the strategy. External tool calling interacts with the data lake API through the gRPC protocol with a response delay of ≤100 ms, enabling real-time access to energy consumption and location data. PropertyGPT automatically generates decision recommendations for four core scenarios in property operation, including facility maintenance, resource allocation, personalized services, and emergency response. These recommendations can be in the form of text or executable instruction sequences. The device state data in the digital twin is analyzed to predict device failure probability and generate maintenance strategies. The energy consumption data and resident behavior data in the digital twin are analyzed to generate energy consumption optimization strategies. The resident behavior data in the digital twin is analyzed to generate personalized service pushes. The emergency scenario data in the digital twin is analyzed to generate emergency response strategies. In the digital twin simulation verification stage, the decision recommendations of the generative AI need to be virtually tested in the digital twin to verify their feasibility, rationality, and effectiveness, including energy consumption change prediction, resident impact assessment, and emergency effectiveness prediction. Device effectiveness simulation is used to predict energy consumption changes after strategy implementation. Resident impact simulation is used to assess the impact of device maintenance or downtime on travel time and convenience. Emergency scenario simulation is used to predict the effectiveness of disposal schemes in fire and water pipe burst scenarios in terms of safety and loss control. The simulation results need to meet acceptable performance and impact indicators to enter the execution phase. After decision execution, feedback and iteration are performed on the execution effect. Real-time actual operation data after strategy execution is collected through the digital twin and fed back to the PropertyGPT model for strategy adjustment and model optimization. If the strategy does not achieve the expected effect, the model should automatically adjust the scheme, such as adding inspection steps or replacing execution personnel. These execution feedbacks are incorporated into subsequent model training to continuously improve the accuracy and adaptability of decisions.

[0031] The decision execution layer efficiently, accurately and safely implements the decision suggestions verified by the generative AI layer through digital twinning in the property physical space, monitors, feeds back and closes the loop of the execution process, and ensures that the strategy achieves the expected effect in the real scene. In the task receiving and dispatching link, the decision execution layer needs to connect the output interface of the generative AI layer, receive the decision tasks that have been verified by digital twinning simulation and meet the performance and impact indicators. These tasks include equipment maintenance job instructions, energy optimization scheduling schemes, personalized service execution plans for residents, and emergency disposal action plans. The system needs to automatically classify and dispatch tasks according to type, priority and geographical location to the corresponding execution subjects, such as equipment maintenance personnel, energy management personnel, customer service personnel and safety management personnel. The task scheduling algorithm adopts an improved genetic algorithm combined with the shortest path priority strategy. The improved genetic algorithm formula is: wherein is the distribution variable, which is iteratively optimized by the genetic algorithm, is the total scheduling cost (time or distance), is the cost (time / distance) of the execution personnel to execute the task , is a 0-1 variable indicating whether the task is executed by the personnel , is the total number of execution personnel, The total number of tasks is optimized based on the task priority, the current location of the executor, and their skill matching. The scheduling calculation time does not exceed 1 second. During the operation implementation and process control, the decision-making execution layer needs to use mobile terminals, wearable devices, or scheduling consoles to push task details, including location, execution steps, and a list of required tools and materials, to the executor or control system in real time. For equipment maintenance tasks, the decision-making execution layer needs to complete component replacement, parameter adjustment, and functional testing operations according to strategic instructions. For energy consumption optimization tasks, the decision-making execution layer needs to implement air conditioning temperature control adjustments, lighting brightness optimization, and equipment operation schedule adjustments through the building control system. For personalized service tasks, the decision-making execution layer needs to push customized notifications, service appointments, and life convenience plans to target residents. For emergency response tasks, the emergency response equipment linkage logic, including fire protection systems, emergency lighting, broadcasting systems, organize personnel evacuation, and deploy emergency resources, is implemented. For example, if a fire detection signal is triggered, the air conditioning fan is first turned off, the fire damper is activated, and the fire water pump and emergency broadcast are turned on. If a water pipe burst is triggered, the main valve is first closed, and then the drainage pump is started and the maintenance team is notified. During the execution process monitoring and status feedback phase, the decision-making execution layer must collect execution status information in real time, including task progress, key node completion status, equipment operation change data, and personnel location and status information. For equipment operations, operating parameters before, during, and after the operation must be collected. For emergency response tasks, evacuation speed, resource arrival time, and risk elimination status must be collected. This status data is transmitted back to the digital twin via the edge gateway, synchronized with the virtual model, and used to dynamically monitor execution results. Execution effect indicators include: energy consumption optimization ratio = (energy consumption before optimization - energy consumption after optimization) ÷ energy consumption before optimization × 100%; household travel impact index = Σ (affected households × average delay) ÷ total number of households. During the result acceptance and feedback closed-loop phase, the decision-making execution layer must formulate acceptance criteria based on the task type, such as the operating stabilization time after equipment maintenance, the energy consumption change ratio after energy consumption optimization, the household feedback score after personalized services, and the safety risk elimination rate after emergency response. The acceptance results must be recorded in the task database and generated into an execution report, which is fed back to the generative AI layer as training data for subsequent strategy optimization. The energy consumption optimization ratio formula is: , is the energy consumption before optimization, is the optimized energy consumption, It is the energy consumption optimization ratio. If it is found that the execution effect does not meet the standard, a secondary execution task must be triggered immediately, or the abnormal information must be reported to the management for manual intervention. The work of the decision-making execution layer forms a closed-loop execution system from task reception, dispatching, implementation, monitoring to feedback, which not only ensures the implementation of generative AI decisions, but also provides high-quality execution effect data support for digital twins and AI models, realizing an efficient closed loop between decision-making and execution.

[0032] Embodiment two: This embodiment is applied to large residential communities, focusing on the realization of emergency event rapid response and risk control. Temperature, pressure, current, voltage, energy consumption, etc. Sensors are deployed in the park for elevators, water pumps, air conditioners, lighting and other equipment. Combined with the access control system, cameras and smart home terminals, the travel trajectory and behavior preference information of the residents are collected. All sensing data is transmitted to the park edge server through LoRaWAN networking. The edge server has RAID1 redundant storage and UPS power supply protection, and performs three sigma method to remove transient outliers. The data layer manages the collected data with multi-dimensional labeling. AES-256 encryption storage and RSA-2048 key exchange are used to protect privacy. The digital twin layer establishes a building and equipment state model in the park and loads an emergency scenario simulation module to support virtual pre-rehearsal of fire, water pipe rupture, etc. After receiving the fire detection signal, the generative AI layer calls the digital twin emergency simulation module, generates an emergency disposal strategy based on historical event data and real-time state, including closing the air conditioner fan, starting the fire valve, opening the fire pump and emergency broadcast, etc. The decision execution layer receives and issues emergency disposal tasks to relevant personnel and systems, monitors the task execution progress and evacuation speed in real time, and updates the execution state to the digital twin for synchronization. The execution effect is evaluated by the safety risk elimination rate and resource arrival time. If it is lower than the threshold, a backup emergency plan is started to ensure effective control of the event.

[0033] Finally, it should be noted that: the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent replacements for some technical features. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A property operation decision support system based on digital twin and generative AI, characterized in that: The system comprises a perception layer, a data layer, a digital twin layer, a generative AI layer and a decision execution layer connected in sequence. The perception layer is deployed in the physical space of the property, used to collect device status, spatial structure information and resident behavior data, and provide original input support for digital twin construction. The data layer is used to converge and store the multi-source heterogeneous data collected by the perception layer, and construct a property data lake. The digital twin layer fuses BIM, GIS, Internet of Things sensor data and resident behavior data, constructs a multi-dimensional high-fidelity digital twin of building structure + device status + resident activity, and has a virtual scene simulation function. The generative AI layer is provided with an inference end, based on the digital twin full-scene data and property historical data, trains a property scene special generative AI model, generates multi-scene decision suggestions and performs virtual simulation verification, and the decision execution layer is used to execute the decision suggestions verified by the digital twin simulation in the physical space of the property, and feed back the execution results to the generative AI layer.

2. The digital twin and generative AI based property operations decision support system according to claim 1, characterized in that: The perception layer comprises multiple types of Internet of Things sensors installed on elevators, air conditioners, water pumps, lighting and electric meters, used to collect temperature, pressure, current, voltage, energy consumption data and fault diagnosis signals. The LoRaWAN wireless protocol networking adopts a star topology structure, and is configured with channel frequency hopping and collision avoidance technology. The edge server supports data redundancy storage and UPS power supply, runs containerized data processing services, adopts outlier rejection and data compression algorithms, transmits data through the MQTT protocol, video data adopts compression encoding format, and voice data adopts efficient encoding format. In privacy protection, the resident ID and location information are desensitized through a hash algorithm.

3. The digital twin and generative AI based property operations decision support system according to claim 1, characterized in that: The data cleaning process of the data layer includes field standardization, missing value processing and outlier detection. Unstructured data processing includes OCR optical character recognition and ASR speech recognition, which converts image, video and voice data into structured information. All data are labeled with multi-dimensional tags according to device number, resident ID and geographic location, and privacy is protected by encrypted storage and key exchange. Redundant storage uses a distributed file system.

4. The digital twin and generative AI based property operations decision support system according to claim 1, characterized in that: The digital twin layer establishes a building structure model through BIM software, including walls, columns, stairs, elevator shafts and spatial topological relationships. Based on a unique ID mapping table, the real-time status of air conditioners, elevators, water meters, monitoring devices is mapped to the digital twin. The resident's travel time, service habit, payment cycle data are collected through access control, mobile phone and smart home appliance terminals to generate a resident behavior portrait. The simulation engine combines physical simulation, discrete event simulation and finite element analysis for device, energy consumption and emergency scene simulation.

5. The digital twin and generative AI based property operations decision support system according to claim 1, characterized in that: The generative AI layer includes a property scene special generative AI model based on the Transformer architecture. The model is trained using property work orders, device data and resident interaction logs, and feature engineering is performed. The inference end uses quantization acceleration. Through API calls to the data lake, energy consumption and location data are obtained. Decision suggestions are generated for facility maintenance, resource allocation, personalized service and emergency response, and are virtually tested in the digital twin.

6. The digital twin and generative AI based property operations decision support system according to claim 1, characterized in that: The decision execution layer is automatically classified by task type, priority, and geographic location, and assigns tasks through a scheduling algorithm. Task details are issued through mobile terminals, wearable devices, or dispatch consoles. There is a linkage logic for emergency response equipment. Execution status information is returned to the digital twin through an edge gateway, and acceptance feedback is provided based on execution effect indicators.

7. The digital twin and generative AI based property operations decision support system according to claim 2, characterized in that: The sensor layout density of the perception layer is determined by device type, and the installation location is optimized through building BIM models and heat distribution simulation to ensure that the sampling location covers the key parts of the equipment. The edge server of the perception layer has multiple network interfaces and wireless communication modules, supporting data redundancy storage and dual time calibration.

8. The digital twin and generative AI based property operations decision support system according to claim 4, characterized in that: The digital twin layer achieves real-time synchronization between the digital twin and the physical space through edge computing and message queue protocols, and supports large-scale device connectivity. The digital twin layer establishes association rules between device status, household demand, and physical space, enabling demand triggering based on state changes. The emergency scenario simulation of the digital twin layer includes virtual simulation of fire and water pipe explosion events.

9. The digital twin and generative AI based property operations decision support system according to claim 5, characterized in that: The generative AI layer labels device data when generating decision recommendations and converts household behavior data into demand preference vectors. When generating emergency response strategies, the generative AI layer calls the emergency scenario simulation module of the digital twin to predict the safety and loss control effectiveness of the disposal scheme.

10. The digital twin and generative AI based property operations decision support system according to claim 6, characterized in that: The execution effect evaluation of the decision execution layer includes device operation stability, energy consumption change ratio, household feedback score, and safety risk elimination rate. If the execution effect is not up to standard, secondary execution tasks or manual intervention will be triggered.

Images