Cloud side-end collaborative intelligent building management and control system and method fused with large language model

Abstract

The invention relates to the technical field of large language models, and discloses a cloud side-end collaborative intelligent building management and control system and method fused with a large language model. The method comprises the following steps: an edge gateway initiates reverse proxy connection to a cloud end and collects original protocol data of each protocol device in a building; performing protocol analysis on the original protocol data to obtain semantic data; receiving semantic data and fusing the semantic data with a vector knowledge base and weather electricity price external data to generate a predictive maintenance work order and an energy consumption optimization strategy; and issuing the energy consumption optimization strategy and the lightweight decision model to an edge gateway, executing local prediction and equipment control, obtaining execution feedback data and reporting the execution feedback data to the cloud. According to the invention, by constructing a zero-configuration security network penetration mechanism based on reverse proxy, the NAT natural penetration capability of active external connection of intranet equipment is realized, the protocol compatibility problem caused by firewall configuration in the prior art is solved, and the intelligent analysis level of an intelligent building management and control system is improved.

Description

Technical Field

[0001] This invention relates to the field of large language model technology, and in particular to a cloud-edge-device collaborative smart building management system and method that integrates large language models. Background Technology

[0002] With the rapid development of smart building construction, the number of intelligent devices such as various automatic control systems, security systems, and energy management systems deployed within buildings is growing exponentially. However, existing control solutions face numerous technical bottlenecks in practical applications. Traditional remote access solutions are cumbersome and inefficient in complex network environments, and commonly used building protocols have poor communication reliability in multi-layer NAT environments, resulting in severely insufficient real-time performance of remote monitoring and control. Smart buildings contain many heterogeneous subsystems from different vendors, using various communication protocols and proprietary interfaces such as BACnet, Modbus, and OPC UA, with varying data formats, forming data silos and hindering inter-system interconnection and global optimization. While existing systems can collect large amounts of data, they mostly remain at the level of basic monitoring and simple rule-based alarms, leading to slow responses from local critical business operations. They lack effective integration and intelligent application of knowledge in the building operations and maintenance field, making it difficult for maintenance personnel to quickly acquire tacit knowledge, resulting in high deployment and maintenance costs. Summary of the Invention

[0003] The main objective of this invention is to provide a cloud-edge-device collaborative smart building management system and method that integrates a large language model. This invention achieves the natural NAT traversal capability of internal network devices actively connecting to external networks by constructing a zero-configuration secure network traversal mechanism based on a reverse proxy. It solves the protocol compatibility problem caused by firewall configuration in the prior art and improves the intelligent analysis level of the smart building management system.

[0004] To achieve the above objectives, this invention provides a cloud-edge-device collaborative smart building management method that integrates a large language model, comprising the following steps: S1: The edge gateway initiates a reverse proxy connection to the cloud and collects the raw protocol data of each protocol device in the building; S2: The protocol adaptation engine in the edge gateway performs protocol parsing on the original protocol data, obtains semantic data, and uploads it to the cloud through the reverse proxy connection; S3: The cloud-based big model engine receives the semantic data and integrates it with the vector knowledge base of the storage device manual and operation and maintenance procedures, as well as external data such as weather and electricity prices, to generate predictive maintenance work orders, energy consumption optimization strategies, and fault diagnosis conclusions. S4: The cloud sends the energy consumption optimization strategy and lightweight decision model to the edge gateway through the reverse proxy connection, performs local prediction and device control, obtains execution feedback data and reports it to the cloud.

[0005] Optionally, in a first implementation of the first aspect of the present invention, step S1 includes: S11: The NPC client in the edge gateway initiates a TCP long connection to the NPS server in the cloud and performs a TLS 1.2 encrypted handshake. After completing two-way authentication with the NPS server through the X.509 digital certificate, the NPS server registers the edge gateway to the connected NPC registry and maintains heartbeat keep-alive, thus obtaining a verified reverse proxy connection. S12: The protocol adaptation engine in the edge gateway calls the BACnet / IP, Modbus TCP, and OPC UA protocol drivers through the reverse proxy connection to send data reading instructions to each protocol device in the building and receive device response messages to obtain the original protocol data.

[0006] Optionally, in a second implementation of the first aspect of the present invention, step S2 includes: S21: The protocol adaptation engine in the edge gateway calls the corresponding protocol parser according to the protocol type of the original protocol data, extracts the object identifier and Present_Value attribute value from the BACnet / IP protocol data, extracts the physical quantity value from the ModbusTCP protocol data according to the register address mapping table, and reads the DataValue corresponding to the NodeId from the OPC UA protocol data to obtain the protocol parsing result; S22: Based on the spatial hierarchy, device classification system, and standard measurement point definition in the unified semantic model, the protocol parsing result is mapped into semantic data containing device identifier, spatial location, measurement point type, value, unit, and timestamp, and then uploaded to the cloud via the reverse proxy connection after attaching the HMAC-SHA256 message authentication code.

[0007] Optionally, in a third implementation of the first aspect of the present invention, step S3 includes: S31: The cloud-based big model engine performs semantic understanding on the semantic data and retrieves semantic vectors from the vector knowledge base of the storage device manual and operation and maintenance procedures, and integrates them with external data such as weather and electricity prices to obtain comprehensive operation data; S32: Based on the comprehensive operation data, extract the physical connection relationship, control logic relationship, spatial topology relationship and functional dependency relationship between the protocol devices, and construct a device semantic association network containing device nodes and association relationship edges; S33: The large model engine identifies faults based on the device semantic association network and generates predictive maintenance work orders, energy consumption optimization strategies, and fault diagnosis conclusions.

[0008] Optionally, in a fourth implementation of the first aspect of the present invention, step S33 includes: The large model engine extracts features from the operating parameters in the device semantic association network and retrieves historical fault cases from the vector knowledge base for comparison, thereby obtaining device health status assessment results and fault risk prediction results. Based on the equipment health status assessment results, a predictive maintenance work order is generated, which includes the remaining life prediction value, fault description, recommended maintenance time, and required spare parts list. An energy consumption optimization strategy is generated by combining the external data of weather and electricity prices. Based on the semantic association network of the devices, the alarm devices are analyzed to obtain fault diagnosis conclusions.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, step S4 includes: S41: The cloud sends the energy consumption optimization strategy and lightweight decision model to the edge gateway through the reverse proxy connection. The edge intelligent decision engine in the edge gateway verifies and loads the lightweight decision model. S42: The edge intelligent decision engine runs the lightweight decision model to perform local prediction and generate device control parameters, converts the control instructions of the energy consumption optimization strategy into protocol messages corresponding to the target device and sends them for execution control, obtains the execution feedback data of the target device and reports it to the cloud through the reverse proxy connection.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, step S41 includes: S411: The cloud performs knowledge distillation, model parameter quantification, and model structure pruning on the initial decision model in the large model engine to obtain a lightweight decision model. The control instructions of the energy consumption optimization strategy and the lightweight decision model are encapsulated by cloud private key signing and sent to the edge gateway through the reverse proxy connection. S412: The edge intelligent decision engine in the edge gateway uses a pre-stored cloud public key to verify the digital signature to confirm that the data source is trustworthy and has not been tampered with, extracts the lightweight decision model and loads it into the local inference engine.

[0011] Optionally, in a seventh implementation of the first aspect of the present invention, it further includes: S5: The large model engine performs intent recognition on the natural language query requests input by the operation and maintenance personnel to extract the query intent, and converts the query intent into a natural language query response; S6: The cloud periodically collects the execution feedback data, fault cases, and operation and maintenance feedback, labels and cleans them, constructs a training dataset, and inputs the training dataset into the large model engine for supervised learning to update the model parameters and complete incremental training.

[0012] Optionally, in an eighth implementation of the first aspect of the present invention, step S6 includes: The cloud periodically collects the execution feedback data, fault cases and operation and maintenance feedback, analyzes the causes of cases with deviations and marks them with fault tags, removes abnormal data and completes the cleaning process to build a training dataset; The training dataset is input into the large model engine, and the loss function between the predicted output and the true label is calculated through supervised learning. The backpropagation gradient descent algorithm is then executed to update the model parameters, thus completing incremental training.

[0013] The present invention also provides a cloud-edge-device collaborative smart building management system that integrates a large language model, used to execute the steps of the above-mentioned cloud-edge-device collaborative smart building management method that integrates a large language model.

[0014] In summary, this invention achieves natural NAT traversal capability for internal network devices to proactively connect to external networks by constructing a zero-configuration secure network traversal mechanism based on a reverse proxy, thus solving the protocol compatibility issues caused by firewall configuration in existing technologies. Through the synergistic effect of a protocol adaptation engine and a unified semantic model, it converts various protocol data from heterogeneous underlying devices into standardized semantic information, breaking down information barriers between different building subsystems and achieving integration and data fusion. Furthermore, by deeply integrating a large-scale language model fine-tuned with knowledge from the building operations and maintenance domain into the cloud-based intelligent analysis layer, it performs deep understanding of semantic data and combines it with equipment manuals, operation and maintenance procedures, and external weather and electricity price data from a vector knowledge base for multimodal fusion. A semantic network of device associations, encompassing physical connections, control logic, spatial topology, and functional dependencies, was constructed. Through knowledge distillation, quantization, and pruning techniques, the large cloud-based model was compressed into a lightweight decision-making model and distributed to the edge gateway, enabling local intelligent decision-making at the edge. This ensures reliable operation of core building functions and rapid emergency response under extreme conditions such as network outages. A natural language interface lowers the barrier to entry for the system, and incremental training based on execution feedback data enables continuous learning and co-evolution of the model, improving the network access efficiency, data fusion capabilities, intelligent analysis level, emergency response speed, human-computer interaction experience, and long-term operational value of the smart building management system. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the steps of a cloud-edge-device collaborative smart building management method that integrates a large language model in one embodiment of the present invention; Figure 2 This is a structural block diagram of a cloud-edge-device collaborative smart building management system that integrates a large language model, according to one embodiment of the present invention.

[0016] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Reference Figure 1 This embodiment provides a cloud-edge-device collaborative smart building management method that integrates a large language model, including the following steps: S1: The edge gateway initiates a reverse proxy connection to the cloud and collects the raw protocol data of each protocol device in the building; S2: The protocol adaptation engine in the edge gateway performs protocol parsing on the original protocol data, obtains semantic data, and uploads it to the cloud through a reverse proxy connection. S3: The cloud-based big model engine receives semantic data and integrates it with the vector knowledge base of storage device manuals and operation and maintenance procedures, as well as external data such as weather and electricity prices, to generate predictive maintenance work orders, energy consumption optimization strategies, and fault diagnosis conclusions. S4: The cloud sends energy optimization strategies and lightweight decision models to the edge gateway through a reverse proxy connection, performs local prediction and device control, obtains execution feedback data, and reports it to the cloud.

[0019] In one example, step S1 includes: S11: The NPC client in the edge gateway initiates a TCP long connection to the NPS server in the cloud and performs a TLS 1.2 encrypted handshake. After completing two-way authentication with the NPS server through the X.509 digital certificate, the NPS server registers the edge gateway to the connected NPC registry and maintains heartbeat keep-alive, thus obtaining a verified reverse proxy connection. S12: The protocol adaptation engine within the edge gateway calls the BACnet / IP, Modbus TCP, and OPC UA protocol drivers through a reverse proxy connection to send data read commands to various protocol devices within the building and receive device response messages to obtain the original protocol data.

[0020] In this example, after the edge gateway powers on, its embedded NPC client module immediately initializes the network communication stack and proactively initiates a TCP long connection request to the pre-configured cloud NPS server address. This connection, as the foundation of the reverse proxy data channel, requires a TLS 1.2 encrypted handshake process during its establishment to ensure the security of data encryption and authentication within the communication channel. During the TLS handshake phase, the NPC client retrieves the device's digital certificate in X.509 format from its protected local storage and ensures that the private key cannot be exported or forged through a hardware security module or secure boot binding mechanism. The NPS server also presents its server certificate for peer verification. Both parties establish a secure session context through certificate chain verification and key negotiation mechanisms, completing the encrypted channel initialization process, including certificate validity verification, two-way authentication, and symmetric key generation. After the handshake is complete, the NPS server writes the edge gateway's unique identifier and connection status into the connected NPC registry and assigns a session identifier. The registry information includes the edge gateway's device fingerprint, access time, connection status, IP address, and session token. To maintain the validity of long-term connections and achieve a highly reliable transmission mechanism, the NPS server will periodically send heartbeat packets to the NPC client. The NPC client must respond to the heartbeat request within a specified time window to maintain connection activity. If a heartbeat is lost or the response times out, a connection reconstruction mechanism is triggered, thereby ensuring that the reverse proxy channel is always available under complex network conditions, forming a persistent reverse proxy connection channel that is authenticated, fully encrypted, and supports NAT traversal. After the authentication channel is established, the edge gateway enters the protocol device data acquisition phase. At this time, the protocol adaptation engine integrated within the edge gateway initiates a heterogeneous protocol data polling process. It retrieves the list of currently managed devices and their corresponding communication parameters from the access configuration table, and schedules protocol drivers such as BACnet / IP, Modbus TCP, and OPC UA according to the communication protocol type adapted by each device. The driver constructs a data read request in the corresponding format according to the standard protocol stack specification. For example, for BACnet / IP devices, it constructs a ReadProperty request to read the Present_Value attribute of the Analog Value object; for Modbus TCP devices, it constructs a message with function code 03 to read the holding register; and for OPC UA devices, it constructs a node browsing tree and accesses the DataValue field of the specified NodeId. The constructed data read command is sent to the target device via Ethernet or serial interface. At the same time, the protocol driver configures the response timeout and retry mechanism to deal with abnormal responses or link fluctuations on the device side.After receiving the read command, the target device generates a response data packet according to its respective protocol stack and returns it to the edge gateway. The protocol adaptation engine receives and initially parses the data content one by one. For different protocol types, it uses the parsing logic in its respective protocol driver to extract the original protocol data fields such as measurement point address, value field, status word, and timestamp. At the same time, it adds a unified identifier and source address mark and stores them in the data preprocessing module of the edge gateway.

[0021] In one example, step S2 includes: S21: The protocol adaptation engine in the edge gateway calls the corresponding protocol parser according to the protocol type of the original protocol data, extracts the object identifier and Present_Value attribute value for BACnet / IP protocol data, extracts the physical quantity value for Modbus TCP protocol data according to the register address mapping table, and reads the DataValue corresponding to NodeId for OPC UA protocol data to obtain the protocol parsing result; S22: Based on the spatial hierarchy, equipment classification system, and standard measurement point definition in the unified semantic model, the protocol parsing results are mapped into semantic data containing equipment identifier, spatial location, measurement point type, value, unit, and timestamp. After attaching the HMAC-SHA256 message authentication code, the data is uploaded to the cloud via a reverse proxy connection.

[0022] In this example, the protocol adaptation engine within the edge gateway determines the protocol type of each data entry based on the protocol identifier information carried in the data, and automatically calls the pre-built protocol parser module to perform protocol field decoding and structure reconstruction accordingly. When receiving BACnet / IP protocol data, the parser deconstructs the NPDU and APDU message portions into an object access request structure. Based on the object identifier and attribute identifier (such as Present_Value), it locates the specific measurement point and extracts information such as the real-time value, object type, and status code of that measurement point. Simultaneously, it records the device instance number and message timestamp of the data source device. When receiving Modbus TCP protocol data, the protocol parser identifies the function code field. If it is a read holding register (function code 0x03) or read input register (function code 0x04), it extracts the corresponding register address and register value from the data frame. Using the register address mapping table configured within the edge gateway, it converts the register location into a measurement point name and unit with physical semantics. For example, it parses the value of register 40010 as "water supply temperature 23.5 degrees Celsius." For OPC UA protocol data, the parser calls OPC through the established session channel. The UA service interface uses NodeId as a key index to read the current DataValue field content of the node, obtaining its corresponding measurement point name, measurement value, data quality status (StatusCode), and timestamp information. The results parsed from different protocol types are standardized into a unified intermediate structure and cached in the parsing result buffer. This structure contains fields such as parsing success flag, original device identifier, protocol source, measurement point code, physical quantity value, unit, and original timestamp. The semantic mapping module configured inside the edge gateway is based on a unified semantic model. The model includes the building space hierarchy (e.g., “Building-A / Floor-3 / Area-North / Room-301”), equipment classification system (e.g., “Air Conditioning Unit / Chiller Pump / Air Valve”), and standard measurement point definitions (e.g., “Temperature / Power / Status”) and unit specifications (e.g., “Celsius / kWh / ON / OFF”). Based on the original device identifier in the protocol parsing result, it looks up the corresponding logical device identifier (deviceId) in the unified semantic model. At the same time, it locates the physical spatial location information (spaceId) of the device by combining the device's installation configuration table. Then, it maps the parsed measurement point code to a standardized measurement point type (pointType) and measurement point name (pointName), and retains its measurement value (value), unit (unit), and collection timestamp (timestamp). The above fields are assembled into a semantic data object, which is in structured JSON format.After assembling the data, the semantic conversion module invokes the edge gateway's built-in security module to perform HMAC-SHA256 digest calculation on the structured data. It then generates a corresponding Message Authentication Code (MAC) using the device's private key and appends this MAC to the end of the data object as an integrity protection field. The semantically encoded data with the authentication digest is then uploaded to the cloud-based unified operations and maintenance platform's data access layer via an established TLS encrypted reverse proxy connection.

[0023] In one example, step S3 includes: S31: The cloud-based large model engine performs semantic understanding on semantic data and retrieves semantic vectors from the vector knowledge base of storage device manuals and operation and maintenance procedures. It then integrates these semantic vectors with external data such as weather and electricity prices to obtain comprehensive operation data. S32: Based on the comprehensive operation data, extract the physical connection relationship, control logic relationship, spatial topology relationship and functional dependency relationship between protocol devices, and construct a device semantic association network containing device nodes and associated relationship edges; S33: The large model engine identifies faults based on the device semantic association network, generating predictive maintenance work orders, energy consumption optimization strategies, and fault diagnosis conclusions.

[0024] In this example, after receiving structured semantic data uploaded by the edge gateway through the data access layer of the operations and maintenance platform, the cloud-based large model engine initiates a semantic understanding process and performs multi-source knowledge fusion. Internally, the large model engine, based on the Transformer architecture, possesses a self-attention mechanism and high-dimensional semantic modeling capabilities. It performs deep semantic encoding on fields such as device identifiers, measurement point types, spatial locations, and numerical change trends in the uploaded data, mapping them into time-sensitive vector representations. It also combines historical data sequences to identify potential trend shifts and anomalies. Based on this, the large model engine calls its vector knowledge base interface to retrieve semantic vectors with high semantic similarity to the current device, measurement point, or abnormal signal from embedded large-scale textual knowledge. The knowledge base is constructed from texts such as equipment manuals, maintenance procedures, operating standards, industry specifications, and historical cases, and is stored in a vector database after being encoded into high-dimensional vector forms through an embedded model. For example, when a chiller unit frequently starts and stops and experiences abnormal fluctuations in outlet water temperature, the large model will perform similarity matching between the extracted fault symptom vectors and keyword vectors in the knowledge base related to "compressor malfunction," "control parameter drift," or "load response delay," and extract relevant diagnostic suggestions as supplementary background semantics. Simultaneously, the large model engine, through configured external data interfaces, obtains real-time weather forecast information (such as changes in temperature, humidity, and wind speed over the past 72 hours) and time-of-use electricity price data from the electricity market. It then performs preprocessing on this unstructured external data, converting it into a vector representation through structured modeling, and integrates it with semantic data and knowledge vectors to form a comprehensive operational data structure covering "equipment operating status, maintenance semantic knowledge, and environmental constraints." Based on comprehensive operational data, the large model engine gradually extracts four core relationships between protocol devices through methods such as entity alignment, semantic rule recognition, and contextual path modeling. These include physical connection relationships extracted based on device data flow analysis (e.g., the hydraulic link formed by "chiller unit → chilled water pump → air conditioning unit"), control logic relationships extracted based on the acquisition command response and control signal links (e.g., the feedback control chain composed of "temperature sensor → controller → damper actuator"), spatial topology relationships derived from spatial hierarchy field parsing (e.g., "device A and device B are both located in the north area of ​​the third floor"), and functional dependency relationships mapped based on knowledge base rules (e.g., "UPS supplies power to the distribution cabinet", "security controller depends on access control status input"). These various relationships are uniformly encoded into a device semantic association network using graph structure modeling methods. In this network, each device instance is defined as a graph node, and each relationship is defined as a labeled graph edge, forming a multimodal, multi-semantic, and multi-scale fused knowledge graph network. The network possesses upstream and downstream path traversal capabilities, key device influence propagation analysis capabilities, and structural consistency discrimination capabilities.The large model engine initiates the fault identification and diagnostic reasoning process on the device semantic association network. Through the anomaly detection module, it performs statistical analysis, time-series modeling, and pattern recognition on the semantic data sequence, extracting operational deviation features and alarm signal triggering information of key equipment. Starting from this, it performs reverse path reasoning and impact propagation simulation on the semantic network, gradually identifying potential root cause devices and impact chains. Simultaneously, the large model engine combines historical fault cases, failure mode features, and maintenance records matched from the vector knowledge base, using semantic comparison and causal chain alignment methods to verify whether the current anomaly is consistent with known fault modes. Based on this, predictive maintenance work orders are generated, clearly indicating the equipment to be repaired, the expected failure type, the recommended maintenance time window, and recommended spare parts information. For scenarios where the operating energy efficiency deviates significantly or the external electricity price fluctuates drastically, the large model engine will combine the energy consumption baseline, comfort threshold, and electricity price strategy to automatically generate time-segmented energy consumption optimization strategies and output control parameter sequences and priority configuration suggestions. When a fault has occurred or multiple devices are abnormally linked, the large model will output diagnostic conclusions, providing the root cause device, trigger path, key measurement points, and handling suggestions. All conclusions are structured to generate a fault analysis report and stored in the decision knowledge base.

[0025] In one example, step S33 includes: The large model engine extracts features from the operating parameters in the device semantic association network and retrieves historical failure cases from the vector knowledge base for comparison, thereby obtaining the device health status assessment results and failure risk prediction results; Based on the equipment health status assessment results, predictive maintenance work orders are generated, which include the remaining life prediction value, fault description, recommended maintenance time, and required spare parts list. Energy consumption optimization strategies are generated by combining external data such as weather and electricity prices. The alarm devices are analyzed based on the device semantic association network to obtain fault diagnosis conclusions.

[0026] In this example, based on a device semantic association network, representative operating parameter sequences are extracted from the semantic operation data of devices bound to key nodes in the network. Utilizing the multi-head attention mechanism and temporal modeling capabilities of the large model engine, high-dimensional features are extracted from these parameters, including dynamic features such as temperature fluctuation amplitude, abnormal current changes, start-stop frequency, vibration signal energy spectral density, and operating efficiency deviation. Simultaneously, graph structure information such as spatial location, device type, and upstream / downstream dependencies are combined to construct a multimodal embedded vector representation. The large model engine calls its embedded semantic vector knowledge base to perform semantic similarity retrieval on the extracted feature vectors. It then retrieves historical fault cases, maintenance records, and failure modes most closely related to the current device's operating state, including typical issues such as bearing wear, electromagnetic actuator jamming, communication timeouts, and control logic mismatch. Based on this, a feature comparison matrix between the current device and historical faults is established. From this matrix, the current device's health status score and fault risk level prediction results are derived, along with the prediction of its remaining fault-free operating life, forming a preliminary status assessment report. The large model engine calls the predictive maintenance work order generation module to automatically construct structured work order content based on the equipment's lifecycle position and predicted failure mode type. This includes a clear remaining life prediction value (e.g., "estimated remaining operating cycle is 14~21 days"), a standardized fault type description (e.g., "compressor rotor eccentricity causes abnormal current fluctuations"), a suggested maintenance time window (e.g., "it is recommended to perform a shutdown inspection after 22:00 this Friday"), and a list of required spare parts retrieved from the knowledge base based on the equipment model and failure mode. The work order is then pushed to the task scheduling module of the cloud-based operation and maintenance management platform. Simultaneously, the large model combines high-temperature warnings from weather data, peak-valley time-of-use structures from electricity price data, and overlays predictions of occupancy and load fluctuation trends to generate energy-saving control strategies for HVAC, lighting, and water supply and drainage systems. It suggests starting pre-cooling of the cold source in advance during low-electricity-price periods and raising the set temperature or shutting down some lighting circuits during peak periods. The optimized strategies are then issued to the edge gateway in the form of parameterized control commands. When a device reports an active alarm or a threshold overrun event is detected by the edge, the large model engine takes the alarm device as the central node and backtracks to analyze the operating status of other nodes in the device's semantic association network that have physical connections, control links, spatial overlaps, or functional dependencies with it. It extracts abnormal resonance areas and reconstructs potential fault path maps. At the same time, it searches the vector knowledge base to see if there are history records of multi-device linkage faults under the same combination conditions. Through graph neural network reasoning, it determines whether the alarm is the result of the main device failure, the performance of downstream linkage response, or a false alarm caused by control logic mismatch. It outputs diagnostic conclusions, including the fault initiating device, the trigger time, the impact path, the presumed cause, and handling suggestions.

[0027] In one example, step S4 includes: S41: The cloud sends the energy optimization strategy and lightweight decision model to the edge gateway through a reverse proxy connection. The edge intelligent decision engine in the edge gateway verifies and loads the lightweight decision model. S42: The edge intelligent decision engine runs a lightweight decision model to make local predictions and generate device control parameters. It converts the control instructions of the energy consumption optimization strategy into the corresponding protocol messages of the target device and sends them down for execution control. It obtains the execution feedback data of the target device and reports it to the cloud through a reverse proxy connection.

[0028] In this example, the cloud-based large model engine encapsulates the optimization strategy with a lightweight decision model obtained through knowledge distillation, pruning, or quantization, forming a complete policy distribution data packet that includes model structure definition, parameter weights, execution cycle, applicable scope, control variables, and a secure signature. This packet is then sent to the target edge gateway via a task push mechanism through an established TLS 1.2 encrypted reverse proxy connection. The connection channel relies on the NPS / NPC mechanism, where the cloud-based NPS server verifies the target gateway's identity and locates its network session, ensuring the policy's reachability and transmission integrity across network environments. Upon receiving the policy data packet, the edge gateway's internal edge intelligent decision engine performs digital signature verification and structural consistency checks on the included lightweight decision model. This ensures the model's source is trustworthy, its structure is tamper-proof, and it is executable under current device resource conditions. After successful verification, the model is registered to the local model scheduling module, and model initialization and loading are completed. The edge intelligent decision engine, based on a preset model invocation cycle and input variables such as real-time device operating data, personnel occupancy status, external temperature and humidity, and real-time electricity prices, invokes a loaded lightweight model to perform forward inference. The model may be a compressed LSTM network, a small Transformer structure, or decision tree-based inference logic. The inference result is a predicted value of the control parameters for a specific target device, such as predicting the cooling load demand for the next hour and outputting the optimal air conditioning set temperature, fresh air opening ratio, and compressor start / stop commands. At the same time, the edge intelligent decision engine integrates the energy consumption optimization strategies issued from the cloud with the model output, and combines the local current operating status and strategy priority identifier to determine the actual set of control commands. The edge gateway's protocol adaptation engine queries the target device's protocol type and physical address based on the target device's identifier, and converts the generated control parameters into message data in the device's corresponding protocol format. For example, the command "set air conditioner temperature to 24℃" is encapsulated as a WriteProperty service frame in the BACnet / IP protocol, or the instruction "set chilled water pump frequency to 42Hz" is encapsulated as a Modbus TCP protocol function code 06 control instruction. This control message is then accurately sent to the target device via interfaces such as Ethernet or RS485 bus. After receiving the control instruction, the device completes a state switch and returns an execution confirmation response. The edge gateway simultaneously initiates a control result monitoring process, sampling key measurement points of the target device to verify whether the control is effective and recording feedback data.All execution feedback data, including device identifiers, set parameters, actual response values, control delays, status codes, and timestamps, will be uniformly encapsulated into structured execution feedback objects. After attaching a message authentication code generated based on the HMAC-SHA256 algorithm, the data will be securely uploaded to the cloud platform via a reverse proxy connection. The cloud-based large model engine will receive the feedback data and perform closed-loop verification, comparing the deviation range between the predicted values ​​and the actual results. If the policy execution deviation is found to be too large or the feedback is abnormal, the system will automatically enter the cause analysis process and, based on the accumulation of deviations, activate the model self-optimization mechanism or trigger the policy regeneration task.

[0029] In one example, step S41 includes: S411: The cloud performs knowledge distillation, model parameter quantification, and model structure pruning on the initial decision model in the large model engine to obtain a lightweight decision model. The control instructions for energy consumption optimization strategy and the lightweight decision model are encapsulated with cloud private key signature and then sent to the edge gateway through reverse proxy connection. S412: The edge intelligent decision engine in the edge gateway uses a pre-stored cloud public key to verify the digital signature to confirm that the data source is trustworthy and has not been tampered with, extracts a lightweight decision model and loads it into the local inference engine.

[0030] In this example, the cloud platform selects an initial decision model suitable for a specific system (such as HVAC system, lighting system, elevator energy management, etc.) from the large model engine based on the control tasks required for the current building operation scenario. The initial model has high parameter complexity and expressive power, suitable for global analysis in the cloud. Considering the limitations of edge computing resources and the need for real-time control, the model is lightweighted. The cloud sequentially executes three major processes: knowledge distillation, model parameter quantization, and structural pruning. Through knowledge distillation technology, a teacher model and a student model system are constructed. The teacher model is the original large model, and the student model is the lightweight model to be generated. By aligning the outputs with training on a large amount of actual operating data, the student model can significantly reduce model complexity while maintaining inference accuracy. Model parameter quantization is performed, mapping high-precision floating-point parameters to fixed-point integers or low-bit fixed-point representations, thereby reducing the model's storage footprint and computational load on the edge chip. Structural pruning is performed, trimming redundant neural network channels, sparse connection weights, and invalid branch structures, compressing the model size and optimizing execution efficiency to obtain a lightweight decision model. The control instructions and lightweight decision-making model of the energy optimization strategy are encapsulated to construct a policy data packet containing model definitions, control parameters, applicable scenarios, validity periods, priorities, and instruction encoding sequences. The platform's built-in cloud private key is used to digitally sign the policy data packet, ensuring its immutability and verifiability during transmission. The signed policy data packet is then distributed to the target edge gateway via an established TLS encrypted reverse proxy connection. Upon receiving the policy data packet, the edge gateway's internally deployed edge intelligent decision engine initiates a security authentication process. It uses a cloud public key, pre-issued and stored in a trusted execution environment (such as a TPM or secure storage area), to perform digital signature verification on the received data packet. Specifically, this involves calculating a hash digest of the data packet content and comparing it with the decrypted signature value decrypted using the public key. If the signature verification is successful, the data source is confirmed to be authentic and untampered; otherwise, the data is immediately discarded and an anomaly is logged. After successful verification, the edge decision engine parses the data packet content, extracts the compressed lightweight decision model file and structure definition, and loads it into the local inference engine. The inference engine automatically constructs the corresponding execution graph and allocates computing resources according to the model framework type (such as TinyLSTM, Quantized Transformer, Lightweight Decision Tree, etc.), enabling the model to complete the prediction task in a low-latency manner in the edge device processor or embedded AI module.

[0031] In one example, it also includes: S5: The large model engine performs intent recognition on the natural language query requests input by operations and maintenance personnel, extracts the query intent, and converts the query intent into a natural language query response; S6: The cloud periodically collects execution feedback data, fault cases, and operation and maintenance feedback, labels and cleans them, builds a training dataset, inputs the training dataset into the large model engine for supervised learning to update model parameters, and completes incremental training.

[0032] In this example, a semantic parsing mechanism centered on natural language understanding is established. When maintenance personnel input maintenance-related natural language queries through platform terminals, mobile apps, or voice interaction systems, the queries are received by the large model engine and enter the intent recognition process. Based on its trained language model capabilities, the large model engine performs token segmentation, syntactic analysis, and contextual understanding on the query text. Combining the building equipment semantic model, typical problem template library, and historical context, it extracts the core query intent contained in the user input. Through semantic mapping technology, the natural language intent is transformed into an internally structured query expression or instruction template. The large model engine, combined with the device semantic association network and real-time data stream, calls the corresponding reasoning logic and data analysis modules to extract matching information from the semantic database, maintenance knowledge graph, fault diagnosis records, or energy consumption optimization strategy library. The query results are then re-converted into a natural language response statement that maintenance personnel can understand, completing a closed-loop response process from query intent recognition, structured transformation, data result extraction to natural language response generation. An automated data collection and model training mechanism is constructed, including execution feedback data during each query interaction, input and output records of the reasoning process, actual operational feedback information generated after equipment execution, and manual confirmation feedback and subjective evaluation text from maintenance personnel. This data is automatically aggregated by the cloud platform according to a preset cycle. Combined with the question intent, response content, execution actions, and actual effect labels, a preliminary labeled dataset is constructed. For historical cases related to faults, the occurrence time, fault type, diagnostic reasoning path, equipment semantic relationship chain, and maintenance processing records are structurally labeled. At the same time, maintenance experts intervene to review and complete the quality verification of samples and label consistency correction, forming a training dataset. After data cleaning, text normalization, outlier removal, and semantic enhancement, the training dataset is fed into the large model engine for incremental supervised learning. Using the parameters of the pre-trained model as the initial state, the language encoder, intent recognition module, semantic reasoning layer, and response generation module are fine-tuned and trained respectively. By optimizing the loss function, the model's generalization ability to new questions and understanding of semantic expressions are improved, realizing the continuous evolution of the language model's capabilities and performance self-enhancement, forming a stable and efficient intelligent question answering and knowledge-assisted decision-making system for building maintenance.

[0033] In one example, step S6 includes: The cloud regularly collects execution feedback data, failure cases, and operation and maintenance feedback. For cases with deviations, it performs cause analysis and labels the failures. After cleaning and removing abnormal data, it constructs a training dataset. The training dataset is input into the large model engine, and the loss function between the predicted output and the true label is calculated through supervised learning methods. The backpropagation gradient descent algorithm is then executed to update the model parameters, completing the incremental training.

[0034] In this example, a data acquisition and closed-loop labeling mechanism is established in the cloud. Various execution feedback data are collected periodically through a reverse proxy connection with the edge gateway, including the device's response status to control commands, the difference between actual and predicted energy consumption, task success or failure, and execution time delay. Simultaneously, device fault alarm records and their contextual data are collected, along with feedback information entered by maintenance personnel on the system platform, manually labeled diagnostic conclusions, and maintenance logs. Data normalization and fusion are completed using device identifiers, timestamps, and spatial locations as primary keys. For cases where historical predictions show significant deviations from actual performance, the cause analysis is performed. Upstream and downstream dependency chains are extracted using the device semantic association network to identify input anomalies, feature expression errors, or semantic matching offsets during model inference. Professional maintenance engineers or automated rule engines accurately label these deviation samples with fault type tags, such as "sensor drift," "control signal loss," and "abnormal device response delay." Abnormal data samples caused by communication interruptions, timing errors, or numerical mutations are removed. After cleaning, a training dataset is constructed. During the training phase, the training dataset is input into the large model engine, and the model is incrementally optimized using supervised learning methods. Each sample contains an input feature vector (such as historical measurement point value sequences, equipment semantic relationships, environmental context features, etc.) and a corresponding true fault label or prediction result. The model calculates its predicted output and constructs a loss function with respect to the true label, such as cross-entropy loss or mean squared error loss. The backpropagation algorithm is used to calculate the partial derivatives of the loss function with respect to each model parameter layer by layer, and the parameters are updated using gradient descent, so that the model gradually converges to a better solution under the new data distribution. During training, the learning rate is automatically adjusted and regularization is applied to prevent overfitting. At the same time, a sliding window strategy is used to retain representative samples from the old data to ensure that the model does not lose its memory of historical patterns. This achieves high-precision learning and real-time adaptation of equipment operating status, fault trends, and energy consumption optimization effects, completing the periodic incremental training task of the large model engine.

[0035] Reference Figure 2This embodiment provides a cloud-edge-device collaborative smart building management system integrating a large language model, used to execute the steps of the aforementioned cloud-edge-device collaborative smart building management method integrating a large language model. In this embodiment, the cloud-edge-device collaborative smart building management system is divided into a cloud layer, an edge layer, and a terminal layer. The cloud layer is located at the top of the architecture and includes three core components: an NPS cluster responsible for receiving and managing connections from various edge gateways, achieving secure and efficient connection routing and management; a unified operation and maintenance management platform providing a web interface or API for operation and maintenance personnel to perform operations such as device monitoring, remote configuration, firmware upgrades, alarm management, and data analysis; and a smart building large model engine, the core intelligent module of this system, responsible for intelligent data analysis, predictive maintenance, energy consumption optimization decisions, fault diagnosis, natural language interaction, and knowledge services. The cloud layer and the edge layer are isolated through a firewall / NAT and establish a TLS secure channel for encrypted communication to ensure the confidentiality and integrity of data transmission. The core of the edge layer is the smart building edge gateway, which integrates multiple functional modules: the NPC client is responsible for establishing a reverse proxy connection with the cloud NPS; the protocol adaptation engine is responsible for parsing and encapsulating different building protocols; the near-end decision engine enables local intelligent decision-making and emergency response; the security module performs device access authentication and communication encryption; and the data cache is used for local data storage and breakpoint resume. The edge layer and the terminal layer are connected through different industrial protocols: the Modbus protocol connects to sensors; the BACnet protocol connects to the air conditioning system; and the OPC protocol connects to the lighting system. In this embodiment, data is collected from terminal devices, converted and preliminarily processed by the edge gateway protocol, and uploaded to the cloud through a secure channel. The cloud-based big data engine performs in-depth analysis and then sends decision commands to the edge. The edge gateway then converts the commands into device protocols to control the terminal devices.

[0036] In this embodiment, the specific implementation of each unit in the above system embodiment is described in the above method embodiment, and will not be repeated here.

[0037] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, system, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, system, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, system, article, or method that includes that element.

[0038] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A cloud-edge-device collaborative smart building management method integrating a large language model, characterized in that, include: S1: The edge gateway initiates a reverse proxy connection to the cloud and collects the raw protocol data of each protocol device in the building; S2: The protocol adaptation engine in the edge gateway performs protocol parsing on the original protocol data, obtains semantic data, and uploads it to the cloud through the reverse proxy connection; S3: The cloud-based big model engine receives the semantic data and integrates it with the vector knowledge base of the storage device manual and operation and maintenance procedures, as well as external data such as weather and electricity prices, to generate predictive maintenance work orders, energy consumption optimization strategies, and fault diagnosis conclusions. S4: The cloud sends the energy consumption optimization strategy and lightweight decision model to the edge gateway through the reverse proxy connection, performs local prediction and device control, obtains execution feedback data and reports it to the cloud.

2. The cloud-edge-device collaborative smart building management method based on a fusion of large language models as described in claim 1, characterized in that, Step S1 includes: S11: The NPC client in the edge gateway initiates a TCP long connection to the NPS server in the cloud and performs a TLS 1.2 encrypted handshake. After completing two-way authentication with the NPS server through the X.509 digital certificate, the NPS server registers the edge gateway to the connected NPC registry and maintains heartbeat keep-alive, thus obtaining a verified reverse proxy connection. S12: The protocol adaptation engine in the edge gateway calls the BACnet / IP, ModbusTCP, and OPC UA protocol drivers through the reverse proxy connection to send data reading instructions to each protocol device in the building and receive device response messages to obtain the original protocol data.

3. The cloud-edge-device collaborative smart building management method based on a large language model as described in claim 1, characterized in that, Step S2 includes: S21: The protocol adaptation engine in the edge gateway calls the corresponding protocol parser according to the protocol type of the original protocol data, extracts the object identifier and Present_Value attribute value from the BACnet / IP protocol data, extracts the physical quantity value from the Modbus TCP protocol data according to the register address mapping table, and reads the DataValue corresponding to the NodeId from the OPC UA protocol data to obtain the protocol parsing result; S22: Based on the spatial hierarchy, device classification system, and standard measurement point definition in the unified semantic model, the protocol parsing result is mapped into semantic data containing device identifier, spatial location, measurement point type, value, unit, and timestamp, and then uploaded to the cloud via the reverse proxy connection after attaching the HMAC-SHA256 message authentication code.

4. The cloud-edge-device collaborative smart building management method based on a fusion of large language models as described in claim 1, characterized in that, Step S3 includes: S31: The cloud-based big model engine performs semantic understanding on the semantic data and retrieves semantic vectors from the vector knowledge base of the storage device manual and operation and maintenance procedures, and integrates them with external data such as weather and electricity prices to obtain comprehensive operation data; S32: Based on the comprehensive operation data, extract the physical connection relationship, control logic relationship, spatial topology relationship and functional dependency relationship between the protocol devices, and construct a device semantic association network containing device nodes and association relationship edges; S33: The large model engine identifies faults based on the device semantic association network and generates predictive maintenance work orders, energy consumption optimization strategies, and fault diagnosis conclusions.

5. The cloud-edge-device collaborative smart building management method based on a fusion of large language models as described in claim 4, characterized in that, Step S33 includes: The large model engine extracts features from the operating parameters in the device semantic association network and retrieves historical fault cases from the vector knowledge base for comparison, thereby obtaining device health status assessment results and fault risk prediction results. Based on the equipment health status assessment results, a predictive maintenance work order is generated, which includes the remaining life prediction value, fault description, recommended maintenance time, and required spare parts list. An energy consumption optimization strategy is generated by combining the external data of weather and electricity prices. Based on the semantic association network of the devices, the alarm devices are analyzed to obtain fault diagnosis conclusions.

6. The cloud-edge-device collaborative smart building management method based on a fusion of large language models as described in claim 1, characterized in that, Step S4 includes: S41: The cloud sends the energy consumption optimization strategy and lightweight decision model to the edge gateway through the reverse proxy connection. The edge intelligent decision engine in the edge gateway verifies and loads the lightweight decision model. S42: The edge intelligent decision engine runs the lightweight decision model to perform local prediction and generate device control parameters, converts the control instructions of the energy consumption optimization strategy into protocol messages corresponding to the target device and sends them for execution control, obtains the execution feedback data of the target device and reports it to the cloud through the reverse proxy connection.

7. The cloud-edge-device collaborative smart building management method based on a large language model as described in claim 6, characterized in that, Step S41 includes: S411: The cloud performs knowledge distillation, model parameter quantification, and model structure pruning on the initial decision model in the large model engine to obtain a lightweight decision model. The control instructions of the energy consumption optimization strategy and the lightweight decision model are encapsulated by cloud private key signing and sent to the edge gateway through the reverse proxy connection. S412: The edge intelligent decision engine in the edge gateway uses a pre-stored cloud public key to verify the digital signature to confirm that the data source is trustworthy and has not been tampered with, extracts the lightweight decision model and loads it into the local inference engine.

8. The cloud-edge-device collaborative smart building management method based on a fusion of large language models as described in claim 7, characterized in that, Also includes: S5: The large model engine performs intent recognition on the natural language query requests input by the operation and maintenance personnel to extract the query intent, and converts the query intent into a natural language query response; S6: The cloud periodically collects the execution feedback data, fault cases, and operation and maintenance feedback, labels and cleans them, constructs a training dataset, and inputs the training dataset into the large model engine for supervised learning to update the model parameters and complete incremental training.

9. The cloud-edge-device collaborative smart building management method based on a fusion of large language models as described in claim 8, characterized in that, Step S6 includes: The cloud periodically collects the execution feedback data, fault cases and operation and maintenance feedback, analyzes the causes of cases with deviations and marks them with fault tags, removes abnormal data and completes the cleaning process to build a training dataset; The training dataset is input into the large model engine, and the loss function between the predicted output and the true label is calculated through supervised learning. The backpropagation gradient descent algorithm is then executed to update the model parameters, thus completing incremental training.

10. A cloud-edge-device collaborative intelligent building management system integrating a large language model, characterized in that, Steps for implementing the cloud-edge-device collaborative smart building management method that integrates a large language model as described in any one of claims 1 to 9.

Images