Method for a building control system, building system, electronic device, program product

Abstract

The application provides a method for a building control system, the building control system comprising a cluster of agents, each agent in the cluster of agents being pre-trained to perform a part of a building management function, the method comprising converting received heterogeneous protocol data into structured data having a preset format according to an adaptive mapping rule; distributing the structured data to a target agent in the cluster of agents; parsing the received structured data by the target agent to generate a decision instruction based on the preset format; and sending the decision instruction based on the preset format to a corresponding execution device, wherein when the execution device does not support the preset format, the decision instruction based on the preset format is adapted into a decision instruction supported by the execution device. Corresponding electronic devices, program products, etc. are also provided.

Description

Technical Field

[0001] This application relates to building control technology, and more specifically, to building control systems with dual links. Background Technology

[0002] The control logic of traditional Building Automation Systems (BAS) is pre-edited and loaded into the system. If modifications are needed, reprogramming and reloading are usually required. In addition, users need to input relevant information including device codes and function codes when interacting with traditional BAS, which has a high operational threshold. Summary of the Invention

[0003] At least one of the above problems can be solved by the method, system, etc. for building control systems provided in this application.

[0004] According to one aspect of this application, a method for a building control system includes a cluster of intelligent agents, each agent in the cluster being pre-trained to perform a part of a building management function. The method includes: converting received heterogeneous protocol data into structured data with a preset format according to an adaptive mapping rule; distributing the structured data to target agents in the cluster; parsing the received structured data by the target agents to generate decision instructions based on the preset format; and sending the decision instructions based on the preset format to a corresponding execution device, wherein, when the execution device does not support the preset format, the decision instructions based on the preset format are adapted to decision instructions supported by the execution device.

[0005] According to another aspect of this application, a building control system is also provided, the system including a gateway, which includes a first gateway configured with an intelligent semantic conversion engine, the intelligent semantic conversion engine being used to convert received heterogeneous protocol data into structured data with a preset format according to an adaptive mapping rule; an intelligent agent cluster including multiple intelligent agents, each of the intelligent agents being pre-trained to perform part of a building management function, a target intelligent agent in the intelligent agent cluster receiving the structured data parsing the received structured data to generate a decision instruction based on the preset format; and the gateway sending the decision instruction to a corresponding execution device.

[0006] According to another aspect of this application, a building system is also provided, the building system comprising: an interaction interface for realizing information interaction between the building system and the outside world, the interaction interface being configured to receive input data and output data; a building control system comprising an intelligent agent cluster, the intelligent agents in the intelligent agent cluster being pre-trained to execute a part of the building management function, for executing the building control method described herein, wherein the heterogeneous protocol data is the input data received by the interaction interface; an execution device for receiving decision instructions issued by the building control system and executing corresponding operations according to the decision instructions; and a controlled device that performs state adjustment or action execution under the operation of the execution device.

[0007] According to a further aspect of this application, an electronic device is also provided, the electronic device including a memory for storing instructions; and a processor for executing the instructions and implementing the methods described herein.

[0008] A program product is also provided, which stores instructions that, when executed, implement the methods described herein. Attached Figure Description

[0009] This application will be more fully understood by referring to the following detailed description of specific embodiments in conjunction with the accompanying drawings, in which the same reference numerals refer to the same elements in the views. Wherein: Figure 1 The structure of an LLM-based intelligent agent is illustrated in a simple way; Figure 2 This is a flowchart of a method for a building control system according to some embodiments of this application; Figure 3 This illustrates the process of converting received data into unified structured data with a preset format according to adaptive mapping rules; Figure 4 This is a process according to some embodiments of this application for converting data in a queue into unified structured data based on adaptive mapping rules; Figure 5 yes Figure 2 An example of the process for step S104; Figure 6 This is a schematic diagram of the structure of a building control system according to an embodiment of this application; Figure 7 This is a schematic diagram of a multi-agent interaction process according to some embodiments of this application; Figure 8 This is a flowchart illustrating the operation of a building control system according to an example of this application; Figure 9 This is a flowchart of task execution according to some embodiments of this application; Figure 10This is a structural schematic diagram of a building system according to some embodiments of this application. Detailed Implementation

[0010] To help those skilled in the art to more accurately understand the subject matter to be protected by this application, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0011] An intelligent agent is an artificial intelligence proxy that can be software, hardware, or a combination thereof. An intelligent agent can perceive changes in its environment (e.g., through sensors or data input), make judgments and decisions based on its learned knowledge and algorithms, and then perform actions to influence the environment or achieve predetermined goals. Intelligent agents based on Large Language Models (LLMs) utilize LLMs and multimodal foundational models to achieve flexible and adaptive decision-making. They can process text, images, and audio, enabling them to perform interactions including natural language understanding, autonomous problem-solving, planning, and reasoning.

[0012] Figure 1 The structure of an LLM-based intelligent agent is illustrated in a simple diagram. For example... Figure 1 As shown, the agent receives the task, i.e., the goal to be achieved. The context enhancement module 10 combines the task requirements with the multimodal information in the environment 12 and supplements external knowledge from the knowledge base 14, providing a more comprehensive context for LLM and avoiding problems such as knowledge lag or insufficient information in the model. In this embodiment, the application environment 12 of the agent is building control. In addition to general LLM knowledge, the external knowledge supplemented by the agent from the knowledge base 14 includes various types of knowledge related to building control, such as building control system configuration, control commands, different building scenarios, historical operation data, etc. The agent can read information from the knowledge base 14 through its Retrieval Enhancement Generation (RAG) module 13, making the agent's reasoning more accurate. The multimodal generation module 16 is based on LLM and supports the understanding, fusion, and generation of multimodal data such as text, audio, and images. The multimodal generation module 16 can process inputs such as images, text, and voice, parse the context-enhanced information, and generate multimodal outputs. The agent's interaction module 18 is used to perform specific operations to achieve the task goal, such as calling tools, generating files, and sending messages. The agent can also collaborate with other agents through the interaction module 18 to complete more complex tasks. The agent's iteration module 19 receives feedback from the outside to adjust subsequent decisions based on the actual results of task execution. The agent may also include a content verification module 17 to verify the compliance and accuracy of its generated content, filter out erroneous or non-compliant outputs, and ensure that the agent is safe and controllable.

[0013] Figure 2This is a flowchart of a method for a building control system according to some embodiments of this application. As an example, the method can be executed by the building control system.

[0014] In step S100, the received heterogeneous protocol data is converted into structured data according to adaptive mapping rules. For example, the gateway of the building control system may include a first gateway, which may be configured with an intelligent semantic conversion engine. The intelligent semantic conversion engine is used to automatically convert the heterogeneous protocol data received by the building control system into unified structured data, that is, to convert data under different protocols input to the building control system into structured data with a preset format. In this application example, the data input to the building control system may come from sensors, such as temperature sensors, current sensors, etc., installed in the building. The data input to the building control system may also come from users, such as control commands input by users, such as voice commands input through voice input devices, control commands input through touch screens in the building, etc. Data from different systems and devices is usually data under different protocols. For example, the data generated by cameras in the building security system is ONVIF protocol data, audio data may be MP3 or AAC data, the data of the air conditioning controller may be Modbus protocol data, and the data of the lighting system may be KNX protocol data, etc. In the example of this application, heterogeneous protocol data is automatically converted into unified, structured data that can be recognized by various electronic devices, modules, or components in the building control system through a smart semantic conversion engine, so that the various electronic devices, modules, or components in the building control system do not need to perform data conversion between protocols on their own.

[0015] In this paper, the structured data transformed according to the adaptive mapping rule is data with a preset format. The purpose of introducing this preset format is to ensure that the input data under different protocols of the building control system are all constructed into data in the preset format. From this perspective, this paper also refers to the data in the preset format as multi-protocol fusion format (MCP) data. In other words, MCP data is structured data with a unified data format formed by transforming and normalizing the heterogeneous protocol data input into the building control system.

[0016] In step S102, the converted structured data is distributed to target agents in the agent cluster. The agent cluster includes multiple agents, which are pre-trained to perform a portion of the building management functions. Specifically, the building control system according to the example of this application is configured with an agent cluster, which includes multiple agents that can collaborate based on data in a preset format. The building management functions include the management of different systems and devices in the building, such as building lighting systems and their devices, air conditioning systems and their devices, automatic fire alarm systems and their devices, water leakage alarm systems and their devices, and electrical safety monitoring and emergency response systems and their devices. "Management" here refers to processing operation instructions for these systems and their devices, monitoring the status of these systems and their devices, and handling fault conditions. By way of example, and not limitation, the agent cluster in the embodiments of this application includes agents for building security-related functions, referred to as security agents; agents for device management, referred to as device management agents; and agents for energy management, referred to as energy optimization agents. It should be noted that intelligent agents can also be set up in other ways, such as according to the division of subsystems in the building control system, that is, one intelligent agent is set up for each subsystem.

[0017] In step S104, the target agent parses the received structured data to generate decision instructions based on a preset format, i.e., decision instructions with a preset format or decision instructions under a preset format. Decision instructions typically include the controlled object and the control operation to be performed. The target agent parses the structured data to obtain the information carried by the data, thereby determining which object to perform which control operation, or in other words, determining the event corresponding to the structured data.

[0018] For example, the original input is a user's voice input: "The conference room on the second floor is too hot, turn it down a bit." The first gateway, equipped with an intelligent semantic conversion engine, converts this input into structured data with a preset format in step S100. The target agent in the agent cluster receives and parses this structured data with the preset format, determines that the controlled object is the air conditioner in the conference room on the second floor, and the control operation to be performed is to lower the temperature by 2 degrees. Based on this, it generates a decision instruction to lower the temperature of the air conditioner in the conference room on the second floor by 2 degrees, and this generated decision instruction is constructed as structured data conforming to the preset format.

[0019] In step S106, a decision instruction based on a preset format is sent to the execution device corresponding to the decision instruction. For example, a decision instruction to lower the temperature of the air conditioner in the second-floor conference room by 2 degrees Celsius is sent to the corresponding air conditioning temperature control component via transmission lines such as buses, so that it can perform the operation of lowering the air conditioning temperature. According to the example of this application, the gateway may also include a second gateway disposed between the intelligent agent and the execution device, so that when the execution device does not support data based on the preset format, the second gateway converts the data based on the preset format into data supported by the execution device. For example, if the execution device does not support the decision instruction based on the preset format generated by the intelligent agent, before transmitting the decision instruction to the execution device, the second gateway first converts the decision instruction based on the preset format into data in a format supported by the execution device, so that the execution device can parse the instruction and operate accordingly.

[0020] In some cases, the structured data to be distributed in S102 involves only one target agent. In such cases, the structured data is only distributed to that target agent. For example, the event corresponding to the data on raising the air conditioner temperature is associated with only one agent. In this case, the structured data can be distributed to the agent associated with the event, that is, to the agent that is the target agent.

[0021] In other cases, the events corresponding to the structured data are associated with multiple target agents. According to some examples of this application, the structured data can be distributed to these multiple target agents respectively. For example, based on MCP-based structured data generated by a smoke sensor, the corresponding event is a fire located in zone 3F-A. This event is associated with security, energy optimization, and equipment management agents. Accordingly, the structured data can be distributed to these agents respectively. Accordingly, in step S104, the security agent receiving the structured data can generate decision instructions to activate sprinklers and unlock passageways; the energy optimization agent can generate decision instructions to cut off emergency power; and the equipment management agent can generate decision instructions to disable faulty elevators. In step S106, the decision instructions generated by each agent are distributed to different execution devices. According to the examples of this application, after each execution device completes its execution, it can report the execution status to the agent cluster. In other examples, the structured data can be distributed to one of the associated agents, for example, the primary associated agent as the target agent. After receiving the data, the target intelligent agent then drives other intelligent agents based on its own parsing results. For example, structured data corresponding to a fire event is distributed to the security intelligent agent. The security intelligent agent generates decision commands to start the sprinkler system and unlock the passage, and also drives the energy optimization intelligent agent and the equipment management intelligent agent.

[0022] This can be performed in the building control system, for example Figure 2The method is illustrated. Specifically, the building control system is equipped with a gateway. The gateway may include a first gateway, which converts data entering the building control system, including heterogeneous protocol data from the device side and user side such as sensors and user commands, into unified structured data that the system can recognize. It may also include a second gateway located between the building control system and the execution devices, used to convert the unified structured data into data under a protocol supported by the execution devices. It should be noted that if the execution devices support data based on a preset format, the second gateway does not need to perform conversion. For example, Figure 2 Step S100 can be executed through the first gateway. When it is necessary to convert the decision instructions generated by the intelligent agent based on a preset format into data under the protocol supported by the execution device in step S106, the conversion can be performed through the second gateway. In this way, the heterogeneous data received by the building control system does not need to be converted by each intelligent agent in the intelligent agent cluster, reducing the workload of each intelligent agent and enabling the intelligent agents to achieve seamless collaborative work.

[0023] Furthermore, after parsing the data, the agents in the agent cluster generate decision instructions based on a preset format. If the execution device supports the preset format, the decision instruction can be directly sent to the execution device. If the execution device does not support the preset format, it is converted into data under a protocol supported by the execution device through a second gateway. In contrast, if the agents directly generate decision instructions under a protocol supported by the execution device, and if the execution device changes its protocol or a new execution device supporting the new protocol is added, the instruction generation logic of the agents needs to be modified, resulting in redundancy and complexity, and also hindering collaboration between agents. However, in this application example, because the data transmitted to the agents has already been converted into a unified preset format in, for example, the gateway, the agents do not need to care which protocol the execution device supports. For the building control system, when adding devices or changing protocols in the future, only the relevant rules of the gateway need to be modified, without changing the agent cluster.

[0024] Figure 3 This illustrates the process of converting received heterogeneous protocol data into structured data with a preset format according to adaptive mapping rules. For example... Figure 3 As shown, in step S300, heterogeneous protocol data is received and preliminarily parsed. For example, data from different devices in the building control system and data from input devices in the building control system are received and preliminarily parsed to determine the data type, the address of the device to be controlled, and other information.

[0025] In step S302, based on the preliminary analysis results, the data is categorized into different queues, each with a different priority. For example, security-related data (hereinafter referred to as security data) is categorized into the security data queue; equipment status-related data (hereinafter referred to as equipment data) is categorized into the equipment management data queue; and energy efficiency-related data (hereinafter referred to as energy efficiency data) is categorized into the energy efficiency data queue. In this example, the security data queue has the first priority, the equipment management data queue has the second priority, and the energy efficiency data queue has the third priority. The first priority is higher than the second priority, and the second priority is higher than the third priority. In this document, the first priority is also referred to as high priority, the second priority as second priority, and the third priority as low priority. Furthermore, the division of queues and the setting of priorities are not limited to this example. For example, the division of queues and / or priorities can be more detailed; for instance, air conditioning data and lighting data within equipment data can be divided into different queues. Also, there can be more priority levels, such as from the first priority to the fifth priority.

[0026] After the data enters the queue, in step S304, based on Time Sensitive Networking (TSN), the hardware resources of the building control system are scheduled preferentially for data in the high-priority queue according to the priority of each queue, so that the data in the high-priority queue is processed first. As an example, the TSN function is configured in advance in the building control system executing this method; or the system uses a device with TSN function, such as a system-on-a-chip (SoC) with integrated TSN function.

[0027] According to the example in this application, the TSN module not only processes data according to queue priority to allocate resources to them, but also performs deterministic transmission scheduling based on a preset hybrid arbitration mechanism during resource allocation. Specifically, for security data in high-priority queues, a strict priority scheduling (SP) strategy can be adopted to ensure transmission by reserving 30% bandwidth and a dedicated buffer channel to ensure latency is less than 10ms; for medium-priority device data, a weighted fair queue (WFQ, weight w=0.5) strategy can be adopted to dynamically allocate 20%-50% of bandwidth resources to ensure latency is less than, for example, 10ms; for low-priority energy efficiency data, a round-robin scheduling (RR) strategy can be adopted to share only the remaining bandwidth resources in the transmission path to ensure latency is less than, for example, 100ms. During execution, for security data in high-priority queues, fixed-priority arbitration (FPA) directly triggers the processor accelerators in the hardware system, such as the NVIDIA Deep Learning Accelerator (NVDLA) available on the market, to preemptively suspend low-priority tasks. For medium-priority device data, time-round arbitration (TRA) monitors the queue status and dynamically allocates processor resources in the hardware system, for example, polling the queue status every 100μs and dynamically allocating ARM CPU cores such as Jetson AGX Orin. For low-priority energy-efficient data, weighted random early detection (WRED) manages hardware system resources, randomly discarding data packets when cache occupancy exceeds, for example, 70%, to avoid congestion. Through differentiated scheduling and resource provisioning by the TSN module, as well as the execution of the hardware-level arbitration mechanism, both low latency and reliability of high-priority data transmission are guaranteed, while also taking into account the transmission efficiency of low-priority data. In some applications, such as a 200-node network environment, the measured average latency is only 18 milliseconds. By using time-sensitive networking to process data in different queues, low-latency data transmission can be ensured and high-priority data can be processed first.

[0028] In step S306, the data in the queue is converted into structured data with a preset format according to the adaptive mapping rules.

[0029] In some examples, Figure 3 The process shown can be executed via a pre-trained intelligent semantic translation engine based on a large language model configured in the first gateway. In some other places... Figure 3The process shown can be partially executed by another large language model-based model or agent independent of the intelligent semantic conversion engine, and partially executed by the intelligent semantic conversion engine. For example, the former executes steps S300 to S304, while the intelligent semantic conversion engine executes step S306. It should be noted that, unless explicitly stated otherwise in the examples below this application, Figure 3 The process shown is performed by the intelligent semantic transformation engine.

[0030] Furthermore, in the examples of the methods in this application, the step numbers are only used to distinguish steps, not to limit the data for each step, for example... Figure 3 Step S306 and step S304 can be performed simultaneously or step S306 can be executed first.

[0031] In the example of this application, an adaptive mapping rule is dynamically generated. This mapping rule indicates that objects and values ​​involved in heterogeneous protocol data should be converted into standard objects and values ​​under a preset data format. For example, the adaptive mapping rule is dynamically generated by the intelligent semantic conversion engine according to the following process: obtaining the input corresponding to the register from the protocol standard library, then reading the object corresponding to the input to generate a mapping relationship between the register and the object; and automatically deriving and generating a value range conversion function based on the value range of the register in the protocol standard library and the value range of the corresponding object type. The protocol standard library is a pre-stored standard specification of different industrial protocols involved in the building control system. For example, when Modbus protocol data, such as register 40001, is received, the intelligent semantic conversion engine first determines from the protocol standard library that the object corresponding to Modbus 40001 register is an analog input. Then, it reads that the analog input type of the Modbus 40001 object is BACnet AnalogInPut, thereby deriving the mapping relationship from 40001 to AnalogInPut. The numerical conversion function is automatically calculated based on the 0-65535 value range of Modbus and the 4-20mA value range of BACnet in the protocol standard library, thus determining the calculation formula for f(x). By adaptively and dynamically generating mapping rules based on the protocol standard library, the corresponding mapping rules can be automatically generated for both new registers and new data in a preset format. If the conversion rules are pre-set instead of being dynamically generated, when encountering a new register (such as 40002), rules need to be manually added, and adaptive automatic generation is not possible. According to the example in this application, the dynamically generated adaptive mapping rules can be termed executable code for direct use in specific conversions.

[0032] Standard objects are standardized carriers for heterogeneous protocol data defined based on a pre-formatted data model. Their purpose is to map data from different protocols such as Modbus and KNX to standardized objects with clear semantics. For example, the value of Modbus register 40001 is mapped to an "AnalogInput" object, allowing other components in the system to directly recognize that the data is an analog quantity acquired by a sensor. Standard objects are the data carriers after protocol conversion. Systems executing instructions (such as BACnet controllers and air conditioning / ventilation control systems) read data and execute corresponding instructions based on these standardized objects. For example, the original data is Modbus register 40001 data with a value range of 0-65535. This data is mapped to a BACnet AnalogInput object by the intelligent semantic conversion engine. The BACnet controller, acting as the executor, reads the corresponding 4-20mA value from this AnalogInput object and can then execute instructions to adjust the air conditioning temperature based on that value.

[0033] Figure 4 This describes a process for dynamically generating adaptive mapping rules according to some embodiments of this application. Heterogeneous protocol data (e.g., data in a queue) can be converted into structured data with a preset format based on the generated adaptive mapping rules.

[0034] In step S400, the intelligent semantic conversion engine identifies the register addresses and register values ​​in the received heterogeneous protocol data. For register addresses, the standard object corresponding to the identified register address in a preset format is determined. Specifically, in step S400, the identified register address is input into a semantic association matrix to determine the standard object corresponding to the address in the preset format based on the matching relationship between the register address and the standard object in the preset format. For register values, in step S402, value domain features (including data type, range, unit, etc.) are extracted and used as the values ​​to be converted. A conversion function from the original value domain to the value domain in the preset format is generated by a dynamic value domain converter. In step S404, based on the standard object in the preset format corresponding to the address determined in step S400 and the conversion function generated in step S402, a complete mapping rule is formed, thereby generating code that can be directly deployed and executed in the building control system to automatically convert various protocol data into structured data based on the preset format.

[0035] For example, the received data is a data frame under the Modbus protocol. The intelligent semantic conversion engine parses this heterogeneous data frame to identify protocol field values ​​and extract register addresses. As an example, the intelligent semantic conversion engine includes a pre-trained large language model, more specifically, a BERT pre-trained model. The intelligent semantic conversion engine calls the BERT pre-trained model to parse the received heterogeneous protocol data (also called heterogeneous protocol frames), identifies register addresses and values, and automatically generates mapping rules adapted to the Modbus protocol as the current heterogeneous protocol. It maps Modbus register 40001 to a target object in a preset format, specifically a BACnet simulated input type object. The value range of 0 to 65535 in the Modbus protocol data frame is converted into a preset format-compatible value range of 4mA to 20mA using a value range conversion function based on the adaptive mapping rule. Based on the BACnet simulated input type object and the 4mA to 20mA value range, structured data in the preset format is formed. During data recognition, when the BERT pre-trained model parses the address of register 40001 in the Modbus protocol frame, it calls the mapping rule to determine that the target object corresponding to this field in the preset format is BACnet AnalogInPut (analog input). During numerical conversion, the BERT pre-trained model reads the value of register 40001 in the Modbus protocol data frame (e.g., 32768), calls the domain conversion function f(x) in the mapping rule, and calculates the value in the preset format corresponding to this value (substituting 32768 into f(x) yields 8mA). Based on the object type (AnalogInPut) and the converted value (8mA) in the preset format as defined by the mapping rule, data in the preset format is formed.

[0036] After the target intelligent agent receives a decision instruction based on a preset format, the first scenario is that during the data parsing process, the target intelligent agent can determine that the operation to be executed only involves the systems, components, and functions managed by this intelligent agent; the second scenario is that during the data parsing process, the target intelligent agent determines that there are collaborative operations involving other intelligent agents. In the case of the second scenario, according to some embodiments of this application, Figure 2 In the method for building control systems shown, step S104 further includes steps S1041 and S1042, as specifically illustrated in... Figure 5 In step S1040, the received structured data is parsed to determine whether the event corresponding to the data is associated with other intelligent agents. If the determination result is negative, the process proceeds to step S106 (see...). Figure 2The target intelligent agent generates decision instructions related to the event, and proceeds to step S1041. In step S1041, other associated intelligent agents are driven to generate relevant decision instructions. Driving here refers to the target intelligent agent driving other intelligent agents based on the event corresponding to the structured data, enabling other intelligent agents associated with the event to be aware of the event and make corresponding decisions. In step S1042, the driven other intelligent agents generate relevant decision instructions. If the judgment result of step S1040 is that no other intelligent agents are involved, only step S106 is reached. For example, when the equipment management intelligent agent parses the data, it finds that the event is an air conditioner compressor failure. This event involves energy optimization. Therefore, the equipment management intelligent agent generates corresponding decision instructions on the one hand, and drives the energy optimization intelligent agent on the other hand, including sharing the air conditioner compressor failure information with the energy optimization intelligent agent. The energy optimization intelligent agent generates relevant decision instructions accordingly, including adjusting weights to reduce the load on the air conditioner.

[0037] In a further example, during the process of the target agent parsing received structured data to generate decision instructions, if data content conflicts occur during data parsing, the target agent, at least based on a knowledge base, eliminates the conflict to generate the decision instructions. According to the example of this application, the knowledge base is a pre-configured and updatable knowledge base, where data items may include device identifiers, attributes, and thresholds, and these identifiers, attributes, and thresholds can be stored in a way that is related to each other. Thus, attributes and related thresholds can be directly indexed based on the device identifier. As an example, data content conflicts may include conflicts triggered by ambiguity in device location, ambiguity in operation type, ambiguity in parameter range, and conflicts triggered by user habits. In a further example, in addition to the knowledge base, conflicts can also be eliminated based on historical data and / or user interaction. In a further example, the method of this application example may also include providing feedback on the ambiguity to the user, and updating the knowledge base when the user explicitly indicates that the ambiguity has been eliminated and the feedback differs from the relevant data content in the knowledge base during interaction with the user. Here, the knowledge base is updated dynamically and incrementally while retaining the original data content. Specifically, during operation, the building control system can add new content to the knowledge base without manual intervention or resetting the original configuration, while retaining the original data. For example, it can update by adding one or more new data entries or by adding a version that includes the new data.

[0038] Figure 6 This is a schematic diagram of a building control system according to an embodiment of this application. Figure 6As shown, the building control system includes a gateway 60 and an agent cluster 62. The gateway 60 includes a first gateway for converting received heterogeneous protocol data into structured data with a preset format according to adaptive mapping rules. This structured data can be sent to the agent cluster. The agent cluster 62 includes multiple agents, which can be pre-trained to perform a portion of the building management functions respectively. The target agent in the agent cluster that receives the structured data parses the received structured data to generate decision instructions based on a preset format; that is, it forms the decision instructions into data with a preset format. The gateway 60 then sends the generated decision instructions to the corresponding execution devices.

[0039] In some examples of this application, the first gateway of gateway 60 includes an intelligent semantic translation engine 600. For instance, the intelligent semantic translation engine 600 for building control can be trained based on a large language model; for example, the intelligent semantic translation engine 600 is a pre-trained language model representation (Bidirectional Encoder Representations from Transformers, BERT) based on a transducer architecture. Alternatively, the intelligent semantic translation engine 600 may include a model for building control trained based on a large language model, such as a BERT model.

[0040] In the example of this application, the intelligent semantic conversion engine 600 is configured to initially parse received heterogeneous protocol data; based on the initial parsing results, the data is categorized into different queues with different priorities; based on Time Sensitive Networking (TSN), according to the priorities of different queues, the hardware resources of the building control system are prioritized for data in higher-priority queues, so that the data in the queues is processed first; the data in the queues is converted into structured data based on a preset format according to adaptive mapping rules. In this example, the queues include a security data queue, an equipment management data queue, and an energy efficiency data queue. The intelligent semantic conversion engine 600 categorizes security-related data into the security data queue; equipment-related data into the equipment management data queue; and energy efficiency-related data into the energy efficiency data queue. For example, and not limitingly, the security data queue is set to first priority, the equipment management data queue to second priority, and the energy efficiency data queue to third priority, wherein the first priority is higher than the second priority, and the second priority is higher than the third priority. According to the example of this application, the intelligent semantic conversion engine 600 can perform the above-described combination of... Figure 3 The described process transforms received data into structured data based on a preset format using adaptive mapping rules. Furthermore, the intelligent semantic transformation engine 60 can categorize data into different queues and allocate hardware resources to the data according to queue priority.

[0041] The adaptive mapping rule is dynamically generated, which indicates the correspondence between objects and values ​​in heterogeneous protocol data and standard objects and values ​​in data with a preset format.

[0042] For example, adaptive mapping rules can be dynamically generated by the intelligent semantic conversion engine 600 according to the following process: The input corresponding to the register is obtained from the protocol standard library, and then the object corresponding to the input is read to generate a mapping relationship between the register and the object; based on the value range of the register in the protocol standard library and the value range of the corresponding object type, a value range conversion function is automatically derived. The protocol standard library pre-stores standard specifications for different industrial protocols involved in the building control system. For example, when Modbus protocol data, such as register 40001, is received, the engine first reads the analog input corresponding to the Modbus 40001 register from the protocol standard library, and then reads the corresponding analog input type BACnet AnalogInPut in the Modbus 40001 object model, thereby deriving a mapping relationship from 40001 to AnalogInPut; the numerical conversion function is based on the value range of 0-65535 for Modbus and the value range of 4-20mA for BACnet in the protocol standard library, automatically calculating the linear conversion formula, that is, determining the calculation formula of f(x). Standard object types are standardized carriers defined for heterogeneous protocol data based on a pre-formatted data model. They are used to map data from different protocols such as Modbus and KNX to standard objects with clear semantics. For example, the value of Modbus register 40001 is mapped to an "AnalogInput" object, allowing other components in the system to directly recognize this data as an analog quantity acquired by a sensor. Standard objects are standardized carriers of protocol-converted data. Systems executing instructions (such as BACnet controllers and air conditioning / ventilation control systems) read data and execute corresponding instructions based on these standard objects. For example, the original data is Modbus register 40001 data with a value range of 0-65535. This data is mapped and converted into a BACnet AnalogInput object by the intelligent semantic conversion engine. The BACnet controller, as the executor, reads the value within the 4-20mA range corresponding to this AnalogInput object and can then execute air conditioning temperature adjustment instructions based on this value.

[0043] In a further example, the intelligent semantic translation engine 600 can perform the above-mentioned combination Figure 4 The process described is used to dynamically generate adaptive mapping rules.

[0044] According to some examples of this application, gateway 60 may also include a second gateway 604. When the execution device receiving the decision instruction does not support the preset format of the data, the second gateway 604 adapts the decision instruction formed based on the preset format into data under the protocol supported by the execution device, and then sends it to the execution device.

[0045] In agent cluster 62, a target agent receiving structured data, during data parsing, if it determines that other agents are cooperating, generates decision instructions itself and drives other agents to generate relevant decision instructions. Driving here means the target agent drives other agents based on the operation to be executed; that is, the target agent drives other agents based on the event corresponding to the received structured data. The driven other agents generate relevant decision instructions. For example, the device management agent in agent cluster 62, as the target agent, receives structured data. When parsing the data, the device management agent finds that the corresponding event is an air conditioner compressor failure, which involves the energy optimization agent, thus determining that the event is associated with the energy optimization agent. Based on this, the device management agent drives the energy optimization agent and can share the air conditioner compressor failure information with it. The energy optimization agent generates decision instructions related to the event based on the air conditioner compressor failure information. These decision instructions may include adjusting weights, reducing air conditioner load, etc. As an example, agents in agent cluster 62 can execute the above-described... Figure 5 The method described.

[0046] The building control system according to an embodiment of this application further includes a knowledge base 70. The knowledge base 70 may include multiple data items, each data item may include a device identifier, an attribute, and a threshold, and the device identifier, the attribute, and the threshold are stored in an associated manner. Agents in the agent cluster 62 are configured to, if a data content conflict occurs while parsing received structured data, resolve the conflict based at least on the knowledge base in order to generate a decision instruction. For example, matters causing data content conflicts include device location ambiguity, operation type ambiguity, parameter range ambiguity, and user habit ambiguity. In a further example, when a conflict occurs, the agent may also resolve the conflict based on historical data or user interaction, or both historical data and user interaction. Each agent is configured to provide feedback to the user regarding the matters causing the conflict, and update the knowledge base based on the user's response to the feedback when the user's response can resolve the ambiguity included in the matters.

[0047] As an example, and not a limitation, the intelligent semantic conversion engine 600 is also configured to convert natural language input by the user into a sequence of device-executable instructions using a Large Language Model (LLM). Thus, if the received input is voice, text, image, or video input from the user, the intelligent semantic conversion engine 600 can convert that input into a sequence of device-executable instructions. As an example, the user can input into the building control system through input devices connected to the building control system, such as switches, buttons, touchscreens, keyboards, voice input, and video input. In some examples, the user can also input through smart terminals such as mobile phones and smartwatches that are connected and communicate with the building control system.

[0048] For example, a user uses a smartphone to input voice commands to turn on the lights in the FA (Focus Area) zone. This voice data is transmitted to the gateway 60 of the building control system. The intelligent semantic conversion engine 600 receives the voice input data and converts it into data with a preset format based on the voice protocol. This data is then distributed to the device management agents in the agent cluster 62. The device management agents generate corresponding decision commands and send them to the lighting controllers in the FA zone. If the controller supports the preset format, the decision command can be sent directly through the gateway 60. If the controller does not support the preset format, the second gateway 604 included in the gateway 60 converts the decision command into data under a protocol supported by the controller and then sends it to the controller for execution.

[0049] In some examples, the intelligent semantic conversion engine 600 distributes structured data with a preset format, formed from the received heterogeneous protocol data, to the primary target agent associated with that data. When the primary target agent parses the structured data and identifies that the event corresponding to the data is associated with multiple agents, the primary target agent drives the other associated agents to generate relevant decision instructions, as in the examples listed above.

[0050] In other examples, when parsing received heterogeneous protocol data, if the intelligent semantic conversion engine 600 identifies that the event corresponding to the data is associated with multiple intelligent agents, the intelligent semantic conversion engine 600 can distribute structured data with a preset format formed based on the data to the multiple intelligent agents associated with the event. For example, when the event corresponding to the data received by the intelligent semantic conversion engine 600 is a fire signal in a building, the structured data formed based on the data can be distributed to security intelligent agents and equipment management intelligent agents.

[0051] In other examples, when the intelligent semantic conversion engine 600 parses received heterogeneous protocol data, if it identifies that the event corresponding to the data is associated with multiple agents, it will automatically distribute the structured data associated with multiple agents to all relevant agents, regardless of whether the intelligent semantic conversion engine 600 distributes the structured data to all relevant agents. When an agent receiving the structured data parses the structured data to generate decision instructions, if it identifies that the event corresponding to the structured data also involves other agents, it will actively drive the other agents, for example, by sharing event information with them by sending structured data. The other agents can then generate decision instructions related to the event. For example, if a security agent receives an alarm signal from a smoke sensor, it will generate decision instructions to instruct relevant execution devices based on the alarm signal, including instructing the sprinkler to spray and instructing the door lock system to unlock the passage. Simultaneously, the smoke alarm signal drives the energy optimization agent and the device management agent, sharing data representing the smoke sensor alarm event with them. The driven energy optimization agent generates a decision instruction to cut off the non-emergency power supply at the smoke sensor, and the device management agent generates a decision to disable the elevator in the area indicated by the smoke alarm signal.

[0052] In the example of this application, the target intelligent agent driving other related intelligent agents means that when the target intelligent agent recognizes that an event is related to other intelligent agents, it sends or shares structured data to other intelligent agents, causing the other intelligent agents to trigger the execution of corresponding decision generation and control logic.

[0053] Figure 6 The building control system shown can execute any of the method examples described above in conjunction with the accompanying drawings. For example, the intelligent semantic conversion engine 600 executes steps S100 and S102; the target intelligent agent in the intelligent agent cluster 62 executes steps S104 and S106; the second gateway 604, located between the intelligent agent cluster 62 and the execution device, converts the decision instructions generated by the intelligent agent based on the preset format into data under a protocol supported by the execution device when the execution device does not support data based on the preset format, and then transmits them. The intelligent semantic conversion engine 600 can execute... Figure 3 and Figure 4 The method shown.

[0054] In the example of this application, the equipment management agent, as one of the agents in the agent cluster, is an equipment health management unit that supports data access in a preset format. Accessed sensor data such as elevator vibration and air conditioning current, regardless of their original format, are converted into data in the preset format before being input to the equipment management agent to achieve fault pre-diagnosis. This fault diagnosis is, for example, a multimodal fault diagnosis fusion, where multi-source sensor data is analyzed and fused through a large model, and after alignment operations such as timestamp synchronization, it is aggregated into the agent cluster. In the equipment management agent, the topology fusion of the equipment is pre-completed during training, including constructing a topology map based on standard objects in a preset format (such as spatial mapping of elevator location and air conditioning area), and predicting the fault impact range (such as elevator bearing faults and associated areas).

[0055] In the example of this application, the energy optimization agent, as one of the agents in the agent cluster, is trained to dynamically adjust the relationship between the system's energy consumption, comfort, and cost. The energy optimization agent supports data access in a preset format. Dynamically acquired multi-source data, such as grid electricity prices, environmental sensor data, and personnel location data, are converted into preset format data and input into the energy optimization agent, which then outputs energy-saving strategies to building equipment (such as air conditioning and lighting). The energy optimization agent directly obtains structured data (such as energy objects (MCP EnergyObjects) in a preset format, including electricity prices, equipment energy consumption, and personnel density) from the bus and performs standard output, that is, converting decision commands into control objects (MCP Control Objects) in a preset format, such as "adjust the air conditioning temperature setpoint by ±2℃", and then sends it to the execution device via the BACnet protocol and bus.

[0056] As one of the agents in an intelligent agent cluster, the security agent is an emergency response unit. It is pre-trained to adopt a hierarchical response mechanism, receiving security events (such as fires and intrusions) and triggering cross-device linkage. As an example, a fire is a level-one event, and an intrusion is a level-two event. The fire event type is an energy event (MCP_Emergency_Event) in a preset format, and the corresponding response actions may include shutting down the fresh air supply, starting the sprinkler system, and unlocking the passage, with a response time delay of no more than 3 seconds. The intrusion event type is a security alert (MCP_Security_Alert) in a preset format, and the corresponding response actions may include laser fencing, alarm linkage, and camera tracking, with a response time delay of no more than 5 seconds.

[0057] The agents do not operate independently but rather collaborate. For example, the energy optimization agent can work in conjunction with the equipment management agent for monitoring equipment status and the security agent for handling emergency events to achieve cross-agent target linkage. In one example, when the equipment management agent detects an equipment anomaly (such as an air conditioning compressor malfunction), the energy optimization agent automatically adjusts the control weights, reducing the load on the faulty equipment to prioritize the normal operation and user comfort of equipment in other areas. As another example, when the security agent triggers a Level 1 fire emergency response, the energy optimization agent can immediately adjust the power supply weights, elevating the emergency power supply to the highest level to ensure power supply for sprinklers, lighting, escape routes, and other equipment.

[0058] Figure 7 This is a schematic diagram of a multi-agent collaboration process according to some embodiments of this application. The interaction process is as follows: Figure 6 The building control system shown is in operation.

[0059] In step S700, the user makes voice input through an interactive interface, such as through the voice input interface in the building control system or through a smart terminal that is connected to the building control system.

[0060] In step S702, a large language model set in the building control system receives user voice input to determine user intent and form user instructions. For example, the large language model analyzing the instructions can be an LLM-based model independent of the intelligent semantic translation engine. This large language model receives user input and can determine user intent based on the user input and the interaction used, thereby forming user instructions. In some examples, step S702 can also be performed by an agent using an LLM-based model. In another example, the LLM-based model performing step S702 to analyze user input is the intelligent semantic translation engine 600; that is, the intelligent semantic translation engine 600 can determine user intent based on the user input and the interaction used, thereby forming user instructions.

[0061] In step S704, the intelligent semantic conversion engine 600 converts user instructions into structured data based on a multi-protocol fusion protocol according to adaptive mapping rules. For example, during the execution of step S704, the intelligent semantic conversion engine 600 also categorizes security-related data into a security data queue; categorizes device status-related data into a device management data queue; and categorizes energy efficiency-related data into an energy efficiency data queue, so as to schedule resources for the data in the queues according to their priorities.

[0062] In an alternative example, during step S702, security-related data can be categorized into a security data queue; equipment status-related data can be categorized into an equipment management data queue; and energy efficiency-related data can be categorized into an energy efficiency data queue, so as to schedule resources for the data in the queues according to the queue priorities.

[0063] If the structured data being converted is in a high-priority security data queue, the intelligent semantic conversion engine 60 will schedule resource allocation based on TSN (Transmission Service Number) to prioritize the transmission bandwidth and other resources in the building control system's transmission links for this data. For example, it will reserve 30% bandwidth and a dedicated buffer channel to ensure its transmission and guarantee a latency of less than 10ms. Furthermore, it can directly trigger the processor's accelerator in a fixed-priority context to pause the processor's execution of low-priority tasks.

[0064] Furthermore, if a conflict occurs during the parsing of user input instructions (in step S702), the conflict can be resolved by retrieving information from the knowledge base; if the conflict cannot be resolved using the knowledge base, it can be resolved through context, historical data, or user interaction. The following factors may lead to conflicts: device location ambiguity, operation type ambiguity, parameter range ambiguity, and user habit ambiguity. Examples of device location ambiguity, operation type ambiguity, parameter range ambiguity, and user habit ambiguity are given below. It should also be noted that, although in combination... Figure 7 In the example, the knowledge base is used to disambiguate the instruction part that accepts user input in step S702. However, the method and process of disambiguating using the knowledge base can be executed in multiple processing stages of the embodiments of this application. For example, it can also be used to disambiguate when the target intelligent agent parses structured data.

[0065] Equipment location ambiguity The user inputs the command "Turn off the conference room air conditioner". To determine the specific location, in step S702, for example, the intelligent semantic conversion engine 60 (or alternatively, an LLM-based intelligent agent set independently of the intelligent semantic conversion engine 60) searches the knowledge base to determine the specific location of the air conditioner targeted by the command. According to the search results, the conference room air conditioner corresponds to two device IDs: the air conditioner in conference room 101 on the 1st floor with ID AC-001, and the air conditioner in conference room 201 on the 2nd floor with ID AC-002. Because the user command does not specify a specific floor, the atomic command cannot be uniquely determined. According to the example of this application, in step S702, the location can be determined by calling the "device ID-attribute" triple in the knowledge base. The "device ID-attribute" triple refers to a triple data structure consisting of device ID, attribute name, and attribute value, such as {AC-001: [location, conference room 101 on the 1st floor], AC-002: [location, conference room 201 on the 2nd floor]}). If calling the "device ID-attribute" triple reveals that "conference room air conditioner" has multiple device mappings, conflict detection is triggered. In conflict detection, user interaction logs can be queried. If 80% of a user's "meeting room" commands in the past 7 days refer to meeting room 301 on the 3rd floor, then the system will prioritize interacting with the user to inquire about... Please confirm that 3 is closed. Is the air conditioner in conference room 301 on the first floor, or in conference room 101 on the first floor or conference room 201 on the second floor? This approach, combined with user habits, reduces interaction costs. For example, if a user requests to turn off the air conditioning in conference room 301 on the 3rd floor, the command can be completed with "conference room air conditioning" as device ID=AC-003 (conference room air conditioning on the 3rd floor), generating the atomic command: [AC-003, operation type=off, parameter value=0]. Furthermore, following the example in this application, the intelligent semantic conversion engine 60 can also incrementally update the association between "conference room air conditioning" and "3rd floor 301" to the knowledge base ("user habit adjusts default value: conference room default=3rd floor 301"); subsequently, when the user inputs "turn off conference room air conditioning," the intelligent agent automatically completes it with "3rd floor 301," eliminating the need for a separate prompt.

[0066] Operation type ambiguity A user inputs the command "It's stuffy and uncomfortable in the meeting room" without specifying any action. For example, the Intelligent Semantic Transformation Engine 60 (or, alternatively, an LLM-based agent set up independently of the Intelligent Semantic Transformation Engine 60) parses this command and finds the action type ambiguous, triggering conflict detection. In the conflict detection, the Intelligent Semantic Transformation Engine 60, combining building domain knowledge ("stuffy" is usually caused by "excessive carbon dioxide concentration" or "excessive humidity"), generates a rhetorical question: "Should we activate the fresh air system (increase ventilation) or lower the air conditioning temperature (relieve the stuffiness)?", guiding the user to choose the action more suited to their actual needs. The user responds by activating the fresh air system. Based on this, the Intelligent Semantic Transformation Engine 60 generates the atomic command: [Fresh Air System ID=VF-001, Action Type=Activate, Parameter Value=50%] (Fresh air volume adjusted to 50%); and implements this as structured data based on MCP for distribution. Furthermore, the intelligent semantic conversion engine 60 can add the association between "stuffy" and "start the fresh air system" to the "scenario-based rule" template ("user needs 'stuffy' are prioritized for association with the fresh air system"); when the user enters "the meeting room is too stuffy", the intelligent agent directly generates the instruction to "start the fresh air system" without asking a question.

[0067] Parameter range ambiguity When a user inputs "adjust the meeting room temperature to 32℃," the Intelligent Semantic Transformation Engine 600 (or, alternatively, an LLM-based agent set independently of the Intelligent Semantic Transformation Engine 600) identifies that this value exceeds the threshold. The Intelligent Semantic Transformation Engine 600 calls the "threshold verification" module in the knowledge base and finds that "32℃" exceeds the temperature limit of "AC-003" (30℃), triggering conflict detection. In the conflict detection, the Intelligent Semantic Transformation Engine 600, combined with building safety standards (GB50314-2015 "Intelligent Building Design Standard" requires air conditioning temperatures to be controlled between 18-28℃), generates a question: "The current meeting room air conditioning temperature threshold is 16-30℃. 32℃ exceeds the safe range. Do you confirm the adjustment? It is recommended to adjust to 28℃." This both highlights the risk and preserves the user's choice. The user responds, "Confirmed adjusting to 32℃." In this scenario, the Intelligent Semantic Conversion Engine 600 initiates an advanced security mechanism for secondary verification and execution. This includes: re-verifying user permissions (ensuring the user has "over-threshold operation" permissions, such as an administrator); invoking "device linkage rules" (e.g., "forcefully start the fresh air system when the temperature exceeds 30℃") to generate a compensation instruction: [VF-001, Operation Type = Start, Parameter Value = 100%] (adjusting the fresh air volume to 100% to alleviate discomfort caused by high temperatures); and converting the original instruction (32℃) and the compensation instruction (starting the fresh air system) together into MCP-based structured data for execution. Furthermore, the Intelligent Semantic Conversion Engine 600 records the user's preference for "allowing over-threshold operations" in the knowledge base ("user allows air conditioning temperature to exceed the threshold to 32℃"). Subsequently, when the user inputs "adjust to 32℃," the Agent executes directly (without requiring a prompt), but the compensation instruction will still be triggered.

[0068] User habit ambiguity For example, if a user inputs "adjust the meeting room temperature to 30℃," the intelligent semantic translation engine 600 (or alternatively, an LLM-based agent set up independently of the intelligent semantic translation engine 600) might recognize that adjusting the temperature to 30℃ is not the user's usual setting; according to historical data, the user's preferred temperature setting is 28℃. The intelligent semantic translation engine 600 can then generate a follow-up question, "Adjust to 30℃ or 28℃?", to confirm the input and ultimately generate a decision instruction based on the user's input.

[0069] Back Figure 7 The structured data obtained by the intelligent semantic conversion engine 600 is distributed as follows: if it is security-related data, it is sent to the security agent 621 in the agent cluster; if it is device status data, it is sent to the device agent 622 in the agent cluster; and if it is energy efficiency-related data, it is sent to the energy optimization agent 623.

[0070] In this example, when the intelligent semantic conversion engine 600 identifies that the event corresponding to the data is associated with multiple agents in the agent cluster during the process of converting heterogeneous protocol data into structured data, it can distribute the resulting structured data to the multiple agents respectively.

[0071] At the security agent 621, the agent 621 can first classify events into levels. According to the example in this application, fire is level 1 and intrusion is level 2; that is, a fire event should be responded to at level 1, and an intrusion event should be responded to at level 2. It should be noted that the level can be set as needed, for example, a fire event can be set to level 0 or level 1. If the event is a fire as a level 1 event, the response actions may include shutting off the fresh air supply, activating the sprinkler system, and unlocking the passage, with a delay of less than 3 seconds. Based on a preset format, the type of a level 1 event can be represented as MCP_Emergency_Event. If the event is an intrusion as a level 2 event, the response actions should include laser fencing, alarm linkage, and camera tracking, with a delay of less than 5 seconds. Based on a preset format, the type of a level 2 intrusion event can be represented as MCP_Security_Alert.

[0072] When the received structured data in a preset format indicates an event type of MCP_Emergency_Event, the security agent 621 generates a decision instruction to initiate a response action. In the example of this application, the security agent 621 can also generate drive instructions to be sent to two other agents related to the event. In some examples, the security agent 621 sends the drive instructions back (step S710) to the intelligent semantic conversion engine 600, which then sends the instructions (S712) to the two other agents related to the instructions. It should be noted that the security agent 621 can also share event information, i.e., structured data, with the other two agents. The sharing methods include sending the event information along with the drive instructions to the other two agents.

[0073] At the device management agent 621, the received structured data is processed to generate decision instructions. If the device management agent 621 receives a drive instruction from the security agent 621, it generates a decision instruction to disable the elevator. The device management agent 621 also generates a fault warning signal and sends it back (S722) to the intelligent semantic conversion engine 600.

[0074] At the energy optimization agent 622, the received structured data is processed to generate decision instructions. If the energy optimization agent 622 receives a drive instruction from the security agent 621, it generates a relevant decision instruction to cut off non-emergency power. The energy optimization agent 622 can then send this instruction back (S723) to the intelligent semantic conversion engine 600.

[0075] In the example of this application, data is stored in the knowledge base in the form of triples. For example, the knowledge base stores data in a "device ID-attribute-threshold" structure, such as: {Conference Room Air Conditioner: [Temperature, 16-30℃]}. This structured storage method allows the intelligent agent to quickly match the device, attribute, and valid range when invoking commands, performing legality verification and security filtering on the commands, automatically rejecting dangerous or invalid commands that exceed the threshold (such as rejecting the abnormal command "adjust the temperature to 50℃"), thereby improving the security, reliability, and execution efficiency of the building control system and preventing equipment damage or malfunctions caused by illegal commands.

[0076] In the example of this application, when the received control command is ambiguous, vague, or incomplete, a natural language questioning interaction mechanism can be automatically triggered to clarify the user's true intention. For example, if the user command "turn off the air conditioner" does not specify a specific floor or device, an LLM-based model or intelligent speech conversion engine can be triggered to ask, "Please confirm that the air conditioner in the conference room on the 1st or 2nd floor is turned off?" This achieves automatic resolution of command ambiguity, accurate identification of control intention, and reliable iteration and traceable management of the knowledge base, thereby improving the system's intelligence level and interaction reliability.

[0077] The LLM-based intelligent semantic transformation engine and the agents in the agent cluster in this application embodiment can both be pre-trained. The training process includes data acquisition and collection, and federated learning configuration. For example, energy data such as temperature, humidity, and CO2 (32-dimensional features with a time span of ≥1 year) can be collected to train the energy optimization agent, and fault datasets including time-series signals such as elevator vibration spectrum and air conditioner current waveform can be used to train the equipment management agent. For the security agent configured with federated learning, differential privacy (ε=1.2) is used to protect model parameters between edge nodes; the cloud aggregator can be deployed in a Software Guard Extensions (SGX) trusted execution environment, where SGX secure aggregation can be performed once every 6 hours.

[0078] Figure 8 This is an example of a building control system operation flowchart according to this application. For example... Figure 8As shown, in step S800, devices supporting different protocols are connected to the building control system. In step S802, data input to the building control system is received, and heterogeneous protocol data is converted into data with a preset format. In step S804, it is determined whether a data anomaly has occurred based on the data. If the determination result of step S804 is that an anomaly has occurred, then proceed to step S806; otherwise, continue executing step S804. In step S806, the intelligent agent cluster 8060 is triggered to handle the event corresponding to the anomaly. Specifically, based on the event corresponding to the anomaly, data with a preset format is distributed to the target intelligent agent in the intelligent agent cluster 8060 that handles the event. The target intelligent agent may be one or more of the following: energy optimization intelligent agent 8060A, security intelligent agent 8060B, and device management intelligent agent 8060C. In step S808, the target intelligent agent generates a decision instruction based on a preset format; that is, the generated decision instruction is data with a preset format. In step S810, the decision instruction is issued to the execution device. In some examples, prior to step S810, the decision instruction may be converted into data under a protocol supported by the execution device and then sent to the execution device. (Combined with...) Figure 8 In the described example, step S802, receiving data input to the building control system can be user input, and also includes data collected from the building system. The data collection period is, for example, 200 milliseconds for critical equipment and 5 seconds for ordinary equipment.

[0079] Figure 9This is a flowchart of task execution according to some embodiments of this application. In step S900, the building control system receives a natural language request from a user, such as keeping the room cool for a meeting at 9:00 AM tomorrow. In step S901, an LLM-based agent receives this request. The agent executes steps S902 and S904 respectively. In step S902, the LLM-based agent, based on the user input, invokes a knowledge base to determine the scene mode as meeting mode, determines the associated equipment including air conditioning, lighting, and fresh air, and determines that the safety rule requires the temperature range to be between 16 and 30 degrees Celsius. In step S904, the LLM-based agent, based on the user input, invokes real-time data, for example, but not limited to, including the current meeting room temperature of 28 degrees Celsius, the number of attendees, and the external weather conditions. In step S906, based on the information obtained in steps S902 and S904, a control strategy is dynamically generated, for example, pre-cooling to 24 degrees Celsius one hour in advance and fine-tuning every 30 minutes during the meeting. In step S908, the generated control strategy is decomposed into an atomic instruction sequence, including instruction 1, instruction 2, and instruction 3. Instruction 1, for example, sets the air conditioner to 24 degrees Celsius; the instruction data triplet is, for example, device ID AC-301 with parameter 24. Instruction 2, for example, sets the light brightness to 70%; the instruction data triplet is, for example, device ID light-301 with parameter 70. Instruction 3, for example, sets the fresh air volume to 60%; the instruction data triplet is, for example, device ID Air-301 with parameter 60.

[0080] Here, step S906 can be performed, for example, by an intelligent semantic conversion engine. Thus, the user's natural language input is converted into structured data in a preset format. During the parsing process in step S906, it is determined that the event corresponding to the data is associated with a device management agent in the agent cluster. Therefore, in step S908, the device management agent that received the structured data further parses the data, including breaking it down into atomic instructions and generating decision instructions for different executing device operations, including instruction 1, instruction 2, and instruction 3.

[0081] In this example, in step S911, the second gateway in the gateway (such as a protocol adaptation engine) converts Command 1, Command 2, and Command 3, which have preset formats, into data under the protocols supported by the corresponding execution devices. For example, Command 1 is converted into air conditioning protocol data, Command 2 is converted into lighting protocol data, and Command 3 is converted into fresh air protocol data.

[0082] As shown in step S912, the converted data is sent to the client. After authorization verification, it can be sent to the Building Performance System (BPS) rules server via Message Queuing Telemetry Transport (MQTT), as shown in step S914. In step S916, each executing device executes the command and reports its status.

[0083] Figure 10 This is a building system according to some embodiments of this application. The building system includes an interaction interface 1000, a building control system 1002, an execution device 1004, and a controlled device 1006. The interaction interface 1000 is used to realize information interaction between the building system and the outside world. The interaction interface is configured to receive input data and output data. The interaction interface includes an input interface, such as one or more of a touch screen, a button panel, a voice acquisition component, an infrared remote control interface, a face recognition module, and a gateway interaction interface. The interaction interface includes an output interface, which can be a display component, a voice playback component, etc. Further, the interaction interface may also include a mobile terminal interface, such as an application that can be used on a mobile terminal such as a mobile phone or tablet. The building system communicates with the mobile terminal through a communication interface, thereby making the application set on the mobile terminal the interaction interface. The building control system 1002 is, for example, the one described above. Figure 6 The building control system described herein. In another example, the building control system 1002 includes a cluster of intelligent agents, each of which is pre-trained to perform a part of the building management functions. Further, the building control system 1002 is configured to perform any of the building control methods described above. An execution device 1004 is used to receive decision commands issued by the building control system 1002 and execute corresponding operations according to the decision commands. The execution device 1004 may include one or more of relays, valves, frequency converters, drives, and electric actuators, and may also include components not listed herein, which are not limited herein. The controlled device 1006, under the operation of the execution device 1004, performs state adjustment or action execution, and may include one or more of building lighting equipment, air conditioning equipment, ventilation equipment, water pumps, elevators, access control equipment, sunshade equipment, and heating equipment.

[0084] According to embodiments of this application, a dynamic mapping architecture based on a preset format is constructed in the building control system. This allows data input into the building system under different protocols to be smoothly transmitted within the building control system after being converted to the preset format. Specifically, through dynamically generated automatic mapping rules, data under different protocols can be converted to the preset format at the system entry point (e.g., the gateway). This eliminates the need for components and modules within the system to perform conversions themselves; they only need to identify and process the data, significantly saving resources within the building control system and accelerating processing speed. Embodiments of this application also employ a TSN scheduling mechanism, prioritizing high-priority data such as security data by categorizing data into queues with different priorities. According to some embodiments of this application, the building control system can respond to user-defined natural language requests, lowering the interaction threshold. Furthermore, embodiments of this application integrate multiple intelligent agents into an intelligent agent cluster, achieving full lifecycle management of the building and effectively solving the problems of protocol fragmentation, isolated decision-making, and privacy leakage in traditional building systems.

[0085] While specific embodiments of this application have been shown and described in detail to illustrate its principles, it should be understood that the technical features of the various embodiments of this application can be combined with each other to form new implementation methods without contradiction. Furthermore, this application can be implemented in other ways without departing from the principles described herein.

Claims

1. A method for a building control system, characterized in that, The building control system includes an agent cluster, wherein the agents in the agent cluster are pre-trained to perform a part of the building management functions, and the method includes: The received heterogeneous protocol data is converted into structured data with a preset format according to adaptive mapping rules; The structured data is distributed to the target agents in the agent cluster; The target intelligent agent parses the received structured data to generate decision instructions based on the preset format; The decision instruction is sent to the corresponding execution device, wherein, when the execution device does not support the preset format, the decision instruction is adapted to a decision instruction supported by the execution device.

2. The method according to claim 1, characterized in that, The step of converting the received heterogeneous protocol data into structured data with a preset format according to adaptive mapping rules includes: Preliminary analysis of the received heterogeneous protocol data; Based on the analysis results, the heterogeneous protocol data are categorized into different queues; Based on time-sensitive networking, the hardware resources of the building control system are scheduled preferentially for data in high-priority queues according to the priority of each queue, so that the data in the high-priority queues can be processed first.

3. The method according to claim 2, characterized in that, The step of classifying the heterogeneous protocol data into different queues includes: Data related to security is categorized into the security data queue; Data related to the equipment is categorized into the equipment management data queue; Data related to energy efficiency is categorized into the energy efficiency data queue.

4. The method according to claim 3, characterized in that, The security data queue has the first priority, the device management data queue has the second priority, and the energy efficiency data queue has the third priority. The first priority is higher than the second priority, and the second priority is higher than the third priority.

5. The method according to claim 1, characterized in that, The step of converting the received heterogeneous protocol data into structured data with a preset format according to adaptive mapping rules includes: using a pre-trained intelligent semantic conversion engine based on a large language model to convert the heterogeneous protocol data into structured data with the preset format according to the adaptive mapping rules; the adaptive mapping rules are dynamically formed according to the following process: Identify register addresses and register values ​​from the heterogeneous protocol data; Based on the identified register address and the preset semantic association matrix, determine the standard object under the preset format corresponding to the register address; Extract the value range characteristics of the register value to generate a conversion function from the original value range to the value range under the preset format; Based on the standard object in the preset format corresponding to the determined address and the conversion function, a complete adaptive mapping rule is formed.

6. The method according to claim 1, characterized in that, The process of parsing the received structured data by the target intelligent agent to generate decision instructions further includes: When the target intelligent agent identifies that the event corresponding to the data is associated with other intelligent agents during the process of parsing the received structured data, it drives the other related intelligent agents to generate relevant decision instructions.

7. The method according to claim 1, characterized in that, The intelligent agent cluster includes security intelligent agents, equipment management intelligent agents, and energy optimization intelligent agents.

8. The method according to claim 1, characterized in that, The method of generating decision instructions by parsing the received structured data through the target intelligent agent also includes: If a conflict is determined to exist in the data content by parsing the received structured data, the target intelligent agent shall eliminate the conflict based at least on a knowledge base. The knowledge base includes multiple data items, each of which includes a device identifier, an attribute, and a threshold, and the device identifier, the attribute, and the threshold are stored in a way that is related to each other. Among them, the issues that cause conflicts in the data content include one or more of the following: ambiguity in device location, operation type, parameter range, and user habits.

9. The method according to claim 8, characterized in that, The target intelligent agent eliminates the conflict based on historical data and / or interaction data with the user.

10. The method according to claim 8, characterized in that, The method further includes: providing feedback to the user regarding the matter causing the conflict, and updating the knowledge base based on the user's response to the feedback when the user's response can eliminate the ambiguity included in the matter.

11. The method according to claim 10, characterized in that, The knowledge base is updated dynamically and incrementally while retaining the original data content.

12. A building control system, characterized in that, The building control system includes: The gateway includes a first gateway configured with an intelligent semantic conversion engine, which is used to convert received heterogeneous protocol data into structured data with a preset format according to adaptive mapping rules; An intelligent agent cluster, comprising multiple intelligent agents, each pre-trained to perform part of a building management function, wherein a target intelligent agent in the cluster that receives the structured data parses the received structured data to generate decision instructions based on the preset format; and The gateway sends the decision command to the corresponding execution device.

13. The building control system according to claim 12, characterized in that, The gateway also includes: The second protocol gateway is used to adapt the decision instruction to a data format supported by the corresponding execution device before sending it to the corresponding execution device.

14. The building control system according to claim 12, characterized in that, The intelligent semantic conversion engine is also configured to perform the method according to any one of claims 2 to 6.

15. The building control system according to claim 12, characterized in that, Each agent in the agent cluster is configured to perform the method according to any one of claims 8 to 11 when acting as a target agent.

16. The building control system according to claim 12, characterized in that, The intelligent agent cluster includes security intelligent agents, equipment management intelligent agents, and energy optimization intelligent agents.

17. A building system, characterized in that, The building system includes: An interactive interface is used to enable information interaction between the building system and the outside world. The interactive interface is configured to receive input data and output data. A building control system comprising a cluster of intelligent agents, wherein the agents in the cluster are pre-trained to perform a part of a building management function for executing the building control method according to any one of claims 1 to 11, wherein the heterogeneous protocol data is the input data received by the interaction interface; An execution device is used to receive decision instructions issued by the building control system and perform corresponding operations according to the decision instructions; A controlled device that performs state adjustment or action execution under the operation of the execution device.

18. The building system according to claim 17, characterized in that, The interactive interface includes one or more of the following: touch screen, button panel, voice acquisition component, voice playback component, display component, mobile terminal interactive interface, remote communication interface, infrared remote control interface, face recognition module, and gateway interactive interface.

19. An electronic device, characterized in that, The electronic device includes: Memory, used to store instructions; A processor for executing the instructions and implementing the method according to any one of claims 1 to 11.

20. A program product, characterized in that, The program product storage instructions, when executed, implement the method according to any one of claims 1 to 11.

Images