Intelligent building remote operation and maintenance management and control system based on large model and cloud edge collaborative architecture

Abstract

The invention relates to the technical field of operation and maintenance management and control, and discloses an intelligent building remote operation and maintenance management and control system based on a large model and a cloud edge collaborative architecture, and the system comprises a configuration subsystem which is used for carrying out the reverse proxy and persistent connection configuration of edge gateway equipment in an intelligent building, and obtaining a TLS encryption bidirectional communication channel; the construction subsystem is used for constructing a cloud-edge-end three-layer cooperative computing architecture; the conversion subsystem is used for performing unified semantic abstraction and conversion on the heterogeneous protocol data to obtain semantic data in a standard JSON (JavaScript Object Notation) format; the rule processing subsystem is used for performing rule processing through a near-end decision engine of the edge gateway to obtain a local intelligent response result; the multi-dimensional analysis subsystem is used for carrying out multi-dimensional analysis through a cloud large model engine to obtain an equipment health score, a fault prediction result and an energy optimization strategy, seamless fusion and interoperation of heterogeneous system data are achieved through the system, and then global optimization and linkage control of the intelligent building are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of operation and maintenance management and control, and particularly to a smart building remote operation and maintenance management and control system based on a large model and a cloud-edge collaborative architecture. Background Art

[0002] Traditional building management systems are difficult to achieve cross-system data interconnection and collaborative control. Operation and maintenance personnel need to frequently switch between different management platforms, and single-point failures may trigger chain reactions, ultimately affecting the overall building operation efficiency and user experience.

[0003] Current building remote operation and maintenance management and control faces multiple challenges: on the one hand, traditional VPN or port mapping has poor communication stability in a multiple NAT environment. The instruction loss rate of the BACnet / IP protocol is as high as 22%, and the communication failure rate of the Modbus-TCP protocol exceeds 30% under strict firewall policy restrictions; on the other hand, operation and maintenance personnel lack a unified view and it is difficult to accurately identify abnormal device states and root causes of faults. The system relies on the cloud for decision-making, resulting in slow local critical business responses in case of network interruptions or cloud platform failures and other abnormal situations, unable to meet the requirements of emergency response time. In addition, traditional analysis methods are difficult to deeply mine the complex associations and trends hidden in building operation data and cannot provide accurate fault prediction and intelligent operation and maintenance decision-making support. Summary of the Invention

[0004] The present invention provides a smart building remote operation and maintenance management and control system based on a large model and a cloud-edge collaborative architecture. The system realizes seamless integration and interoperability of heterogeneous system data, and further realizes global optimization and linkage control of smart buildings.

[0005] In a first aspect, the present invention provides a smart building remote operation and maintenance management and control system based on a large model and a cloud-edge collaborative architecture. The smart building remote operation and maintenance management and control system based on a large model and a cloud-edge collaborative architecture includes: A configuration subsystem for performing reverse proxy and persistent connection configuration on edge gateway devices in a smart building to obtain a TLS encrypted two-way communication channel; A construction subsystem for constructing a cloud-edge-end three-layer collaborative computing architecture based on the TLS encrypted two-way communication channel. The cloud-edge-end three-layer collaborative computing architecture includes a cloud large model engine, an edge gateway, and a terminal device layer; A conversion subsystem for performing unified semantic abstraction and conversion on heterogeneous protocol data through the edge gateway to obtain standard JSON format semantic data; A rule processing subsystem for inputting the standard JSON format semantic data into the proximal decision engine of the edge gateway for rule processing to obtain a local intelligent response result; The multi-dimensional analysis subsystem is used to transmit the standard JSON format semantic data and the local intelligent response results to the cloud large model engine through the TLS encrypted bidirectional communication channel for multi-dimensional analysis, so as to obtain the device health score, the fault prediction result and the energy optimization strategy.

[0006] Combined with the first aspect, in the first implementation manner of the first aspect of the present invention, the configuration subsystem is specifically used for: Deploy a network proxy server cluster in the cloud to obtain an NPS server cluster, and install a network proxy client NPC in the edge gateway device in the intelligent building; Authenticate the TCP / TLS long connection request initiated by the edge gateway device through the NPS server cluster to obtain an active connection registration table; Establish a reverse proxy channel for the edge gateway device based on the active connection registration table to obtain a NAT penetration communication link, and perform TLS 1.2 encryption processing on the NAT penetration communication link to obtain a TLS encrypted bidirectional communication channel.

[0007] Combined with the first aspect, in the second implementation manner of the first aspect of the present invention, the construction subsystem is specifically used for: Deploy and configure the cloud infrastructure to obtain a cloud large model engine; Deploy a building-specific protocol adaptation engine, a protocol abstraction and semantic conversion module, a data preprocessing and caching module, an edge security gateway and a proximal decision-making engine in the edge gateway; Connect HVAC equipment, intelligent lighting equipment, elevator control equipment, video surveillance equipment, intrusion alarm equipment, access control management equipment and energy metering equipment to the edge gateway to obtain a terminal device layer; Configure the data closed-loop process of the cloud large model engine, the edge gateway and the terminal device layer through the TLS encrypted bidirectional communication channel to construct a cloud-edge-terminal three-layer collaborative computing architecture.

[0008] Combined with the first aspect, in the third implementation manner of the first aspect of the present invention, the conversion subsystem further includes: A docking unit for docking the protocol abstraction and semantic conversion module in the edge gateway with the protocol adaptation engine to obtain a receiving channel for the original protocol data; A layering unit for performing a hierarchical design on the unified semantic model based on the concept of point cloud digital twin, and constructing a five-layer structure model including a space model layer, a device model layer, a measurement point model layer, a service model layer and an association model layer; An extraction unit, configured to identify the data format and extract valid values from the heterogeneous protocol data of BACnet / IP and Modbus-TCP parsed by the protocol adaptation engine through the receiving channel, so as to obtain the original device measurement values and control parameters; A mapping unit, configured to perform standardized mapping on the original device measurement values and control parameters based on the five-layer structure model to obtain semantically unified device data; An encapsulation unit, configured to encapsulate the semantically unified device data in JSON format to obtain standard JSON format semantic data.

[0009] Combined with the first aspect, in the fourth implementation manner of the first aspect of the present invention, the mapping unit is specifically configured to: Define the building physical space structure for the space model layer, and construct a set of standardized space identifiers including buildings, floors, areas, rooms and their hierarchical relationships; Define the description of device types, attributes, functions and states for the device model layer, and construct a set of standardized device identifiers including HVAC devices, lighting devices, elevator systems, security devices, and energy metering devices; Define standard identifiers, data types, units, value ranges, and alarm thresholds for the measurement point model layer to obtain a standardized measurement point description library; Based on the standardized measurement point description library, construct a mapping table between the object identifiers and attribute identifiers in the BACnet / IP protocol and the register addresses and standard measurement point identifiers in the Modbus-TCP protocol to obtain a conversion dictionary; According to the conversion dictionary, perform protocol-specific identifier replacement and data type conversion on the original device measurement values to obtain device parameters with unified identifiers and data types; Perform unit standardization processing and validity verification on the device parameters with unified identifiers and data types, and at the same time associate them with the set of standardized space identifiers and the set of standardized device identifiers to obtain semantically unified device data.

[0010] Combined with the first aspect, in the fifth implementation manner of the first aspect of the present invention, the rule processing subsystem is specifically configured to: Deploy a local status maintenance module, a preset rule execution module, a lightweight inference module and a breakpoint resumption module in the proximal decision engine of the edge gateway; Input the standard JSON format semantic data into the local status maintenance module for status update processing to obtain a real-time device status database; Perform boolean expression condition matching on the parameter values in the real-time device status database to obtain a sequence of operation instructions to be executed; Input the abnormal parameters in the real-time device status database into the lightweight inference module for knowledge distillation neural network calculation, perform local anomaly detection and optimization adjustment based on the parameters sent from the cloud, and obtain intelligent optimization control instructions; Perform priority sorting on the operation instruction sequence and the intelligent optimization control instruction according to the event type to obtain an execution instruction set with priority sorting; Track the execution status of the execution instruction set with priority sorting, and locally cache important data during the execution process through the breakpoint resumption module to ensure the operation of the local decision-making mode during network interruption. After the network resumes, synchronize with the cloud in the order of priority to obtain the local intelligent response result.

[0011] Combined with the first aspect, in the sixth implementation manner of the first aspect of the present invention, the multi-dimensional analysis subsystem further includes: A deployment unit for deploying a multi-functional large model engine based on the LLaMA architecture with deep fine-tuning, including a data processing module, an intelligent analysis module, a decision-making generation module, a natural language interaction module, and a model optimization module, in the cloud large model engine; A data processing unit for inputting the standard JSON format semantic data and the local intelligent response result into the data processing module through the TLS encrypted bidirectional communication channel, and performing data cleaning, standardization, time series alignment, and feature extraction processing to obtain target time series data; An intelligent analysis unit for inputting the target time series data into the intelligent analysis module, applying deep learning algorithms to perform energy consumption pattern recognition, device performance evaluation, abnormal behavior detection, and fault feature extraction, and obtaining a device health score; A decision-making generation unit for inputting the device health score, historical maintenance records, and professional knowledge base data into the decision-making generation module, performing fault diagnosis, root cause analysis, and predictive maintenance decision calculation, and obtaining a fault prediction result; A multi-objective optimization unit for performing multi-objective optimization algorithm calculation through the decision-making generation module based on the fault prediction result to obtain an energy optimization strategy.

[0012] Combined with the first aspect, in the seventh implementation manner of the first aspect of the present invention, the deployment unit is specifically used for: Adjust the architecture and perform initialization configuration on the original LLaMA pre-trained model in the cloud large model engine to construct an initialized LLaMA base model suitable for the building field; Perform data cleaning, structured processing, and annotation on building equipment manuals, historical operation and maintenance records, industry standard documents, and maintenance cases to construct a building operation and maintenance training corpus containing equipment parameters, fault characteristics, maintenance strategies, and control logic; Supervise and fine-tune and reinforce the learning training of the initialized LLaMA base model based on the building operation and maintenance training corpus, and selectively update the model weights by applying parameter-efficient fine-tuning technology to obtain a pre-fine-tuned model adapted to the building field; Integrate the building professional knowledge graph data into the pre-fine-tuned model adapted to the building field through knowledge enhancement technology, execute knowledge distillation and model compression algorithms, and at the same time perform multi-task training on the pre-fine-tuned model adapted to the building field to obtain a large model for the vertical field; Build a multi-functional large model engine based on the LLaMA architecture with deep fine-tuning, which includes a data processing module, an intelligent analysis module, a decision-making generation module, a natural language interaction module, and a model optimization module, based on the large model for the vertical field.

[0013] Combined with the first aspect, in the eighth implementation manner of the first aspect of the present invention, the intelligent building remote operation and maintenance management and control system based on the large model and the cloud-edge collaborative architecture further includes: An execution subsystem for setting warning thresholds and alarm thresholds according to the device health score to obtain device health management trigger conditions; calculating the remaining service life of the device based on the fault prediction result and historical fault data, and generating a predictive maintenance work order; generating a rolling optimization control strategy according to the energy optimization strategy combined with weather forecast data, energy prices, predicted people flow, and historical usage patterns, and converting the rolling optimization control strategy into a dynamic control instruction; analyzing the device access mode and network traffic characteristics in the building to obtain a security protection strategy; classifying the predictive maintenance work order, dynamic control instruction, and security protection strategy into three working modes: automatic execution mode, recommended approval mode, and manual intervention mode according to the priority to obtain a classified execution strategy set; sending the dynamic control instruction in the classified execution strategy set to the corresponding edge gateway for execution through the TLS encrypted two-way communication channel, and converting the control instruction into the corresponding device protocol format through the edge gateway to obtain the execution result of the terminal device.

[0014] In the technical solution provided by the present invention, a TLS-encrypted two-way communication channel configured through reverse proxy and persistent connection solves the problems of high loss rate of BACnet / IP protocol instructions in a multi-NAT environment for traditional VPNs or port mapping, and high communication failure rate of the Modbus-TCP protocol under strict firewall policy restrictions. It ensures the confidentiality and integrity of data transmission and effectively resists various network threats including protocol-layer attacks. The five-layer structure unified semantic model constructed based on the concept of point cloud digital twin solves the technical problem that it is difficult for devices from different manufacturers in traditional smart buildings to interoperate due to different communication protocols and private interfaces, and realizes seamless fusion and interoperability of heterogeneous system data. The proximal decision-making engine of the edge gateway implements a four-level response mechanism, solving the problem that in abnormal situations such as network interruption or cloud platform failure in the existing system, relying on the cloud for decision-making leads to slow response of local critical services, and ensuring the continuous and stable operation of critical services such as emergency dispatching of elevators. The cloud-edge-end three-layer collaborative computing architecture follows a closed-loop process of "terminal collection-edge processing-cloud analysis-edge execution-terminal response", breaking the information barrier between heterogeneous systems and solving the problem that traditional building systems form "data islands" due to different protocols and data formats and cannot achieve global optimization and linkage control. The cloud-based large model engine, by integrating a building professional knowledge base, has the ability of in-depth semantic understanding of multi-source heterogeneous data, can associate with the building operation and maintenance knowledge base, and solves the technical problem that traditional analysis methods are difficult to deeply mine the complex associations and trends hidden in building operation data and cannot provide accurate fault prediction and intelligent operation and maintenance decision support. Through standardized access, automated configuration, and remote maintenance capabilities, the system deployment complexity and operation and maintenance costs are significantly reduced, and the workload of network configuration, protocol docking, and system debugging, which originally accounted for a relatively high proportion of the total project working hours, is greatly reduced, reducing the total cost of ownership of the system.

[0015] Other features and advantages of the present invention will be described in the following specification, and part of them will become obvious from the specification or be understood by implementing the present invention. The objectives and other advantages of the present invention are achieved and obtained by the structures specifically pointed out in the specification, claims, and drawings.

[0016] To make the above objectives, features, and advantages of the present invention more obvious and understandable, the following preferred embodiments are specifically given in conjunction with the accompanying drawings and described in detail as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of an embodiment of a smart building remote operation and maintenance management and control system based on a large model and a cloud-edge collaborative architecture in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0019] As used in the embodiments of the present invention, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but optionally further includes other steps or units not listed, or optionally further includes other steps or units inherent to these processes, systems, products, or devices.

[0020] For ease of understanding of this embodiment, first, a detailed introduction will be given to a smart building remote operation and maintenance management and control system based on a large model and a cloud-edge collaboration architecture disclosed in the embodiments of the present invention. As Figure 1 shown, this system includes: A configuration subsystem 101, configured to perform reverse proxy and persistent connection configuration on the edge gateway devices in the smart building to obtain a TLS-encrypted two-way communication channel; Specifically, the construction of the network infrastructure is completed in the cloud, that is, a cluster of network proxy servers with high availability and load balancing capabilities, namely the NPS server cluster, is deployed through the cloud platform. Each NPS server receives and manages connection requests initiated by each edge gateway device inside the building and has unified security authentication and routing management functions. At the same time, a dedicated network proxy client NPC is pre-installed on each edge gateway device deployed inside each smart building. As the initiator of active communication, the NPC will automatically initiate a long connection request based on the TCP / TLS protocol to the NPS server in the cloud after the gateway device is powered on or the network is restored. The identity identification and security authentication of this request initiation process are achieved through means such as the device's unique identity and certificate. On this basis, the NPS server uniformly executes the authentication process for all received long connection requests, and performs two-way verification of the identity of the edge gateway device based on the X.509 certificate system or the dynamic certificate mechanism. Only devices that pass the authentication are allowed to establish a connection. After successful authentication, the NPS server records key information such as the unique identifier of the corresponding edge gateway, the current communication channel status, the last heartbeat time, and the connection establishment time in an active connection registry in real time. This registry is dynamically maintained on the server side and reflects the communication status between all edge gateways and the cloud. Based on the information in the active connection registry, the cloud system assigns a unique reverse proxy channel to each edge gateway device and establishes a two-way data communication link. The core of the reverse proxy mechanism is that the edge gateway NPC actively maintains the external connection channel to the NPS server and uses the NPS cluster in the cloud for unified traffic transfer and routing, enabling the cloud operation and maintenance platform to seamlessly access any device inside the building. Even if there are complex network isolations, NAT conversions, or multi-level firewalls inside the building, there is no need to configure local port mapping, nor to modify the user-side network rules. After the reverse proxy channel is established, all upstream and downstream communication data is forwarded through the NAT penetration link, and each link must be encrypted by TLS 1.2 to ensure that all data streams transmitted on the public network have end-to-end data confidentiality and integrity, effectively preventing man-in-the-middle attacks and data leakage risks, and obtaining a TLS-encrypted two-way communication channel.

[0021] The construction subsystem 102 is used to construct a cloud-edge-end three-layer collaborative computing architecture based on the TLS-encrypted two-way communication channel. The cloud-edge-end three-layer collaborative computing architecture includes a cloud large model engine, edge gateways, and a terminal device layer; Specifically, the cloud infrastructure is standardized in deployment and parameterized in configuration. By deploying a high-performance computing cluster on the cloud platform and building and fine-tuning a large model engine dedicated to the intelligent building scenario on it, the cloud-based large model engine integrates a vast amount of data samples related to building operation and maintenance, equipment knowledge bases, industry specifications, and historical maintenance records, and integrates natural language processing and knowledge reasoning capabilities, enabling the cloud to have the capabilities of global data analysis, predictive maintenance decision-making, and intelligent question-and-answer support. At the same time, in the edge gateway of each intelligent building, multi-level functional components are integrated and deployed, including a building-specific protocol adaptation engine that can automatically identify and parse various mainstream and proprietary communication protocols such as BACnet / IP, Modbus, OPC UA / DA, GB28181, etc. Then, through the protocol abstraction and semantic conversion module, the collected underlying heterogeneous protocol data is real-time converted into structured and unified standard semantic data. At the same time, a data preprocessing and caching module is deployed in the edge gateway to achieve real-time cleaning, aggregation, noise reduction, and intelligent caching of local data, reducing bandwidth pressure and ensuring data integrity and continuity in the case of network fluctuations or short-term disconnection. On this basis, the edge gateway integrates the edge security gateway function to achieve access authentication, data encryption and decryption, and access control of terminal devices, ensuring that all data streams entering and leaving the edge gateway comply with the system security policy and effectively defending against potential network attacks and data tampering risks; while the proximal decision-making engine pre-installs multi-level business rules and emergency control logic, enabling the edge gateway to independently make critical business responses according to the local real-time status in the case of disconnection or abnormality, such as fire alarm linkage, elevator emergency control, etc., thereby enhancing the business resilience and security reliability of the system. With the gradual deployment completion of the above-mentioned edge function modules, it is possible to support various terminal facilities such as heating, ventilation, and air conditioning equipment, intelligent lighting equipment, elevator control equipment, video surveillance equipment, intrusion alarm equipment, access control management equipment, and energy metering equipment to be connected to the edge gateway in a wired or wireless manner, thus forming the terminal device layer. Through the TLS encrypted two-way communication channel established in the aforementioned configuration subsystem, the cloud-based large model engine, the edge gateway, and the terminal device layer are connected to build a cloud-edge-terminal three-layer collaborative computing architecture, and the data flow follows a complete closed-loop from real-time collection of terminal devices, edge intelligent preprocessing, in-depth analysis by the cloud-based large model, and then automatic response at the edge or terminal.

[0022] The conversion subsystem 103 is used to perform unified semantic abstraction and conversion on heterogeneous protocol data through the edge gateway to obtain standard JSON format semantic data; Specifically, a data docking mechanism is established inside the edge gateway. That is, through the docking unit, the protocol abstraction and semantic conversion module is docked with the protocol adaptation engine to obtain the original protocol data receiving channel. This receiving channel listens to and receives the underlying protocol data from various devices at the building site in real time. At the same time, based on the concept of point cloud digital twin, a systematic hierarchical design is carried out on the semantic model of building information. That is, driven by the hierarchical unit, a five-layer structure model composed of a space model layer, a device model layer, a measurement point model layer, a service model layer, and an association model layer is constructed. The space model layer is used to describe the physical space structure of the building, including spatial objects such as floors, areas, rooms, and their spatial nesting relationships; the device model layer defines the types, key attributes, and function labels of various devices in a standardized manner, facilitating unified management and classified invocation of different devices; the measurement point model layer is refined to the specific measurement points or control points of each device, including detailed information such as data type, unit, range, and alarm threshold; the service model layer carries the operating status, control mode, priority policy, and linkage service rules of the device, effectively associating device data with the actual building service scenario; the association model layer clarifies the multi-level and multi-dimensional mapping relationships between space, devices, and measurement points, realizing the free flow and linkage of data among the physical space, functional modules, and business requirements. Based on the semantic model of the above hierarchical structure, the extraction unit identifies the format and extracts the valid values of the heterogeneous protocol data such as BACnet / IP and Modbus-TCP obtained by real-time parsing of the protocol adaptation engine through the original protocol data receiving channel. This step involves the compatible adaptation and data verification of various message formats, and extracts the original measurement values, device status, and adjustable control parameters of each device, including multiple information such as temperature and humidity, current and voltage, access control status, energy consumption data, and operating mode. Under the guidance of the five-layer hierarchical structure model, the mapping unit performs a standardized semantic mapping on the above device original measurement values and control parameters to obtain semantically unified device data. The encapsulation unit structurally encapsulates the data that has completed semantic standardization in the standard JSON format to generate standard JSON format semantic data that is easy to transmit, analyze, and share across systems.

[0023] Based on the concept of point cloud digital twin, a systematic physical space structure definition is carried out for the entire spatial model layer. According to the actual distribution and physical nesting relationship of buildings, floors, regions, and rooms, a set of standardized spatial identifier coding rules are formulated, and a set of spatial identifiers that clearly express spatial hierarchy, attribution, and location is constructed in the system. The device model layer is standardized, that is, a unified description specification for the independent type, attributes, function labels, and operating status of each type of device is formulated, and a standardized device identifier set including HVAC devices, lighting devices, elevator systems, security devices, and energy metering devices is constructed. In the measurement point model layer, the mapping unit sets standard identifiers for all collectible and controllable data points, and clarifies the data type, unit, value range, and alarm threshold of each measurement point, and constructs a standardized measurement point description library. Based on the aforementioned standardized measurement point description library, the mapping unit establishes a two-way mapping relationship between the object identifier and attribute identifier in the BACnet / IP protocol, as well as the register address in the Modbus-TCP protocol, and the standard measurement point identifier one by one, and generates a conversion dictionary based on this to realize the two-way conversion of the protocol to semantics of the data point. During the actual operation process, according to the conversion dictionary, the protocol-specific identifiers of the original measured values of the collected devices are replaced and the data type is converted. All data points are finally presented in the system with unified measurement point identifiers and clear data types, thus eliminating the obstacles of protocol fragmentation and data heterogeneity. On this basis, the mapping unit performs unit standardization processing on the converted device parameters. For example, parameters such as temperature, pressure, and electric energy are all unified into international standard units to avoid data chaos and business risks caused by inconsistent units. At the same time, the mapping unit strictly verifies the data validity to ensure that each collected or generated device data is within the physically allowed range and logical threshold, and abnormal, invalid, or distorted data is removed in real time. All device parameters processed through the above are finally provided with unified measurement point identifiers, standardized data types and units, and are multi-dimensionally associated with spatial identifiers and device identifiers to form semantically unified device data.

[0024] The rule processing subsystem 104 is used to input the standard JSON format semantic data into the proximal decision engine of the edge gateway for rule processing to obtain the local intelligent response result; Specifically, a set of local functional module systems for intelligent autonomy and high-resilience service guarantee is deployed inside the proximal decision-making engine of the edge gateway. This system includes a local state maintenance module, a preset rule execution module, a lightweight inference module, and a breakpoint resumption module. The system receives and parses the standard JSON format semantic data output from the protocol conversion subsystem in real time, inputs the structured device data into the local state maintenance module, and utilizes its high-frequency dynamic maintenance ability for various terminal device states to continuously update and persist key information such as device operation parameters, alarm status, online situation, and mode switching, forming a real-time device state database. The preset rule execution module performs conditional matching between the parameters in the real-time device state database and the Boolean expressions in the rule library based on the preset business scenario rules, and automatically filters out the operation instruction sequences to be executed. For complex anomalies or optimization requirements that cannot be fully described by simple Boolean conditions, the abnormal parameters or multi-dimensional abnormal features detected in the real-time device state database are input into the lightweight inference module, and the built-in knowledge distillation neural network model is used for local anomaly detection and pattern recognition. At the same time, the dynamic parameters and optimization strategies sent from the cloud are locally adapted and fine-tuned to achieve more forward-looking and intelligent local control. For example, when the device energy consumption suddenly rises but does not trigger the alarm threshold, the model adjusts the control parameters in advance and optimizes the operation strategy, effectively improving the device operation efficiency and reliability. According to the urgency of the event type, business priority, and impact range, the above-obtained operation instruction sequences and intelligent optimization control instructions are sorted by priority to form a local priority-sorted execution instruction set to be executed. To ensure the traceability of the execution process and business continuity, all execution instructions are tracked for status and logged for execution locally, and the breakpoint resumption module caches the key execution status, important device data, and control results at different levels, so that even when the network fluctuates or even breaks, the edge gateway can continuously guarantee the independent operation of the local decision-making mode and perform millisecond-level local intelligent processing on key services such as fire linkage, elevator safety, and security response. When the network resumes, the breakpoint resumption module synchronizes the local cached operation history, data status, fault logs, etc. with the cloud efficiently according to the instruction priority and event urgency, ensuring the global consistency, data integrity, and analysis continuity of the cloud-edge-end system and obtaining the local intelligent response results.

[0025] The multi-dimensional analysis subsystem 105 is used to transmit the standard JSON format semantic data and local intelligent response results to the cloud large model engine through the TLS encrypted two-way communication channel for multi-dimensional analysis, obtaining the device health score, fault prediction results, and energy optimization strategies.

[0026] Specifically, the deployment of a deeply customized large model engine is completed in the cloud. This process is based on a multi-functional large model engine with LLaMA as the infrastructure and enhanced by in-depth fine-tuning and industry knowledge. It specifically includes a data processing module, an intelligent analysis module, a decision-making generation module, a natural language interaction module, and a model optimization module. Through the role of the deployment unit, this large model engine has the ability to handle large-scale parameters and recognize complex patterns, and combines the knowledge base, industry specifications, and equipment historical work order data in the building operation and maintenance field to achieve professional and intelligent in-depth analysis and prediction optimization capabilities. During the actual operation of the system, the data processing unit inputs the semantic data in the standard JSON format reported by the edge gateway through the TLS encrypted two-way communication channel, as well as the local intelligent response results formed after edge intelligent decision-making, into the data processing module of the large model engine in real-time or in batches. This module performs automated data cleaning on multi-source heterogeneous data, including outlier removal, duplicate data merging, and noise interference filtering, and performs standardized mapping processing on the data structure according to unified data standards and protocol rules. It performs time series alignment on the collected multi-dimensional time series data to ensure that data from different devices, different sampling frequencies, and cross-protocols can be accurately synchronized under the same time line. At the same time, it performs in-depth feature extraction by combining multi-dimensional tags such as building space, equipment type, and business scenarios, and converts the original data into target time series data with global spatio-temporal consistency. The intelligent analysis unit inputs the target time series data into the intelligent analysis module, and uses the built-in deep learning network structure of the large model to perform energy consumption pattern recognition, equipment performance evaluation, abnormal behavior detection, and multi-level fault feature extraction on the operation data of equipment groups and various subsystems. Through the vertical and horizontal analysis of energy consumption and performance, it identifies high-energy consumption patterns in each region and atypical states in equipment operation, and captures early abnormal signals and trend-based fault hazards. The intelligent analysis results are output in the form of equipment health scores, and each device is assigned a real-time updated health index to help the system quantify the operation and maintenance risks and priorities in real-time. On this basis, the decision-making generation unit inputs the above-mentioned equipment health scores and multi-source heterogeneous information such as historical maintenance records and professional knowledge base data into the decision-making generation module. This module relies on knowledge reasoning, data association, and the attribution ability of the large model to automatically carry out fault diagnosis, root cause analysis, and predictive maintenance decision-making calculations, enabling the system to discover potential faults in advance and accurately locate the causes. At the same time, it generates personalized and forward-looking maintenance suggestions and operation instructions for different devices and scenarios, improving the accuracy of fault warning and the efficiency of maintenance response. The multi-objective optimization unit uses the fault prediction results as constraints, and executes multi-objective optimization algorithms through the decision-making generation module. Considering multiple operation and maintenance objectives such as equipment health status, energy consumption level, user comfort, system security, and economic cost, it intelligently calculates and generates energy optimization strategies for the entire building or specific subsystems.The strategy includes fine-grained control schemes such as the device start-stop sequence, temperature control parameter adjustment, and load dynamic allocation, and adapts to external environmental changes and business requirements in real time. Finally, the intelligent achievements of the multi-dimensional analysis subsystem are accurately sent to the edge gateway and terminal devices through the closed-loop link of cloud-edge collaboration, enhancing the self-perception, self-optimization, and self-adaptation capabilities of smart buildings.

[0027] In the cloud large model engine, the original LLaMA pre-trained model is architecturally adjusted and initialized with configuration, including optimizing structural elements such as the original model parameters, number of network layers, activation functions, etc. Combining with the specific scenarios in the building field, the input-output interfaces of the model, the context window during inference, as well as the inference accuracy and speed are adaptively adjusted to construct a dedicated initialized LLaMA base model for building operation and maintenance tasks. To ensure that the model has the knowledge depth and semantic expression ability in the building field, around real engineering operation and maintenance scenarios, multi-source heterogeneous knowledge materials such as building equipment manuals, historical operation and maintenance records, industry standard documents, and typical maintenance cases are systematically cleaned and structured. Through artificial intelligence annotation or automated information extraction methods, the original text of various documents is summarized into structured entities and attributes such as equipment parameters, fault characteristics, maintenance strategies, control logics, etc., and then a building operation and maintenance training corpus is constructed. Based on the building operation and maintenance training corpus, the initialized LLaMA model is supervised and fine-tuned and trained with reinforcement learning. In the supervised learning stage, through the "input-output" pair samples manually annotated, the model's understanding and inference results of professional problems are continuously corrected, enabling the model to accurately learn the functional parameters, fault diagnosis paths, maintenance processes, and operation suggestions of building equipment. In the reinforcement learning stage, an evaluation and reward mechanism is introduced to promote the model to continuously optimize its performance in tasks such as policy generation, intelligent analysis, and human-computer interaction. At the same time, combined with parameter-efficient fine-tuning techniques (such as LoRA, Adapter, etc.), only the key weights inside the model are selectively updated, ensuring the model's rapid absorption of professional knowledge, reducing the computing power consumption and the hardware burden during model training, and obtaining an efficient pre-fine-tuned model applicable to the building vertical domain. Through knowledge enhancement technology, the organized building professional knowledge graph data is deeply integrated with the pre-fine-tuned model. This process includes structural knowledge transfer based on the knowledge distillation algorithm, embedding knowledge such as equipment associations, process flows, and experience rules in the graph into the multi-layer neural network of the model, and using model compression technology to refine the large model into a lightweight version with a compact structure, easy to deploy and perform local inference, ensuring the efficiency and usability of the model in the actual operation and maintenance environment. At the same time, to improve the multi-dimensional task adaptability of the model, it is trained with multiple tasks, covering various core operation and maintenance scenarios such as data analysis, equipment prediction, policy generation, natural language interaction, intelligent question answering, etc., enabling the model to support complex decision-making inferences and efficient human-computer conversations, and obtaining a large model for the vertical domain. Based on the above technical link and optimization process, a multi-functional large model engine supported by the deeply fine-tuned LLaMA architecture is constructed in the cloud large model engine. This engine integrates five core units, namely a data processing module, an intelligent analysis module, a decision-making generation module, a natural language interaction module, and a model optimization module, realizing a full-process closed-loop of human-machine collaboration from raw data collection to complex intelligent decision-making, from fault self-diagnosis to automatic strategy optimization and then to natural language interaction, and deeply integrating building operation and maintenance knowledge and AI intelligent capabilities.

[0028] Establish an early warning and alarm mechanism based on equipment health scores in the system architecture. That is, for the health score results of the large model of each building equipment, according to the equipment type, operation risk, and operation and maintenance requirements, set multi-level early warning thresholds and alarm thresholds to obtain a set of health management trigger conditions refined to equipment and scenarios. When the health score of the equipment drops to the early warning threshold, the system starts the corresponding maintenance prompt, and when it drops to the alarm threshold, it automatically enters the high-priority operation and maintenance process to ensure early warning and early intervention of equipment failures. Based on the health management system, dynamically combine the fault prediction results generated by the large model analysis module with historical fault data, and through the equipment remaining service life algorithm, evaluate the future risk window and maintenance time of each equipment, and then automatically generate predictive maintenance work orders, including equipment risk assessment, recommended maintenance time window, required spare parts, and specific operation steps, to achieve the upgrade from passive repair to active maintenance, and reduce unplanned downtime and operation and maintenance costs. At the same time, intelligently automate the implementation process of the energy optimization strategy, and dynamically integrate the energy optimization strategy generated by the cloud large model engine based on multi-dimensional goals such as equipment energy consumption, building comfort, and economy, combined with weather forecast data, energy prices, predicted pedestrian flow, and historical usage patterns, to form a rolling optimization control strategy. This strategy dynamically adjusts the operation parameters of the main energy-consuming equipment in the building, such as air conditioner start and stop, temperature setting, lighting time period, and elevator dispatching, to ensure the best balance between energy utilization rate and operation comfort. For the convenience of actual control execution, automatically convert the rolling optimization control strategy into dynamic control instructions recognizable by the device layer to ensure the immediacy and consistency of end-to-end control. In terms of ensuring equipment safety, deeply analyze the access modes and real-time network traffic characteristics of all equipment inside the building, and based on technical means such as anomaly detection, protocol compliance, and behavior profiling, dynamically generate security protection strategies, including access authentication, traffic throttling, instruction whitelist, attack protection rules, etc., to effectively prevent potential network threats and malicious intrusions. To meet the operation and maintenance requirements in different business scenarios, the execution subsystem automatically divides tasks into three working modes: automatic execution mode (such as self-repair of faults and self-adjustment of energy consumption), recommended approval mode (such as manual confirmation required for the maintenance of high-value equipment), and manual intervention mode (such as manual intervention required for emergency security events or complex faults), to form a classification execution strategy set. The entire classification execution strategy set securely sends the dynamic control instructions to the corresponding edge gateway through the TLS encrypted two-way communication channel. The edge gateway combines its own protocol adaptation and control engine to accurately convert the upper-layer control instructions into the corresponding device protocol format to achieve precise scheduling and real-time response to terminal equipment. The feedback results after the equipment execution are transmitted back to the cloud along the original channel to achieve closed-loop tracking and dynamic optimization of the entire process.

[0029] In one embodiment, the configuration subsystem 101 is specifically configured to: Deploy a network proxy server cluster in the cloud to obtain an NPS server cluster, and install a network proxy client NPC in the edge gateway device within the intelligent building; Authenticate the TCP / TLS long connection requests initiated by the edge gateway device through the NPS server cluster to obtain an active connection registry; Establish a reverse proxy channel for the edge gateway device based on the active connection registry to obtain a NAT penetration communication link, and perform TLS 1.2 encryption processing on the NAT penetration communication link to obtain a TLS encrypted two-way communication channel.

[0030] Specifically, guided by the global requirements of building intelligent operation and maintenance, a network proxy server cluster, namely the NPS (Network Proxy Server) server cluster, is deployed in a highly reliable and highly available physical or virtual machine environment in the cloud. This cluster has the ability to scale horizontally and load balancing characteristics, and at the same time supports redundant deployment to ensure automatic switching and service recovery in case of a single point of failure. The main role of the server cluster is to provide unified, manageable, and elastically scalable reverse proxy and secure communication services for a large number of edge gateway devices in many buildings. Each NPS server is pre-installed with a reverse proxy service program that supports multi-threading and high concurrency, has encryption capabilities of TLS 1.2 and above, and can open specified access ports to the outside world to handle external connection requests from edge gateways. Inside the smart building, one or more edge gateway devices are deployed for each building or each device concentration area. These edge gateways need to be embedded with a dedicated network proxy client NPC (NetworkProxy Client). As an independent daemon process, this client starts with the gateway main system and automatically initiates a long TCP connection request after detecting the network reachability of the NPS server cluster in the cloud at regular intervals. Each NPC has a unique device ID and identity credentials, and automatically submits the certificate through the TLS handshake process when initiating a connection to achieve two-way identity authentication. After receiving the connection request from the NPC, the NPS server cluster verifies the client's identity according to the pre-established authentication mechanism, including processes such as certificate validity verification, certificate chain check, and revocation list check. Only the edge gateways that pass the authentication are allowed to establish a persistent connection and enter the subsequent data channel allocation process. When the identity authentication is passed, the NPS server writes key information such as the unique identifier, network address, communication status, last heartbeat timestamp, and certificate information of the current NPC into an actively maintained active connection registry. This registry is stored distributively in the cloud and supports multi-node concurrent updates to ensure accurate tracking of the communication status of all edge gateways in scenarios of multi-point deployment and high-concurrency requests. The information in the registry reflects the online status of each building edge gateway in real time, providing basic data support for the cloud management platform to achieve remote monitoring, health inspection, data scheduling, and fault location. The active connection registry can also automatically detect situations such as connection timeouts and abnormal disconnections. If it is found that a certain NPC heartbeat packet is lost or there is no data interaction for a long time, its status will be automatically marked as offline or abnormal, and relevant operation and maintenance personnel will be notified to conduct a timely investigation. Based on the active connection registry, the cloud NPS server cluster allocates a dedicated reverse proxy channel for each authenticated and online edge gateway as needed.The cloud NPS server reserves a set of virtual channel identifiers and a dynamic communication port pool for reverse data forwarding. Whenever remote access to the devices inside a certain building is required, the operation and maintenance management platform, through the NPS cluster routing management unit, assigns a unique session channel to the target NPC, encapsulates the downstream instruction data into an encrypted data packet, and forwards it to the target NPC through the existing TLS long connection of the NPS server. After receiving the data packet, the NPC automatically sends the instruction to the specific terminal device inside the building according to the protocol adaptation and device distribution rules. The response data of the device is also actively uploaded to the NPS server by the NPC to complete the end-to-end two-way communication and obtain the NAT penetration communication link. For each reverse proxy channel established through NAT penetration, the NPS and the NPC use an encryption protocol of TLS 1.2 or a higher version to perform end-to-end encryption on the transmission link. TLS encryption not only encrypts and protects the communication content, but also supports two-way authentication, message integrity verification, and replay attack prevention mechanisms, which can effectively prevent network threats such as man-in-the-middle attacks, data hijacking, forged instructions, and session hijacking. The establishment process of the TLS connection includes key negotiation, certificate exchange, encryption suite agreement, and communication handshake, and after successful establishment, it maintains a long-term stable connection in a secure data stream mode. The NPS server also supports automatic certificate update, revocation, and rotation mechanisms to ensure that there are no security risks caused by certificate expiration or leakage during long-term operation. To ensure the real-time performance and high availability of communication, the NPS cluster supports automatic load balancing, distributes connection requests from different buildings to each node, and avoids bottlenecks caused by excessive pressure on a single server. At the same time, the NPC client implements a heartbeat mechanism and a disconnection reconnection function. If it detects an abnormal interruption of the TLS long connection, it automatically attempts to re-establish a secure connection with other server nodes in the NPS cluster to ensure that the data channel is uninterrupted and the operation and maintenance link is always unblocked. Through the above steps, a TLS-encrypted two-way communication channel is obtained.

[0031] In one embodiment, the construction subsystem 102 is specifically configured to: Deploy and configure the cloud infrastructure to obtain a cloud large model engine; Deploy a building-specific protocol adaptation engine, a protocol abstraction and semantic conversion module, a data preprocessing and caching module, an edge security gateway, and a proximal decision-making engine in the edge gateway; Connect heating, ventilation, and air conditioning equipment, intelligent lighting equipment, elevator control equipment, video surveillance equipment, intrusion alarm equipment, access control management equipment, and energy metering equipment to the edge gateway to obtain the terminal device layer; Configure the data closed-loop process of the cloud large model engine, the edge gateway, and the terminal device layer through the TLS-encrypted two-way communication channel to build a cloud-edge-end three-layer collaborative computing architecture.

[0032] Specifically, deploy and configure the cloud infrastructure, and adopt a highly available and highly scalable data center environment as the hosting platform. On this basis, use modern cloud computing means such as containerization technology, microservice architecture, and high-performance distributed storage to build a large model engine suitable for the vertical field of building operation and maintenance. The large model engine is based on a deep neural network, uses an advanced large model such as LLaMA or others as the kernel, and is targeted for fine-tuning and multi-task optimization training through an industry knowledge graph and a large amount of building operation and maintenance corpus, enabling it to have multiple capabilities such as understanding complex building system time-series data, semantic modeling of cross-protocol data, fault prediction and diagnosis, energy consumption optimization decision-making, and natural language interaction. The cloud large model engine not only needs to be deeply integrated with the operation and maintenance platform and data service APIs, but also needs to deploy multiple modules such as data reception and storage, knowledge management, intelligent analysis, control strategy generation, and model self-evolution to ensure real-time parallel processing, deep reasoning, and efficient feedback of massive building data, and support the operation and maintenance needs of large-scale, multi-building, and multi-subsystem across the network. At the same time, at the on-site edge side of each smart building, deploy a fully functional edge gateway. The gateway is not only a bridge between on-site devices and the cloud large model, but also undertakes multiple responsibilities such as protocol compatibility, data abstraction, security protection, and local intelligence. The edge gateway is embedded with a building-specific protocol adaptation engine, which supports automatic identification, message parsing, and real-time communication of mainstream standard protocols such as BACnet / IP, Modbus, OPC UA, GB28181, and vendor private protocols, ensuring that various devices from different manufacturers and different subsystems can be seamlessly connected to the unified management system. The supporting protocol abstraction and semantic conversion module further converts the collected raw protocol data into a structured and unified semantic model, and establishes a five-layer model of space, equipment, measurement points, services, and associations based on the concept of point cloud digital twin, mapping the underlying heterogeneous "dialect" data into structured data with standard tags, units, and timestamps. The data preprocessing and caching module performs preprocessing tasks such as cleaning, noise reduction, anomaly detection, and aggregation statistics on the raw data locally, effectively reducing the network transmission pressure, and at the same time performing data buffering and breakpoint resumption locally to provide technical guarantees for business continuity and data integrity in extreme scenarios such as network fluctuations and disconnections. The edge security gateway module authenticates device access, executes communication encryption and access control policies, and real-time identifies, isolates, and alarms abnormal traffic and abnormal instructions to ensure the coordinated control of network space security and physical device security in the entire building site. The proximal decision-making engine, based on the local cache status, preset rules, emergency linkage logic, optimization parameters sent from the cloud, and lightweight AI models, realizes local real-time fault response, linkage control, and optimization decision-making. Even in the case of temporary disconnection from the cloud, it can still ensure the autonomous management and millisecond-level emergency response of core services such as fire protection, elevators, and security.After the edge gateway architecture is completed, various intelligent terminal devices are connected to the edge gateway to form a terminal device layer, covering multiple subsystems such as heating, ventilation, and air conditioning systems, intelligent lighting circuits, elevator group control units, high-definition video surveillance cameras, intrusion alarm detectors, access control management systems, and energy metering instruments. The access methods of the devices are various wired or wireless solutions such as Ethernet, RS485, CAN bus, Zigbee, LoRa, or WiFi, ensuring full compatibility and easy expansion between the physical layer and the network layer of the devices. Each terminal device is centrally managed by the edge gateway. The device operating status, real-time parameters, event data, and control commands are all subject to protocol adaptation and preliminary processing locally at the edge, achieving full data collection, full instruction coverage, full event tracking, and full linkage response. A data closed-loop process for the cloud large model engine, edge gateway, and terminal device layer is configured through a TLS encrypted two-way communication channel. The TLS 1.2 and above protocols ensure the confidentiality, integrity, and identity authenticity of the data during transmission. All device-collected data, alarm information, and local optimization results are encrypted in real time by the edge gateway and uploaded to the cloud model engine. The cloud conducts multi-dimensional cleaning, time series alignment, in-depth analysis, and strategy optimization on the data. Based on the intelligent analysis results and global operation objectives, rolling optimization control strategies, fault warnings, predictive maintenance suggestions, and dynamic security protection rules are generated and then securely sent down to the corresponding edge gateway through the TLS channel. After receiving the instructions sent down by the cloud, the edge gateway automatically converts the control strategy into a local protocol recognizable by the terminal device in combination with the actual operating conditions and protocol compatibility of the local device, ensuring accurate execution within the shortest delay. At the same time, the edge gateway sends the execution results, on-site feedback, and abnormal data back to the cloud again, forming a complete data and control closed loop of "terminal-edge-cloud-edge-terminal".

[0033] In one embodiment, the conversion subsystem 103 further includes: A docking unit for docking the protocol abstraction and semantic conversion module in the edge gateway with the protocol adaptation engine to obtain a receiving channel for the original protocol data; A layering unit for performing a layered design on the unified semantic model based on the concept of point cloud digital twin, and constructing a five-layer structure model including a space model layer, a device model layer, a measurement point model layer, a service model layer, and an association model layer; An extraction unit for identifying the data format and extracting valid values from the heterogeneous protocol data of BACnet / IP and Modbus-TCP parsed by the protocol adaptation engine through the receiving channel to obtain the original device measurement values and control parameters; A mapping unit for performing a standardized mapping on the original device measurement values and control parameters based on the five-layer structure model to obtain device data with unified semantics; An encapsulation unit for encapsulating the device data with unified semantics in JSON format to obtain standard JSON format semantic data.

[0034] Specifically, an efficient data flow mechanism is implemented inside the edge gateway platform of the smart building. The docking unit deeply docks the protocol abstraction and semantic conversion module inside the gateway with the actual protocol adaptation engine. The protocol adaptation engine integrates the message parsing capabilities of multiple building control and automation protocols such as BACnet / IP, Modbus-TCP, OPC UA, GB28181, etc. These protocols have many differences in actual applications, including data frame structure, addressing method, data type and object mapping. Through the process arrangement of the docking unit, the protocol adaptation engine pushes the original protocol data stream collected from the site directly to the semantic conversion module to form a high-throughput, low-latency original protocol data receiving channel. After receiving the multi-protocol original data stream on site, the system relies on the hierarchical unit to realize the hierarchical design and dynamic maintenance of the unified semantic model of the building. This process is based on the concept of point cloud digital twins as the theoretical core. Point cloud digital twins mean one-to-one mapping of physical devices and virtual models, as well as structured and hierarchical expression of multi-dimensional information such as space, equipment, measurement points, business, and associations. The spatial model layer abstracts the physical space of the building, clarifies the hierarchy, ownership and spatial identifiers of physical spaces such as buildings, floors, areas and rooms, and provides a unified coordinate system for the spatial positioning, regional management and cross-building business linkage of all equipment data. The equipment model layer performs structural modeling according to the type, brand, main function, topological ownership and other standards of the equipment, and realizes the unified description and identification of heterogeneous equipment such as HVAC equipment, lighting control units, elevator systems, security subsystems, and energy metering terminals. The measurement point model layer defines the monitoring and control points of the equipment in more detail. Each measurement point is described in detail with unique ID, standard unit, data type, collection frequency, alarm threshold and other elements, providing a high-precision reference for data collection and abnormal monitoring. The business model layer carries complex business information such as business logic, operating status, control mode, priority strategy, linkage rules, etc. between equipment and space, and measurement points, and realizes the full expression of intelligent operation and maintenance scenarios such as equipment life cycle, operation scenario, energy consumption allocation, and event flow. The association model layer ultimately connects all mappings and relationships between space, equipment, measurement points and business, allowing the system to complete data aggregation, event tracing and global optimization from a multi-dimensional perspective. After the layered semantic model is established, the extraction unit uses the aforementioned original protocol data receiving channel to process the multi-protocol data stream transmitted by the protocol adaptation engine in real time. The unit quickly identifies the data format and extracts valid values of mainstream protocols such as BACnet / IP and Modbus-TCP, involving accurate parsing of protocol fields such as message headers, function codes, object identifiers, register addresses, and data areas.For BACnet / IP, map object types, object instance numbers, attribute IDs, etc. to the actual monitoring parameters or control points of the device; for Modbus-TCP, extract the original measurement values and control parameters such as temperature, humidity, current, voltage, status quantities, etc. according to the register address and register type (such as holding register, coil, input register, etc.). The extraction unit must be compatible with issues such as private extensions of devices from different manufacturers, unit conversions, and byte order differences, ensure the original accuracy and usability of the data, and have a complete fault tolerance and logging mechanism for abnormal data, null data, verification failures, etc. The original measurement values and control parameters after extraction should all carry metadata such as complete spatio-temporal tags, device ID, measurement point ID, data quality tags, etc. After obtaining the original device data, the mapping unit performs standardized mapping according to the unified five-layer structure model. The mapping unit searches the mapping rule library in the hierarchical model and converts underlying protocol features such as BACnet object identifiers and Modbus register addresses into unified measurement point identifiers, data types, and business meanings. The rules in the business model layer combine the measurement point with scenarios such as energy consumption analysis, device scheduling, and environmental linkage to obtain device data with unified semantics. The encapsulation unit encapsulates the information such as the mapped device measurement values, control parameters, spatio-temporal metadata, data quality tags, etc. in a unified JSON data structure. Each standard JSON format semantic data contains key fields such as device unique identifier, measurement point identifier, acquisition timestamp, standard unit, data value, data quality, spatial attribution, business tags, etc.

[0035] In one embodiment, the mapping unit is specifically used for: Define the building physical space structure for the spatial model layer and construct a set of standardized spatial identifiers including buildings, floors, areas, rooms and their hierarchical relationships; Define the device type, attributes, functions and status descriptions for the device model layer and construct a set of standardized device identifiers including HVAC devices, lighting devices, elevator systems, security devices, energy metering devices; Define the standard identifiers, data types, units, value ranges, and alarm thresholds for the measurement point model layer to obtain a standardized measurement point description library; Based on the standardized measurement point description library, construct a mapping table for the object identifiers and attribute identifiers in the BACnet / IP protocol and the register addresses and standard measurement point identifiers in the Modbus-TCP protocol to obtain a conversion dictionary; Replace the protocol-specific identifiers of the device original measurement values according to the conversion dictionary and perform data type conversion to obtain device parameters with unified identifiers and data types; Perform unit standardization processing and validity verification on device parameters with unified identifiers and data types, and at the same time associate them with the set of standardized spatial identifiers and the set of standardized device identifiers to obtain semantically unified device data.

[0036] Specifically, guided by the concept of point cloud digital twin, a hierarchical description of the spatial model layer and a standardized identifier system are constructed around the spatial hierarchical structure of the actual building scenario. Each intelligent building project contains one or more buildings, each building is divided into several floors, and each floor is further subdivided into multiple physical areas, such as machine rooms, office areas, public corridors, etc. Each area is further refined into specific rooms, independent functional units or equipment concentration points. In order to achieve the uniqueness and traceability of spatial objects, coding rules are defined for each level of spatial objects, and all spatial objects are archived into a standardized spatial identifier set. This set clearly defines the ID, name, type of each spatial object, and maintains hierarchical relationships (such as parent-child relationships, spatial subordination, cross-region mapping, etc.), realizing the tree-like or graph-structured storage of spatial information. The device model layer is described and defined in terms of device type, attributes, functions, and status. All intelligent devices in the building are classified by type, and the attributes, functions, and operating status of the devices are clarified. All HVAC devices (such as air conditioners, fan coil units, chillers), intelligent lighting devices (such as loop controllers, light sources), elevator systems (such as carriages, door machines, control cabinets), security devices (such as access control card readers, alarm hosts, cameras), energy metering devices (such as smart meters, water meters, heat meters), etc. are respectively established into standardized device categories, and a unique device identifier is assigned to each device. The identifier is automatically generated in combination with meta-information such as device type, building location, and installation sequence. Each device object has a complete set of attributes, including manufacturer, communication protocol, main functions, installation location, current operating status (such as ON / OFF, alarm / normal, offline / online), etc., which are summarized into a standardized device identifier set. The measurement point model layer is defined in terms of standard identifiers, data types, units, value ranges, and alarm thresholds. For each type of device, all measurable and controllable measurement points are listed, such as temperature, humidity, pressure, air volume, current, voltage, power, switch status, alarm flag, etc., and a standard identifier is assigned to each measurement point. Each measurement point definition contains parameter information such as data type (such as integer, floating point, boolean), data unit (such as °C, %RH, kW, A, V), value range (such as -40 - 100 °C, 0 - 100%), alarm threshold (such as temperature exceeding 30 °C triggers an alarm), etc., which are aggregated into a standardized measurement point description library. In order to achieve seamless conversion of protocol layer data to the standard measurement point model, based on the aforementioned measurement point description library, a mapping table, that is, a conversion dictionary, is established for the object identifiers and attribute identifiers in the BACnet / IP protocol and the register addresses in the Modbus-TCP protocol. This conversion dictionary stipulates the correspondence between each protocol field and the standard measurement point identifier. Through the conversion dictionary, the protocol parsing engine accurately assigns each piece of raw protocol data collected from the device to the standard measurement point system, realizing protocol-independent full-scenario data adaptation.During the data flow process, the protocol adaptation engine first replaces the protocol-specific identifiers and converts the data types of the collected original measurement values according to the conversion dictionary to obtain device parameters with unified identifiers and data types. The device parameters with unified identifiers and data types are subjected to unit standardization processing and validity verification. All data is converted into the physical quantity units uniformly specified by the platform to avoid unit inconsistencies caused by different protocols, manufacturers, or regions. Validity verification includes data range checking, data format inspection, duplicate collection deduplication, missing value handling, and outlier filtering to ensure that each piece of data is physically and operationally reasonable and truly reflects the actual state and operating environment of the device. The device parameters that have been uniformly processed in terms of identifiers, data types, units, time-space, and service tags are multi-dimensionally associated with the standardized spatial identifier set and the standardized device identifier set to obtain semantically unified device data.

[0037] In one embodiment, the rule processing subsystem 104 is specifically configured to: Deploy a local status maintenance module, a preset rule execution module, a lightweight inference module, and a breakpoint resumption module in the proximal decision engine of the edge gateway; Input the standard JSON format semantic data into the local status maintenance module for status update processing to obtain a real-time device status database; Perform Boolean expression condition matching on the parameter values in the real-time device status database to obtain a sequence of operation instructions to be executed; Input the abnormal parameters in the real-time device status database into the lightweight inference module for knowledge distillation neural network calculation, perform local anomaly detection and optimization adjustment based on the parameters sent from the cloud to obtain intelligent optimization control instructions; Perform priority sorting on the sequence of operation instructions and the intelligent optimization control instructions according to the event type to obtain an execution instruction set with priority sorting; Track the execution status of the execution instruction set with priority sorting, and locally cache important data during the execution process through the breakpoint resumption module to ensure the operation of the local decision-making mode during network interruption. After the network is restored, synchronize with the cloud in the order of priority to obtain a local intelligent response result.

[0038] Specifically, a decision-making module system for real-time autonomy and high robustness is locally deployed at the edge gateway. It includes four core components: the local status maintenance module, the preset rule execution module, the lightweight inference module, and the breakpoint resumption module. The local status maintenance module serves as the data foundation for the entire intelligent decision-making process. It continuously obtains the semantic data stream in the standard JSON format from the protocol conversion subsystem and writes this structured data into the device status database in real time according to multiple dimensions such as devices, measurement points, time, and space. This database contains the latest operating parameters, status bits, alarm signals, and mode information of all building terminal devices, and performs version management, timestamp synchronization, and redundancy check for each data change to ensure that the edge side can always grasp the most complete and real-time on-site operation situation even when the network environment fluctuates. The status maintenance module also supports the periodic or event-driven data snapshot mechanism, which is convenient for historical status tracing and intelligent fault tracking. Based on the continuous dynamic maintenance of the real-time device status database, the preset rule execution module, as the local decision-making master, automatically performs rule matching and condition discrimination on the various device parameters in the database through the built-in business rule library, Boolean logic parsing engine, and status linkage table. The module supports multi-level nesting, condition priority, periodicity, and parallel judgment triggered by events. After each round of condition discrimination, a set of operation instruction sequences that need to be executed immediately or delayed are generated. These instructions cover various scenarios such as device switch control, parameter adjustment, security linkage, and energy consumption optimization. For complex abnormal scenarios that cannot be fully characterized by Boolean conditions alone, have multi-factor implicit relationships, or require in-depth reasoning, the system automatically inputs the data parameters determined to be abnormal in the real-time device status database into the lightweight inference module. This module is based on the local knowledge distillation neural network and integrates the device health scoring model, abnormal diagnosis model, and policy optimization model sent by the cloud large model through knowledge transfer. It can perform edge-side reasoning on multi-source features such as device operating status, environmental variables, and historical abnormal behaviors, identify high-order business risks such as device performance degradation, potential fault hazards, and abnormal energy consumption growth, and calculate the optimal intelligent optimization control instructions at the current moment in combination with the optimal operating parameters and adjustment strategies sent by the cloud. The operation instruction sequences and intelligent optimization control instructions are sorted according to their priorities according to the event type. According to multiple weight parameters such as event type, business impact scope, real-time requirement, and security level, all control instructions to be executed are dynamically sorted. For example, fire and elevator emergency stop instructions involving personal safety require the highest priority response; those involving energy consumption optimization and comfort fine-tuning are postponed or processed in batches. The sorted execution instruction set is sent to the device execution module in descending order, and the execution status of each instruction (such as sent, device response, abnormal feedback, etc.) is written into the status database in real time.To adapt to extreme operating environments such as common network fluctuations and interruptions on the edge side, the breakpoint resumption module locally caches all high-priority execution instructions and their feedback statuses, core device operation parameters, and exception alarm information. It uses a hierarchical data queue, transaction log, and retransmission mechanism to record each issued instruction and each receipt locally, without relying on real-time cloud synchronization. When the network is detected to be disconnected or the edge is disconnected from the cloud, the breakpoint resumption module automatically switches to the local autonomous decision-making mode to ensure that the edge gateway can independently and continuously complete device control, alarm processing, and service self-healing. When the network resumes, the breakpoint resumption module synchronizes the locally cached execution results, exception reports, status snapshots, and other data to the cloud in an orderly manner according to the instruction priority, data generation time sequence, and business urgency, minimizing information loss and business fragmentation, and supporting the cloud to perform a complete backtrace and accurate verification of the edge operation process.

[0039] In one embodiment, the multi-dimensional analysis subsystem 105 further includes: A deployment unit for deploying a multi-functional large model engine based on the deep fine-tuned LLaMA architecture, which includes a data processing module, an intelligent analysis module, a decision generation module, a natural language interaction module, and a model optimization module, in the cloud large model engine; A data processing unit for inputting standard JSON format semantic data and local intelligent response results into the data processing module through a TLS-encrypted bidirectional communication channel, performing data cleaning, standardization, time series alignment, and feature extraction processing to obtain target time series data; An intelligent analysis unit for inputting the target time series data into the intelligent analysis module, applying deep learning algorithms to perform energy consumption pattern recognition, device performance evaluation, abnormal behavior detection, and fault feature extraction to obtain a device health score; A decision generation unit for inputting the device health score, historical maintenance records, and professional knowledge base data into the decision generation module, performing fault diagnosis, root cause analysis, and predictive maintenance decision calculation to obtain a fault prediction result; A multi-objective optimization unit for performing multi-objective optimization algorithm calculation through the decision generation module based on the fault prediction result to obtain an energy optimization strategy.

[0040] Specifically, the cloud-based large model engine is centered around the LLaMA architecture. Through in-depth fine-tuning and vertical domain knowledge enhancement, it not only has basic language understanding and reasoning capabilities but also can perform multi-task adaptation and business intelligence expansion for the intelligent building operation and maintenance scenario. During actual deployment, a distributed multi-node cluster mode is adopted, leveraging high-performance GPUs and elastic storage to achieve high-concurrency processing capabilities for massive data. The engine is internally divided into five core modules, including a data processing module, an intelligent analysis module, a decision-making generation module, a natural language interaction module, and a model optimization module. Each module can operate independently or be called collaboratively to form an efficient closed-loop of complete data flow, analysis flow, decision-making flow, and feedback flow. The model optimization module provides continuous impetus for the online learning, self-adjustment, and version upgrade of the large model, supporting the continuous integration and evolution of new knowledge and new business scenarios. On this basis, standard JSON format semantic data and local intelligent response results are input into the data processing module through a TLS-encrypted bidirectional communication channel. This module performs multi-dimensional cleaning on all raw data, automatically eliminating duplicate data, outliers, and missing segments to ensure that each input has spatio-temporal integrity and business relevance. The data standardization mechanism automatically adapts to and unifies issues such as protocol heterogeneity, data units, data types, and namespaces, constructing a standardized data pool across protocols, manufacturers, spaces, and businesses for the entire system. The data processing module performs alignment processing on data streams with different sampling frequencies, different transmission delays, and different sources through a high-precision time series alignment algorithm, reconstructing the data trajectories of each device, measurement point, and spatial object on a unified timeline to improve the accuracy of subsequent analysis. Using deep learning tools such as autoencoders, temporal convolutional networks, and clustering analysis, the original data stream is transformed into highly distinguishable statistical features, behavior patterns, and event labels. The intelligent analysis unit, as the carrier of the core intelligence of the system, inputs the target time series data into the intelligent analysis module. With the help of various algorithms such as deep neural networks, time series modeling, clustering analysis, and contrast learning, it performs global energy consumption pattern recognition on various energy-consuming devices, key subsystems, and spatial operation states within the building, automatically distinguishing various modes such as normal operation, abnormally high energy consumption, and atypical energy use, and effectively locating energy-saving potential points. The intelligent analysis module continuously evaluates the device performance. By comparing historical and real-time data, it promptly identifies phenomena such as device performance degradation, increased maintenance requirements, or decreased operating efficiency. The abnormal behavior detection model captures early minor anomalies in a large amount of time series data, such as sudden changes in sensor data, fluctuations in device parameters exceeding thresholds, and abnormal spatial energy consumption distributions. Through in-depth fault feature extraction and similarity attribution, a comprehensive health score is generated for each device, which dynamically reflects the device's health status, failure probability, and maintenance priority. Based on the above analysis results, the decision-making generation unit realizes a data-driven intelligent decision-making closed-loop.The device health score, historical maintenance records, and professional knowledge bases (such as manufacturer manuals, maintenance cases, industry standards, etc.) are input into the decision-making generation module. This module uses knowledge reasoning and deep fusion as the engine to automatically complete fault diagnosis, root cause analysis, and predictive maintenance decision calculations. Through comprehensive analysis of the fault chains of similar devices, historical maintenance trajectories, and knowledge base rules, the system locates the root cause of the current anomaly and deduces the optimal treatment measures. At the same time, based on the remaining life, risk level, and maintenance cycle of the device, it generates predictive maintenance suggestions and work order plans with clear priorities. Based on the fault prediction and maintenance strategies, the multi-objective optimization unit improves the intelligent level of the overall building operation and energy consumption management. This unit inputs the fault prediction results and multi-dimensional constraints such as energy consumption, business requirements, comfort goals, and external environment (such as weather, energy prices, and pedestrian flow prediction) into the decision-making generation module, and calls multi-objective optimization algorithms (such as genetic algorithms, multi-objective particle swarm optimization, constraint programming, etc.) to intelligently calculate the energy optimization strategy with the lowest comprehensive energy consumption, the best comfort, the smallest risk, and the highest economy. These strategies are specific to the start-stop scheduling, parameter setting, and load distribution at the device level, and dynamically adapt to the building operation environment and business requirements, realizing the rolling optimization of energy consumption and the real-time adjustment of device operation strategies. All the above decision results, optimization strategies, and analysis reports are fed back to the operation and maintenance personnel, managers, or intelligent agents in a colloquial and knowledge-based manner through the natural language interaction module, realizing efficient human-machine collaboration and business transparency. The model optimization module regularly collects the data and execution results in all business processes, and automatically conducts self-supervised fine-tuning and online transfer learning to ensure the continuous self-evolution and improvement of the business adaptation ability of the entire large model engine during long-term operation.

[0041] In one embodiment, the deployment unit is specifically used for: Adjust the architecture and perform initialization configuration on the original LLaMA pre-trained model in the cloud large model engine to construct an initialized LLaMA basic model suitable for the building field; Perform data cleaning, structured processing, and annotation on building device manuals, historical operation and maintenance records, industry specification documents, and maintenance cases to construct a building operation and maintenance training corpus containing device parameters, fault characteristics, maintenance strategies, and control logic; Based on the building operation and maintenance training corpus, perform supervised fine-tuning and reinforcement learning training on the initialized LLaMA basic model, and selectively update the model weights using parameter-efficient fine-tuning technology to obtain a pre-fine-tuned model adapted to the building field; Integrate the building professional knowledge graph data into the pre-fine-tuned model adapted to the building field through knowledge enhancement technology, execute knowledge distillation and model compression algorithms, and at the same time perform multi-task training on the pre-fine-tuned model adapted to the building field to obtain a large model for the vertical field; Build a multi-functional large model engine based on the LLaMA architecture with deep fine-tuning, which includes a data processing module, an intelligent analysis module, a decision-making generation module, a natural language interaction module, and a model optimization module, based on large models in vertical domains.

[0042] Specifically, in the cloud large model engine, a high-performance and highly versatile original LLaMA pre-trained model is selected as the base. This model has large-scale parameters, cross-domain versatility, as well as text understanding and reasoning capabilities. Architectural adjustments and initial configurations are made to the LLaMA pre-trained model. This includes adaptively adjusting parameters such as the model input length, encoding method, inference depth, number of attention heads, and context window to make it suitable for building professional scenarios such as long documents, tables, and complex nested structures. At the same time, according to the characteristics of building operation and maintenance data, customized preprocessors and post-processing modules are configured to ensure that the model can efficiently and accurately understand and generate the required content when faced with protocol data, device parameter descriptions, work order logic, and time series event reasoning. Through the above architectural adjustments, a LLaMA base model suitable for building domain analysis, question answering, attribution, and reasoning tasks is obtained. To make the large model highly professional, practical, and capable of business implementation, a high-quality training corpus for building operation and maintenance is constructed. This involves batch collection and processing of various types and sources of documents such as building equipment manuals, control system specifications, equipment wiring diagrams, fault repair cases, work order processing records, industry technical specifications, and on-site commissioning logs. Paper and unstructured documents are converted into readable text through automated tools such as OCR, text parsing, and structured scripts, and then deep data cleaning is carried out, including removing redundancy, filtering errors, unifying naming, and standardizing terms. After that, the original text content is structurally annotated, and key information such as device types, models, parameters, typical faults, cause analysis, repair measures, and control processes is converted into standardized data entities and attribute relationships, forming a multi-dimensional, labeled training sample library containing device parameters, fault characteristics, maintenance strategies, and control logic. Based on the building operation and maintenance training corpus, supervised fine-tuning and reinforcement learning joint training are implemented on the initialized LLaMA base model. In the supervised fine-tuning phase, through "input-output" paired samples, with expert annotations as the label standard, the model is guided to quickly converge from the original general knowledge to building professional knowledge, strengthening its understanding ability of tasks such as device attributes, state attribution, fault identification, and strategy generation. In the reinforcement learning stage, an autoregressive training, self-supervised reward and punishment mechanism is introduced to optimize the model's performance in actual decision-making and intelligent interaction, promoting the model to generate solutions that more conform to the logic of operation and maintenance experts and real business scenarios. To improve training efficiency and save computing power, the entire fine-tuning process uses parameter-efficient fine-tuning techniques such as LoRA, Adapter, Prompt Tuning, etc., only updating some model weights or newly added parameter layers, and keeping the remaining weights frozen, reducing the video memory and training time overhead while ensuring performance improvement. After fine-tuning and reinforcement learning, a pre-fine-tuned model adapted to the building domain is obtained. The building professional knowledge graph data is integrated into the pre-fine-tuned model adapted to the building domain through knowledge enhancement techniques.The knowledge graph is composed of entity nodes (such as devices, spaces, faults, events, parameters) and relationship edges (such as subordination, association, causality, control), supplementing the structured logic and rules that are difficult to capture in text training. Through knowledge enhancement techniques, such as knowledge injection, entity embedding, knowledge retrieval-enhanced reasoning, etc., the knowledge graph is integrated with the pre-fine-tuning model. During the training phase, knowledge triples are supplemented as input to enable the model to understand the complex logic and business links between devices; during the inference phase, the knowledge graph is called for auxiliary retrieval and inference correction. To ensure that the model can operate efficiently with limited resources, combined with knowledge distillation and model compression algorithms, the knowledge of the large model is transferred to the lightweight model, reducing the inference latency and deployment cost. At the same time, through the multi-task training mechanism, various tasks such as device attribute recognition, fault diagnosis, policy optimization, text generation, and natural language interaction are mixed for training, enhancing the model's multi-scenario adaptability, inference depth, and human-computer dialogue ability, forming a vertical domain large model for the exclusive building intelligent operation and maintenance scenario. Based on the vertical domain large model, a multi-functional large model engine for the production environment is constructed. The engine internally integrates a data processing module, an intelligent analysis module, a decision-making generation module, a natural language interaction module, and a model optimization module, realizing the full-process closed-loop capabilities from underlying data collection, time-series feature extraction, device health assessment, energy consumption anomaly analysis, predictive maintenance decision-making, fault root cause tracing, energy strategy optimization, to natural language report generation, expert knowledge Q&A, and model dynamic evolution. The data processing module is responsible for the efficient processing of raw semantic data and local responses, ensuring data cleaning, standardization, feature engineering, and time-series alignment; the intelligent analysis module undertakes tasks such as device status recognition, pattern attribution, risk warning, and operation trend analysis, providing accurate input for subsequent decision-making; the decision-making generation module automatically completes fault location, root cause tracing, policy issuance, and work order generation based on the analysis results, knowledge graph, and historical cases, realizing proactive maintenance and global optimization; the natural language interaction module supports the operation and maintenance personnel to interact with the system in a natural language manner, obtaining multi-scenario services such as intelligent suggestions, data analysis, event tracing, and policy explanations; the model optimization module regularly schedules all business data, operation logs, and interaction feedback, automatically performing self-supervised fine-tuning and online reinforcement learning, realizing model self-evolution and continuous enhancement of business adaptation capabilities.

[0043] In one embodiment, the intelligent building remote operation and maintenance management and control system based on the large model and the cloud-edge collaborative architecture further includes: An execution subsystem is used to set warning thresholds and alarm thresholds according to the device health score to obtain device health management trigger conditions; calculate the remaining service life of the device based on the fault prediction results and historical fault data to generate predictive maintenance work orders; generate a rolling optimization control strategy according to the energy optimization strategy in combination with weather forecast data, energy prices, predicted pedestrian flows, and historical usage patterns, and convert the rolling optimization control strategy into dynamic control instructions; analyze the device access patterns and network traffic characteristics in the building to obtain security protection strategies; divide the predictive maintenance work orders, dynamic control instructions, and security protection strategies into three working modes: automatic execution mode, recommended approval mode, and manual intervention mode according to the priority to obtain a classified execution strategy set; send the dynamic control instructions in the classified execution strategy set to the corresponding edge gateway for execution through a TLS encrypted two-way communication channel, and convert the control instructions into the corresponding device protocol format through the edge gateway to obtain the execution results of the terminal devices.

[0044] Specifically, based on the health score results generated by the cloud-based large model intelligent analysis module for all key devices, the execution subsystem automatically sets multi-level warning thresholds and alarm thresholds for different types of devices. Parametrize the threshold logic to form a device health management trigger condition table, enabling real-time automatic discrimination of each update of the health score and ensuring a sensitive response to early risks of the device. When the score approaches the warning value, the system immediately pushes maintenance suggestions or generates a risk warning report. If the score drops below the alarm threshold, the emergency response and fault troubleshooting process will be immediately initiated. Based on the device health score and its changing trend, the execution subsystem combines the cloud-based large model fault prediction results and historical fault samples, and uses the remaining useful life calculation model to accurately predict the actual life window of each device. The remaining useful life calculation model integrates multi-dimensional data such as health score, device operating environment, usage intensity, historical maintenance, and fault statistics of similar devices, and dynamically adjusts parameters. Whenever the RUL of a device approaches the maintenance limit or shows a sudden decline, the system automatically generates a predictive maintenance work order. The work order content includes device identification, predicted fault type, priority, recommended maintenance time window, required spare parts and tools, maintenance responsible person, etc. At the same time, it intelligently schedules the optimal time to avoid peak hours and critical business windows, effectively improving the pertinence and planning of maintenance and reducing the downtime loss caused by sudden failures. Intelligently and dynamically optimize energy management. By integrating the energy optimization strategy issued by the cloud-based large model and the real-time collected data such as weather forecast, energy price, predicted pedestrian flow, and building historical energy consumption, the execution subsystem adopts a sliding window and multi-objective adaptive optimization algorithm to generate a rolling optimization control strategy covering systems such as HVAC, lighting, elevators, and power in real time. All control strategies are automatically converted into dynamic control instructions after generation, including underlying operation commands such as specific device start / stop, parameter setting, operation time period switching, and load distribution, ensuring the efficient implementation of the energy strategy and minimizing the operating cost. The execution subsystem conducts real-time security situation monitoring. By continuously analyzing the device access patterns and network traffic characteristics within the building, it identifies potential threats such as abnormal device access, abnormal instruction issuance, protocol violations, and abnormal data packet traffic. The system uses technical means such as anomaly detection, rule engines, behavior profiling, and threat intelligence libraries to dynamically generate security protection strategies, including measures such as abnormal access blocking, enhanced device authentication, control command whitelisting, traffic rate limiting, abnormal log recording, and alarm linkage. When the system faces the simultaneous issuance of multiple tasks such as maintenance work orders, optimization control, and security protection, the execution subsystem determines the priority of all tasks in multiple dimensions such as the scope of influence, emergency level, and business continuity requirements. The tasks are divided into three modes: automatic execution mode, recommended approval mode, and manual intervention mode.Among them, the automatic execution mode is applicable to maintenance and optimization instructions with low risk, clear boundaries, and safe and reversible operations (such as air conditioner energy-saving mode switching, automatic restart of non-high-risk equipment); the recommended approval mode is for important maintenance decisions or those with a large impact, such as elevator maintenance and significant adjustments to energy consumption strategies, which require online confirmation or approval by professional personnel before being issued; the manual intervention mode is specifically used for scenarios where direct intervention by operation and maintenance engineers is necessary, such as emergency safety incidents, core equipment failures, and major system changes. The system automatically generates a classified execution policy set, setting the trigger conditions, operation procedures, approval links, and fallback mechanisms for each task. After the classified policy integration is completed, all dynamic control instructions, work order tasks, and security policies are uniformly sent to the corresponding edge gateways through a TLS-encrypted two-way communication channel. TLS ensures the confidentiality, integrity, and identity authentication during the instruction transmission process, preventing man-in-the-middle attacks, instruction tampering, and illegal access. After receiving the classified execution policy set, the edge gateway combines its own protocol adaptation engine and protocol conversion module to automatically parse the upper-layer instructions into the protocol formats supported by the target terminal devices (such as Modbus, BACnet, OPC, custom APIs, etc.), ensuring that each device can accurately, efficiently, and safely respond to the control commands. The execution feedback, status changes, and abnormal responses of the devices are returned to the edge gateway in real time and synchronized to the cloud, realizing the full-process closed-loop tracking of business execution. Operation and maintenance personnel and the large model intelligent engine make subsequent policy adjustments, model retraining, and management reviews based on these execution results, forming a self-learning, self-adaptive, and self-optimizing operation and maintenance decision-making mechanism.

[0045] Among them, in the process of implementing the intelligent operation and maintenance control process based on device health scoring, fault prediction, and energy optimization strategies, it includes implementing fixed-time resource allocation optimization control, including: performing multi-agent system modeling on the edge gateways in the cloud-edge-end three-layer collaborative computing architecture, regarding each edge gateway as an agent node with disturbed dynamic characteristics, and obtaining a multi-input-multi-output nonlinear uncertain multi-agent network topology; modeling the global equality constraint and local inequality constraint for the resource allocation in each area of the building, taking device operating parameters, energy quotas, maintenance resources, etc. as decision variables, and obtaining a set of resource allocation constraint conditions; designing a fixed-time high-order extended state observer based on the sign function theory to estimate the uncertain dynamic parameters and external disturbances in the multi-agent system in real time, and obtaining the fixed-time convergence estimation values of the system state and disturbances; constructing a time-switching controller through the output feedback backstepping design process, deploying the controller in the edge gateway, and obtaining a resource optimization control unit with two-stage decision-making capabilities; processing the global equality constraint in the set of resource allocation constraint conditions to ensure the balance of energy and computing resource allocation among agents, and obtaining an initial resource allocation plan that satisfies the equality constraint; extracting the gradient information of the inequality constraint in the set of resource allocation constraint conditions based on the ε-exact penalty function, and optimizing and adjusting the initial resource allocation plan in combination with the gradient information to obtain a resource allocation optimization plan that simultaneously satisfies the equality constraint and the inequality constraint; verifying the resource allocation optimization plan using the Lyapunov stability criterion to ensure that all signals remain actually fixed-time stable, and the error between the resource allocation output of all edge gateways and the theoretical optimal solution is kept within the set threshold range, and obtaining a final resource allocation plan with stability guarantee; converting the final resource allocation plan into device-level dynamic control instructions and a predictive maintenance work order priority sequence, and sending the dynamic control instructions to the corresponding edge gateways for execution through a TLS encrypted two-way communication channel, and obtaining the actual execution results of the resource optimized allocation.

[0046] In the embodiments of the present invention, a TLS-encrypted two-way communication channel configured through reverse proxy and persistent connection solves the problems of high loss rate of BACnet / IP protocol instructions in a multi-NAT environment and high communication failure rate of Modbus-TCP protocol under strict firewall policy restrictions in traditional VPN or port mapping, ensuring the confidentiality and integrity of data transmission and effectively resisting various network threats including protocol layer attacks. The five-layer structure unified semantic model constructed based on the concept of point cloud digital twin solves the technical problem that it is difficult for devices from different manufacturers in traditional smart buildings to interoperate due to different communication protocols and private interfaces, realizing seamless fusion and interoperability of heterogeneous system data. The proximal decision-making engine of the edge gateway implements a four-level response mechanism, solving the problem that in abnormal situations such as network interruption or cloud platform failure in the existing system, relying on the cloud for decision-making leads to slow response of local critical services, ensuring the continuous and stable operation of critical services such as emergency dispatching of elevators. The cloud-edge-end three-layer collaborative computing architecture follows a closed-loop process of "terminal collection-edge processing-cloud analysis-edge execution-terminal response", breaking the information barrier between heterogeneous systems and solving the problem that traditional building systems form "data islands" due to different protocols and data formats and cannot achieve global optimization and linkage control. The cloud-based large model engine has the ability to deeply understand the semantics of multi-source heterogeneous data by integrating the building professional knowledge base, and can associate with the building operation and maintenance knowledge base, solving the technical problem that traditional analysis methods are difficult to deeply mine the complex associations and trends hidden in building operation data and cannot provide accurate fault prediction and intelligent operation and maintenance decision support. Through standardized access, automated configuration and remote maintenance capabilities, the system deployment complexity and operation and maintenance costs are significantly reduced, and the workload of network configuration, protocol docking and system debugging is greatly reduced from a relatively high proportion of the total project man-hours, reducing the total cost of ownership of the system.

[0047] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the foregoing system embodiments and will not be elaborated herein.

[0048] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the system described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0049] As described above, the above embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of various embodiments of the present invention.

Claims

1. A smart building remote operation and maintenance management and control system based on a large model and a cloud-edge collaborative architecture, characterized in that, Including: A configuration subsystem for performing reverse proxy and persistent connection configuration on edge gateway devices in a smart building to obtain a TLS-encrypted two-way communication channel; A construction subsystem for constructing a cloud-edge-end three-layer collaborative computing architecture based on the TLS-encrypted two-way communication channel, where the cloud-edge-end three-layer collaborative computing architecture includes a cloud large model engine, an edge gateway, and a terminal device layer; A conversion subsystem for uniformly semantically abstracting and converting heterogeneous protocol data through the edge gateway to obtain standard JSON format semantic data; A rule processing subsystem for inputting the standard JSON format semantic data into the proximal decision engine of the edge gateway for rule processing to obtain a local intelligent response result; A multi-dimensional analysis subsystem for transmitting the standard JSON format semantic data and the local intelligent response result to the cloud large model engine through the TLS-encrypted two-way communication channel for multi-dimensional analysis to obtain device health scores, fault prediction results, and energy optimization strategies.

2. The intelligent building remote operation and maintenance control system based on the large model and cloud-edge collaborative architecture according to claim 1, wherein The configuration subsystem is specifically used for: Deploying a network proxy server cluster in the cloud to obtain an NPS server cluster, and installing a network proxy client NPC in edge gateway devices in the smart building; Performing identity authentication on TCP / TLS long connection requests initiated by edge gateway devices through the NPS server cluster to obtain an active connection registry; Establishing a reverse proxy channel for edge gateway devices based on the active connection registry to obtain a NAT penetration communication link, and performing TLS 1.2 encryption processing on the NAT penetration communication link to obtain a TLS-encrypted two-way communication channel.

3. The intelligent building remote operation and maintenance management and control system based on the large model and cloud-edge collaborative architecture according to claim 2, characterized in that, The construction subsystem is specifically used for: Deploying and configuring cloud infrastructure to obtain a cloud large model engine; Deploying a building-specific protocol adaptation engine, a protocol abstraction and semantic conversion module, a data preprocessing and caching module, an edge security gateway, and a proximal decision engine in the edge gateway; Connecting heating, ventilation, and air conditioning equipment, intelligent lighting equipment, elevator control equipment, video surveillance equipment, intrusion alarm equipment, access control management equipment, and energy metering equipment to the edge gateway to obtain a terminal device layer; Configuring the data closed-loop process of the cloud large model engine, the edge gateway, and the terminal device layer through the TLS-encrypted two-way communication channel to construct a cloud-edge-end three-layer collaborative computing architecture.

4. The intelligent building remote operation and maintenance management and control system based on the large model and the cloud-edge collaborative architecture according to claim 1, characterized in that, The conversion subsystem further includes: A docking unit for docking the protocol abstraction and semantic conversion module in the edge gateway with the protocol adaptation engine to obtain a receiving channel for raw protocol data; A layering unit for performing hierarchical design on the unified semantic model based on the concept of point cloud digital twin to construct a five-layer structure model including a space model layer, a device model layer, a measurement point model layer, a service model layer, and an association model layer; An extraction unit for identifying the data format and extracting valid values from heterogeneous protocol data of BACnet / IP and Modbus-TCP parsed by the protocol adaptation engine through the receiving channel to obtain device raw measurement values and control parameters; A mapping unit for performing standardized mapping on the original device measurement values and control parameters based on the five-layer structure model to obtain device data with unified semantics; An encapsulation unit for encapsulating the device data with unified semantics in JSON format to obtain standard JSON format semantic data.

5. The intelligent building remote operation and maintenance management and control system based on the large model and cloud-edge collaborative architecture according to claim 4, characterized in that, The mapping unit is specifically used for: Defining the building physical space structure for the space model layer, and constructing a set of standardized space identifiers including buildings, floors, areas, rooms and their hierarchical relationships; Defining the description of device types, attributes, functions and states for the device model layer, and constructing a set of standardized device identifiers including HVAC devices, lighting devices, elevator systems, security devices, and energy metering devices; Defining standard identifiers, data types, units, value ranges, and alarm thresholds for the measurement point model layer to obtain a standardized measurement point description library; Based on the standardized measurement point description library, constructing a mapping table for object identifiers and attribute identifiers in the BACnet / IP protocol and register addresses and standard measurement point identifiers in the Modbus-TCP protocol to obtain a conversion dictionary; Replacing the protocol-specific identifiers of the original device measurement values according to the conversion dictionary and performing data type conversion to obtain device parameters with unified identifiers and data types; Performing unit standardization processing and validity verification on the device parameters with unified identifiers and data types, and associating them with the set of standardized space identifiers and the set of standardized device identifiers to obtain device data with unified semantics.

6. The intelligent building remote operation and maintenance management and control system based on the large model and cloud-edge collaborative architecture according to claim 1, wherein, The rule processing subsystem is specifically used for: Deploying a local status maintenance module, a preset rule execution module, a lightweight inference module and a breakpoint resumption module in the proximal decision engine of the edge gateway; Inputting the standard JSON format semantic data into the local status maintenance module for status update processing to obtain a real-time device status database; Performing Boolean expression condition matching on the parameter values in the real-time device status database to obtain a sequence of operation instructions to be executed; Inputting the abnormal parameters in the real-time device status database into the lightweight inference module for knowledge distillation neural network calculation, performing local anomaly detection and optimization adjustment based on the parameters issued by the cloud to obtain intelligent optimization control instructions; Performing priority sorting on the operation instruction sequence and the intelligent optimization control instructions according to the event type to obtain an execution instruction set with priority sorting; Tracking the execution status of the execution instruction set with priority sorting, and locally caching important data during the execution process through the breakpoint resumption module to ensure the operation of the local decision-making mode during network interruption, and synchronizing with the cloud in priority order after the network is restored to obtain a local intelligent response result.

7. The intelligent building remote operation and maintenance management and control system based on the large model and cloud-edge collaborative architecture according to claim 1, characterized in that, The multi-dimensional analysis subsystem further includes: A deployment unit for deploying a multi-functional large model engine based on the depth-fine-tuned LLaMA architecture including a data processing module, an intelligent analysis module, a decision generation module, a natural language interaction module and a model optimization module in the cloud large model engine; A data processing unit, configured to input the standard JSON format semantic data and the local intelligent response result into the data processing module through the TLS encrypted two-way communication channel, and perform data cleaning, standardization, time series alignment, and feature extraction processing to obtain target time series data; An intelligent analysis unit, configured to input the target time series data into the intelligent analysis module, and apply deep learning algorithms to perform energy consumption pattern recognition, device performance evaluation, abnormal behavior detection, and fault feature extraction to obtain a device health score; A decision-making generation unit, configured to input the device health score, historical maintenance records, and professional knowledge base data into the decision-making generation module, and perform fault diagnosis, root cause analysis, and predictive maintenance decision calculation to obtain a fault prediction result; A multi-objective optimization unit, configured to perform multi-objective optimization algorithm calculation through the decision-making generation module based on the fault prediction result to obtain an energy optimization strategy.

8. The intelligent building remote operation and maintenance management and control system based on the large model and cloud-edge collaborative architecture according to claim 7, characterized in that, The deployment unit is specifically configured to: Adjust the architecture and perform initialization configuration on the original LLaMA pre-trained model in the cloud large model engine to construct an initialized LLaMA base model suitable for the building field; Perform data cleaning, structured processing, and annotation on building equipment manuals, historical operation and maintenance records, industry specification documents, and maintenance cases to construct a building operation and maintenance training corpus containing equipment parameters, fault features, maintenance strategies, and control logic; Perform supervised fine-tuning and reinforcement learning training on the initialized LLaMA base model based on the building operation and maintenance training corpus, and selectively update the model weights using parameter-efficient fine-tuning technology to obtain a pre-fine-tuned model adapted to the building field; Integrate building professional knowledge graph data into the pre-fine-tuned model adapted to the building field through knowledge enhancement technology, perform knowledge distillation and model compression algorithms, and simultaneously perform multi-task training on the pre-fine-tuned model adapted to the building field to obtain a vertical domain large model; Construct a multi-functional large model engine based on the deep fine-tuned LLaMA architecture, including a data processing module, an intelligent analysis module, a decision-making generation module, a natural language interaction module, and a model optimization module, based on the vertical domain large model.

9. The intelligent building remote operation and maintenance management and control system based on the large model and cloud-edge collaborative architecture according to claim 1, characterized in that, The intelligent building remote operation and maintenance management and control system based on the large model and cloud-edge collaborative architecture further includes: An execution subsystem, configured to set warning thresholds and alarm thresholds according to the device health score to obtain device health management trigger conditions; calculate the remaining service life of the device based on the fault prediction result and historical fault data to generate a predictive maintenance work order; generate a rolling optimization control strategy according to the energy optimization strategy in combination with weather forecast data, energy prices, predicted pedestrian flow and historical usage patterns, and convert the rolling optimization control strategy into dynamic control instructions; analyze the device access mode and network traffic characteristics in the building to obtain a security protection strategy; divide the predictive maintenance work order, dynamic control instructions and security protection strategy into three working modes of automatic execution mode, recommended approval mode and manual intervention mode according to the priority to obtain a classified execution strategy set; send the dynamic control instructions in the classified execution strategy set to the corresponding edge gateway for execution through the TLS encrypted bidirectional communication channel, and convert the control instructions into the corresponding device protocol format through the edge gateway to obtain the execution results of the terminal device.

Images