Control device, control system, and control method for air conditioner

Abstract

This control device comprises: a communication interface that performs data communication; and a control unit that controls an inference process in which a prompt is inputted to a large-scale language model and an output sentence is generated by the large-scale language model. The large-scale language model is trained through machine learning so as to generate the output sentence on the basis of the inputted prompt. The control unit: acquires, via the communication interface, text that is included in the prompt as text indicating a user request to an air conditioner and is inputted to the large-scale language model; acquires operation state data indicating the operation state of the air conditioner on the basis of an output sentence generated through the inference process from a first prompt including the text; and changes the operation state of the air conditioner via the communication interface on the basis of an output sentence generated from a second prompt including the operation state data.

Description

Control Device, Control System, and Control Method for Air Conditioner

[0001] The present disclosure relates to a control device, a control system, and a control method for an air conditioner.

[0002] Patent Document 1 discloses a portable terminal device that remotely operates an external device such as an air conditioner in response to voice detected via a microphone unit. When a user emits a "power on" voice, the portable terminal device of Patent Document 1 transmits a control command for turning on the power of the air conditioner to the air conditioner.

[0003] Japanese Patent Application Laid-Open No. 2007-135008

[0004] Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan. (2022). ReAct: Synergizing Reasoning and Acting in Language Models. arXiv preprint arXiv:2210.03629.

[0005] The present disclosure provides a control device, a control system, and a control method that can easily control an air conditioner according to user demands and operating states related to the air conditioner.

[0006] A control device for an air conditioner according to one aspect of this disclosure comprises a communication interface for data communication and a control unit that controls an inference process that inputs a prompt to a large-scale language model and generates an output sentence by the large-scale language model. The large-scale language model is trained by machine learning to generate an output sentence based on the input prompt. The control unit obtains text indicating the user's request to the air conditioner, which is included in the prompt and input to the large-scale language model, via the communication interface. The control unit obtains operating state data indicating the operating state of the air conditioner based on an output sentence generated by the inference process from a first prompt containing the obtained text. The control unit changes the operating state of the air conditioner via the communication interface based on an output sentence generated by the inference process from a second prompt containing the obtained operating state data.

[0007] An air conditioner control system according to one aspect of this disclosure comprises an air conditioner, an input interface for inputting data, and a control unit that controls an inference process that inputs prompts to a large-scale language model and generates output sentences using the large-scale language model. The large-scale language model is trained by machine learning to generate output sentences based on input prompts. The control unit obtains text via the input interface that indicates a user's request to the air conditioner and is included in the prompt to be input to the large-scale language model. The control unit obtains operating state data indicating the operating state of the air conditioner based on an output sentence generated by the inference process from a first prompt containing the obtained text. The control unit controls the air conditioner to change its operating state based on an output sentence generated by the inference process from a second prompt containing the obtained operating state data.

[0008] A control method for an air conditioner according to one aspect of the present disclosure includes the steps of: a control unit that controls an inference process that inputs a prompt to a large language model and generates an output sentence by the large language model, acquiring text that indicates a user's request to the air conditioner and is included in the prompt and input to the large language model; acquiring operating state data indicating the operating state of the air conditioner based on an output sentence generated by the inference process from a first prompt including the acquired text; and controlling the air conditioner to change the operating state of the air conditioner based on an output sentence generated by the inference process from a second prompt including the acquired operating state data.

[0009] According to the control device, control system, and control method disclosed herein, it is possible to easily control an air conditioner according to the user's requests and operating conditions.

[0010] A block diagram showing the configuration of a control system according to Embodiment 1 of this disclosure. A block diagram showing the configuration of a server in the control system. A block diagram showing the configuration of a terminal device in the control system. A diagram for explaining the functional configuration of a server in the control system of Embodiment 1. A diagram for explaining the operation of a control system. A flowchart illustrating the operation of a server in Embodiment 1. A diagram for explaining the language processing unit in the server of Embodiment 1. A diagram for explaining the prompts of a large-scale language model in the control system. A diagram for explaining the language processing unit of a server in a modified example of Embodiment 1. A diagram for explaining the agent prompt in a modified example of Embodiment 1. A diagram for explaining the search prompt in a modified example of Embodiment 1. A diagram for explaining the control prompt in a modified example of Embodiment 1. A flowchart illustrating the operation of a server in a modified example of Embodiment 1. A diagram for explaining the server in the control system of Embodiment 2. A diagram for explaining the server in the control system of Embodiment 3. A block diagram showing the configuration of a control system according to a further modified example of Embodiment 1.

[0011] The embodiments of this disclosure will be described below with reference to the drawings.

[0012] (Embodiment 1) 1. Diagram 1 is a block diagram showing the configuration of an air conditioner control system 1 according to Embodiment 1 of the present disclosure. The control system 1 of this embodiment includes an air conditioner 10, a server 8, and a terminal device 7. In this system 1, the air conditioner 10, the server 8, and the terminal device 7 are connected to each other so as to be able to communicate data. This system 1 can be applied, for example, to the control of the air conditioner 10 by text input by a user as a request to the air conditioner 10. In this system 1, for example, when text is input by the terminal device 7, the air conditioner 10 is controlled according to a control command to the air conditioner 10 generated by the server 8 in accordance with the text. The server 8 of this embodiment is an example of a control device for the air conditioner 10.

[0013] 1-1. The air conditioner 10 illustrated in diagram 1 includes a storage unit 11, a control unit 12, a communication unit 13, an indoor temperature sensor 14, and an indoor humidity sensor 15. The air conditioner 10 may further include various sensors, such as an outdoor temperature sensor, to perform various functions. The air conditioner 10 may also include a display for displaying visual information to the user.

[0014] The air conditioner 10 includes an indoor unit 20 and an outdoor unit 40. A storage unit 11, a control unit 12, a communication unit 13, an indoor temperature sensor 14, and an indoor humidity sensor 15 may be provided in the indoor unit 20. The air conditioner 10 further includes a ventilation device 17, which enables supply ventilation by bringing outdoor air into the room and / or exhaust ventilation by discharging indoor air to the outside.

[0015] The indoor unit 20 is equipped with an indoor heat exchanger 22 that exchanges heat with indoor air, and a fan 24 that draws indoor air into the indoor unit 20 and blows out the indoor air into the room after heat exchange with the indoor heat exchanger 22.

[0016] The outdoor unit 40 is equipped with an outdoor heat exchanger 42 that exchanges heat with outdoor air, and a fan 44 that draws outdoor air into the outdoor unit 40 and blows out the outdoor air to the outside after it has exchanged heat with the outdoor heat exchanger 42. The outdoor unit 40 is also equipped with an indoor heat exchanger 22, an outdoor heat exchanger 42, a compressor 46, an expansion valve 48, and a four-way valve 50 that execute the refrigeration cycle.

[0017] The indoor heat exchanger 22, outdoor heat exchanger 42, compressor 46, expansion valve 48, and four-way valve 50 are each connected by refrigerant piping through which the refrigerant flows. In cooling and dehumidifying (weak cooling) operation, the air conditioner 10 performs a refrigeration cycle in which the refrigerant flows sequentially from the compressor 46 through the four-way valve 50, outdoor heat exchanger 42, expansion valve 48, and indoor heat exchanger 22 before returning to the compressor 46. In heating operation, the air conditioner 10 performs a refrigeration cycle in which the refrigerant flows sequentially from the compressor 46 through the four-way valve 50, indoor heat exchanger 22, expansion valve 48, and outdoor heat exchanger 42 before returning to the compressor 46.

[0018] In this embodiment, the ventilation device 17 is installed outdoors together with the outdoor unit 40. In addition to its ventilation function, the ventilation device 17 can dehumidify or humidify the indoor air in the control space by supplying, for example, dehumidified outdoor air or outdoor air containing moisture to the control space.

[0019] The air conditioner 10 has, for example, a cooling function, a heating function, a dehumidifying function, an air cleaning function, and / or a ventilation function by a ventilation device 17. The air conditioner 10 may also have a dehumidifying function and / or a humidifying function using the ventilation device 17. The air conditioning operation performed by the air conditioner 10 can be selected by the user. These functions and the operating modes that perform the various functions can be freely combined (for example, a cooling dehumidification function, a heating humidification function, a cooling ventilation mode, etc.). For example, in cooling mode, the air conditioner 10 can perform a weak cooling operation (also called compressor-type dehumidification or refrigeration cycle dehumidification) and a dehumidification operation by throttling the expansion valve 48, in addition to the cooling operation.

[0020] The storage unit 11 is a recording medium for recording various information and control programs, and may also be a memory that functions as a work area for the control unit 12. The storage unit 11 can be implemented as, for example, flash memory, RAM (Random Access Memory), ROM (Read Only Memory), other storage devices, or a combination thereof as appropriate.

[0021] The storage unit 11 may store references or thresholds used for various controls of the air conditioner 10. The storage unit 11 may also store information acquired from various sensors, such as the indoor temperature sensor 14. Information acquired from the terminal device 7, the server 8, or an external information source outside the control system 1 may also be stored in the storage unit 11. The information acquired in this way can be read by the control unit 12 when controlling the air conditioner 10.

[0022] The storage unit 11 may store a computer program (sometimes abbreviated as "program" in this disclosure) that causes the air conditioner 10 to execute the control method in this embodiment. The storage unit 11 may also include a non-temporary computer-readable storage medium on which the computer program is stored.

[0023] The control unit 12 is a controller that is responsible for controlling at least some of the functions of the air conditioner 10. The control unit 12 includes a general-purpose processor such as a CPU, MPU, MCU, FPGA, DSP, or ASIC that realizes predetermined functions by executing a program. The control unit 12 can realize various controls in the air conditioner 10 by calling and executing a control program stored in the storage unit 11. The control unit 12 can also cooperate with the storage unit 11 to read and write data stored in the storage unit 11. The control unit 12 is not limited to a processor that realizes predetermined functions through the cooperation of hardware and software, but may also be a hardware circuit specifically designed to realize predetermined functions.

[0024] The control unit 12 receives various control commands (for example, activation commands for various operating modes of the air conditioner 10, and / or temperature setting commands related to heating and cooling control) from the server 8, etc., via the communication unit 13. Based on the set values ​​corresponding to these control commands and the detected values ​​received from various sensors (for example, indoor humidity and / or outdoor humidity), the control unit 12 controls each component of the air conditioner 10 so that the air conditioner 10 performs its cooling, heating, and ventilation functions. The set values ​​indicate, for example, target values ​​for controlling the operating mode and / or temperature of the air conditioner 10. The set values ​​are stored in the storage unit 11 and updated by the control commands.

[0025] The communication unit 13 communicates with the air conditioner 10 and external devices such as the terminal device 7 and the server 8, or external information sources, in accordance with a predetermined standard to send and receive data. The predetermined standard includes, but is not limited to, Wi-Fi®, IEEE 802.2, IEEE 802.3, 3G, 4G, LTE, intranet, extranet, LAN, ISDN, VAN, CATV network, virtual private network, telephone line network, mobile communication network, satellite communication network, infrared and / or Bluetooth®. For example, the control unit 12 can cooperate with the server 8 and / or terminal device 7 via the communication unit 13.

[0026] The indoor temperature sensor 14 detects the temperature of the indoor air drawn into the indoor unit 20 from the control space. The indoor humidity sensor 15 detects the humidity of the indoor air. In one embodiment, each sensor 14 and 15 is installed in the indoor unit 20 at the air intake port that draws in indoor air. The information detected by each sensor 14 and 15 is stored in the storage unit 11 and later used by the control unit 12 or transmitted to the terminal device 7 or server 8.

[0027] In addition to the indoor temperature sensor 14 and the indoor humidity sensor 15, the air conditioner 10 may be equipped with various sensors to acquire various information from outside the air conditioner 10 in order to perform its function. For example, the air conditioner 10 may include an outside air temperature sensor that detects the outside air temperature of the controlled space. These sensors can be used to acquire various information necessary for operating the air conditioner 10.

[0028] 1-2. Server Configuration Diagram 2 is a block diagram showing the configuration of Server 8 in Control System 1. Server 8 is composed of an information processing device, such as a computer. The Server 8 illustrated in Figure 2 includes a control unit 80, a storage unit 81, and a communication unit 82.

[0029] The control unit 80 includes, for example, a CPU or MPU that works in cooperation with software to realize predetermined functions. The control unit 80 controls the overall operation of the server 8 and executes an information processing method in the control system 1, such as the control method of this embodiment. The control unit 80 reads data and programs stored in the storage unit 81, performs various calculations, and realizes various functions.

[0030] The control unit 80 executes a control program that includes, for example, a set of instructions for realizing each of the above functions. The control program may be provided from a communication network such as the Internet and stored in the memory unit 81, or it may be stored on a portable recording medium. The control unit 80 may be a dedicated electronic circuit or a hardware circuit such as a reconfigurable electronic circuit designed to realize each of the above functions. The control unit 80 may be composed of various semiconductor integrated circuits such as a CPU, MPU, GPU, GPGPU, TPU, microcontroller, DSP, FPGA, and ASIC.

[0031] The storage unit 81 is a storage medium that stores programs and data for realizing the functions of the server 8. The storage unit 81 includes, for example, an HDD or an SSD. The storage unit 81 stores, for example, various databases (DBs) described later.

[0032] The storage unit 81 includes, for example, RAM such as DRAM or SRAM, and temporarily stores (i.e., holds) data. The storage unit 81 may function as a work area for the control unit 80, or it may include a storage area in the internal memory of the control unit 80.

[0033] The communication unit 82 is a module (circuit) that connects to an external device according to a predetermined communication standard in wired or wireless communication. The predetermined communication standard includes, for example, USB, HDMI®, IEEE 802.11, Wi-Fi, and Bluetooth. The communication unit 82 may also connect to a server 8 on a communication network such as the Internet. The communication unit 82 is an example of an acquisition unit that receives various information from an external device or a communication network. The communication unit 82 may also constitute an input interface for inputting various information and / or an output interface for outputting various information in the control system 1.

[0034] In server 8, the acquisition of various information is not limited to the communication unit 82, but may also be achieved through cooperation with various software in the control unit 80, etc. The control unit 80 may acquire various information by reading the information stored in various storage media (for example, storage unit 81) into the work area of ​​the control unit 80.

[0035] 1-3. Diagram 3 of the terminal device configuration is a block diagram showing the configuration of the terminal device 7 in the control system 1. The terminal device 7 is composed of an electronic device such as a smartphone, tablet, or PC (personal computer). The terminal device 7 illustrated in Figure 3 includes a control unit 70, a storage unit 71, a communication unit 72, an operation unit 73, a display unit 74, a microphone 75, and a speaker 76.

[0036] The control unit 70 includes, for example, a CPU or MPU that works in cooperation with software to realize predetermined functions. The control unit 70 controls, for example, the overall operation of the terminal device 7. The control unit 70 reads data and programs stored in the storage unit 71, performs various calculations, and realizes various functions.

[0037] The control unit 70 executes a program that includes a set of instructions for realizing each of the above functions. This program may be provided via a communication network such as the Internet, or it may be stored on a portable recording medium. The control unit 70 may also be a dedicated electronic circuit or a hardware circuit such as a reconfigurable electronic circuit designed to realize each of the above functions. The control unit 70 may be composed of various semiconductor integrated circuits such as a CPU, MPU, GPU, GPGPU, TPU, microcontroller, DSP, FPGA, and ASIC.

[0038] The storage unit 71 is a storage medium that stores programs and data necessary to realize the functions of the terminal device 7. The storage unit 71 includes, for example, an HDD or SSD, and stores the above-mentioned programs and various data. The storage unit 71 also includes, for example, RAM such as DRAM or SRAM, and temporarily stores (i.e., holds) data. The storage unit 71 may function as a work area for the control unit 70, or it may include a storage area in the internal memory of the control unit 70.

[0039] The communication unit 72 is a module (circuit) that connects to an external device according to a predetermined communication standard in wired or wireless communication. The predetermined communication standard includes, for example, USB, HDMI, IEEE 802.11, Wi-Fi, and Bluetooth. The communication unit 72 may also connect to a communication network such as the Internet, or to a terminal device 7.

[0040] The operation unit 73 is a general term for the operating components that the user operates. The operation unit 73 may include, for example, a touch panel superimposed on the display unit 74 for inputting various touch operations. The operation unit 73 may also include physical buttons or switches provided on the terminal device 7, or it may include a connection unit that communicates with an external input device to receive operation signals. A keyboard, mouse, or touchpad may be used as the input device. The operation unit 73 may also include various GUIs such as virtual buttons, icons, cursors, software keyboards, and objects displayed on the display unit 74. The operation unit 73 may constitute the input interface of the control system 1.

[0041] The display unit 74 is composed of, for example, a liquid crystal display or an organic EL display. The display unit 74 may display various types of information, such as various GUIs for operating the operation unit 73 and information input from the operation unit 73. When using an external display device outside the terminal device 7, for example, the display unit 74 may be an output interface circuit for video signals and the like compliant with the HDMI standard or the like.

[0042] The microphone 75 includes, for example, one or more microphone elements built in the terminal device 7. The microphone 75 generates voice data based on a voice signal indicating the collected voice and outputs the voice data to the control unit 70. The microphone 75 is an example of an input interface for inputting voice data. The terminal device 7 may include a connection part such as a terminal for connecting an external microphone instead of or in addition to the built-in microphone 75.

[0043] The speaker 76 includes, for example, one or more speaker elements built in the digital camera 100 and outputs voice to the outside of the terminal device 7 under the control from the control unit 70. The terminal device 7 may include a connection part for connecting an external speaker or earphone instead of or in addition to the built-in speaker 76. In the control system 1, the speaker 76 is an example of an output interface for presenting information to the user. In the control system 1, the microphone and / or the speaker may be provided not only in the terminal device 7 but also, for example, in the air conditioner 10.

[0044] The configurations of the air conditioner 10, the server 8, and the terminal device 7 as described above are examples, and these configurations are not limited to the above examples. For example, each of the control units 12, 80, 70 may include two or more processors. Also, different data and programs may be stored in each of the storage units 11, 81, 71 from the above examples. The control method of the present embodiment may be executed by any one of the control units 12, 80, 70 or may be executed by a collaborative operation. Also, the control method of the present embodiment may be executed in distributed computing.

[0045] 1-4. Functional Configuration of the Server The configuration for generating control commands to the air conditioner 10 in the server 8 of this system 1 in response to text input by the user from the terminal device 7 will be explained with reference to Figure 4. Figure 4 is a diagram illustrating the functional configuration of the server 8 in the control system 1 of this embodiment. The server 8 of this embodiment includes, for example, a language processing unit 83, a search execution unit 84, and a control execution unit 85 as the functional configuration of the control unit 80. This functional configuration is realized, for example, by various programs stored in the storage unit 81. Furthermore, the storage unit 81 in the server 8 of this embodiment stores the user DB 86 and the equipment DB 87.

[0046] The language processing unit 83 of the control unit 80 controls the processing in the search execution unit 84 and the control execution unit 85 by performing various natural language processing operations, such as inference processing that generates output sentences from prompts using a large-scale language model (LLM). The LLM is acquired as a pre-trained model that generates output sentences based on input prompts by machine learning using relatively large-scale training data. The LLM is implemented using, for example, a sequence transformation model that includes a self-attention mechanism, and may include various parameters and programs.

[0047] In the server 8 of this embodiment, the control unit 80, which functions as a language processing unit 83, sends a prompt to an external information processing device that stores the LLM, for example via the communication unit 82, and obtains the output sentence generated by the LLM's inference processing based on the prompt from that device. The external information processing device is, for example, a server device in cloud computing. In this embodiment, the control unit 80 of the server 8 inputs a prompt to the LLM and obtains the output sentence by controlling the LLM's inference processing in the external server device, etc., via the communication network. In the control system 1, the storage location of the LLM is not limited to the above example, and the LLM may be stored in the storage unit 81 of the server 8, for example.

[0048] The language processing unit 83 acquires a request message as text indicating the user's request from the terminal device 7 through data communication via the communication unit 82, for example. Also, the language processing unit 83 acquires user information for identifying the user, for example, from the terminal device 7 in association with the request message. The user information may be an identifier in the data communication between the terminal device 7 and the server 8. The language processing unit 83 acquires device information of the air conditioner 10 associated with the user identified by the user information from the user DB 86 that manages and associates the air conditioner 10 used by each user for a plurality of users and a plurality of air conditioners 10, for example. The device information includes, for example, a device ID or the like as information for identifying the air conditioner 10.

[0049] The search execution unit 84 executes a data search based on a query generated by the language processing unit 83, for example. The search execution unit 84 acquires device operation data corresponding to the query from the device DB 87, for example. The device operation data includes various data indicating the operation state of the air conditioner 10, and includes, for example, setting values such as the on / off of the power supply in the air conditioner 10, the operation mode, temperature, air volume, and air direction in the air conditioning operation. When the air conditioner 10 has a timer function, the device operation data may include the setting value of the timer. In the server 8, the device DB 87 accumulates the device operation data sequentially received from the air conditioner 10 via the communication unit 82, for example.

[0050] The control execution unit 85 generates a control command for the air conditioner 10 based on a control parameter generated by the language processing unit 83, for example. The control parameter is generated to include, for example, the setting value changed in the device operation data and the changed value. The control execution unit 85 transmits a control command to the air conditioner 10 via the communication unit 82 and receives a response indicating a control result such as success or failure of the control according to the control command from the air conditioner 10, for example. The function of the control execution unit 85 can be implemented by, for example, an API (Application Programming Interface) for controlling the air conditioner 10.

[0051] Furthermore, the language processing unit 83 generates a response message indicating an answer to the user's request message based on the equipment operation data and response acquired by the search execution unit 84 and the control execution unit 85, respectively. The generated response message is transmitted to the terminal device 7 via the communication unit 82.

[0052] 2. Operation The operation of the control system 1 of this embodiment, which is configured as described above, will be explained below.

[0053] 2-1. Overview of Operation Diagram 5 is a diagram illustrating the operation of the control system 1. This system 1 enables air conditioning control in response to natural language requests from the user by making the LLM function like an autonomous agent through prompting.

[0054] First, the system 1 receives a request message entered by the user via the terminal device 7. In the example shown in Figure 5, the text "hot" is entered as the user's request message. For example, the terminal device 7 receives this user input via the operation unit 73 through an application for controlling the air conditioner 10. The terminal device 7 then sends the text data representing this text as a request message to the server 8.

[0055] This system 1 utilizes a technology that combines simulated thinking and decision-making through LLM inference processing, such as disclosed in Non-Patent Document 1, and the server 8 executes a process that causes the LLM to repeatedly perform these simulated thinking and decision-making processes.

[0056] In server 8, the control unit 80, acting as a language processing unit 83, inputs prompts, including request messages, to the LLM, and as a result of simulated thinking through the LLM's inference process, determines the action to take as an autonomous agent in response to the request message. In this system 1, the actions selected by the inference process are defined by the prompt description, such as searching the equipment DB 87 and controlling the air conditioner 10. For example, the language processing unit 83, through inference processing, decides to first understand the operating status of the air conditioner 10 by obtaining equipment operation data by searching the equipment DB 87. In this system 1, prompts are created in advance to allow the LLM to simulate such thinking and actions, and are stored, for example, in the storage unit 81 of server 8. In the example in Figure 5, the power on / off, operating mode, and temperature settings of the air conditioner 10 are obtained from the equipment DB 87 as equipment operation data.

[0057] The language processing unit 83 of server 8 performs further inference processing using LLM based on a prompt that includes equipment operation data obtained from the equipment DB 87 in addition to the request message, and determines the next action. In the example in Figure 5, when the equipment operation data observed as a result of searching the equipment DB 87 indicates that the power is on, the operating mode is cooling operation, and the set temperature is 26°C, it is decided that the next action is to lower the set temperature to 25°C in cooling operation by controlling the air conditioner 10. Server 8 generates a control command to the air conditioner 10 based on the control parameters determined by the language processing unit 83, for example, to realize this control of the air conditioner 10.

[0058] Server 8 transmits a control command to the air conditioner 10 and receives a response from the air conditioner 10 indicating the control result. Furthermore, based on prompts that include, for example, the history of inputs to and outputs from the LLM in addition to the response, Server 8 uses the language processing unit 83 to generate a response message indicating the control content of the air conditioner 10 executed in response to, for example, the user's request message. The system 1 displays the response message received from Server 8 on the display unit 74 of the terminal device 7.

[0059] As described above, System 1 controls the inference process by LLM on the server 8 based on prompts to perform data retrieval from the equipment DB 87 and determine the control parameters of the air conditioner 10. This allows the LLM to obtain equipment operation data from the equipment DB 87 to understand the operating status of the air conditioner 10, even if the user's request message in natural language is ambiguous, and to determine the control parameters of the air conditioner 10 according to the operating status. With System 1, the air conditioner 10 can be flexibly controlled according to its current operating status in order to meet the user's requests.

[0060] 2-2. Server Operation The operation of Server 8 in System 1 will be explained using Figures 6 to 8.

[0061] Figure 6 is a flowchart illustrating the operation of the server 8 in Embodiment 1. The process shown in Figure 6 is initiated, for example, when a communication connection is established between the server 8 and the terminal device 7 via the communication units 82 and 72 in the control system 1. Each process shown in the flowchart of Figure 6 is executed, for example, by the control unit 80 of the server 8.

[0062] First, the control unit 80, for example as a language processing unit 83, receives the user's request message transmitted from the terminal device 7 via the communication units 72 and 82 (S1). For example, the terminal device 7 obtains text data indicating the request message through text input in the operation unit 73.

[0063] Figure 7 is a diagram illustrating the language processing unit 83 in the server 8 of Embodiment 1. In this embodiment, for example, in step S1, the language processing unit 83 acquires user information that identifies the user in association with the user's request message from the terminal device 7.

[0064] The control unit 80 obtains equipment information for the air conditioner 10 corresponding to the user in the user information by referring to the user DB 86 in the storage unit 81 (S2).

[0065] The control unit 80, acting as the language processing unit 83, includes the request message acquired in step S1 and the device information acquired in step S2 in the prompt, and inputs the prompt to the LLM (S3).

[0066] The language processing unit 83 of this embodiment further includes, for example, an agent unit 30 and a distribution mechanism 34 as functional components, as shown in Figure 7. The agent unit 30 comprehensively controls various processes in the language processing unit 83. The agent unit 30 illustrated in Figure 7 includes an LLM execution unit 61, a prompt acquisition unit 62, and a post-processing mechanism 63.

[0067] The LLM execution unit 61 causes the LLM to perform inference processing based on the prompt obtained by the prompt acquisition unit 62 and obtains the output statement generated by the inference processing. The prompt acquisition unit 62 obtains the prompt to be input to the LLM by the LLM execution unit 61 by reading a prompt stored in, for example, the storage unit 81 of the server 8 and embedding a request message and device information, etc., into the prompt. The post-processing mechanism 63 parses the output statement from the LLM obtained by the LLM execution unit 61 and extracts, for example, predetermined values ​​included in the output statement.

[0068] The distribution mechanism 34 distributes the processing to be performed by the control unit 80 to either the search execution unit 84 or the control execution unit 85, based on the parsing results of the output statement from the LLM by the post-processing mechanism 63. Depending on whether the processing determined as the next action in the output statement is data retrieval or control of the air conditioner 10, the distribution mechanism 34 inputs the query contained in the output statement to the search execution unit 84 or inputs the control parameters contained in the output statement to the control execution unit 85.

[0069] In step S3, the language processing unit 83, acting as a prompt acquisition unit 62, acquires a prompt including a request message, and, acting as an LLM execution unit 61, inputs the prompt to the LLM, thereby acquiring an output statement generated from the prompt through inference processing.

[0070] Figure 8 is a diagram illustrating prompts in the LLM in the control system 1. Figure 8 illustrates prompt D1 input to the LLM by the prompt acquisition unit 62. For example, as shown in Figure 8, prompt D1 includes control information F1 that associates the user's request with the control of the air conditioner 10, in addition to a description that teaches the LLM its role as an agent unit 30. Furthermore, prompt D1 in this embodiment includes explanatory texts E1 and E2 that teach the LLM the processing performed by the search execution unit 84 and the processing performed by the control execution unit 85, respectively, as actions that can be executed by the agent unit 30.

[0071] At prompt D1, the output format of the output statement generated by the LLM's inference process is specified, for example, as shown as output format information F2 in Figure 8. By specifying this output format, the LLM can be made to simulate the actions and thoughts of an autonomous agent. Furthermore, prompt D1 shown in Figure 8 is provided with an input field C1 in which a request message is embedded and an input field C2 in which device information is embedded. For example, by writing "Thought" at the end of prompt D1, which specifies that thoughts should be output in the output format, it is possible to explicitly tell the LLM to start generating an output statement from the results of the thought simulation.

[0072] For example, in Figure 8, prompt D1 can cause the LLM to generate an output statement that includes the selection result from the action options shown in "Action" of the output format information F2, and the result of that action shown in "Observation". The action options include, for example, data retrieval by the search execution unit 84 and control of the air conditioner 10 by the control execution unit 85. The action selection result and the result of the selected action are extracted as predetermined values ​​by, for example, the post-processing mechanism 63 of the language processing unit 83.

[0073] The control unit 80, acting as a post-processing mechanism 63, parses the output statement from the LLM and, for example, depending on whether the action selection result in the output statement is a data search, inputs the query included in the output statement as a result of that action to the search execution unit 84 as a distribution mechanism 34 (S4). The control unit 80, acting as a search execution unit 84, searches the equipment operation data corresponding to the query from the equipment DB 87 and obtains the search results (S4). In the example in Figure 7, the equipment operation data of the search results is input from the search execution unit 84 to the agent unit 30.

[0074] The control unit 80 acquires a new prompt that includes the past prompts and corresponding output statements input to the LLM in step S3 as input / output history, and further includes the search results from step S4, and inputs this prompt to the LLM (S5). The control unit 80, for example as a prompt acquisition unit 62, generates a prompt that adds the search results as "Observation" in the output format specified by prompt D1 in Figure 8, and "Thought" in the same way as in the example in Figure 8, to the past input / output history. In step S5, the control unit 80, as an LLM execution unit 61, inputs this prompt including the past input / output history and the search results of the equipment operation data to the LLM, and acquires the output statement generated by inference processing from this prompt.

[0075] The control unit 80, for example, depending on whether the selected action in the parsed output statement, which is processed by the post-processing mechanism 63, is to control the air conditioner 10, inputs the control parameters generated in the output statement as a result of that action to the control execution unit 85 via the distribution mechanism 34 (S6). The control unit 80, as the control execution unit 85, transmits the control command generated based on the control parameters to the air conditioner 10 via the communication unit 82 (S6). The control unit 80 causes the air conditioner 10 to change its operating state based on this control command. The control unit 80, as the control execution unit 85, receives a response from the air conditioner 10 via the communication unit 82 indicating the control result corresponding to the control command, and inputs it to the agent unit 30, for example.

[0076] The control unit 80 obtains a prompt that includes, for example, past prompts and corresponding output statements input to the LLM in step S5 as input / output history, and the control result from step S6, and inputs this prompt to the LLM (S7). The control unit 80 generates a prompt that adds the response from the air conditioner 10 as "Observation" in the output format illustrated in Figure 8, and "Thought," to the past input / output history, for example, in the same way as in step S5. In step S7, the control unit 80 obtains an output statement generated by inference processing from the prompt that includes this past input / output history and the control result from the air conditioner 10.

[0077] The control unit 80, for example, uses the sorting mechanism 34 to determine whether the output text parsed by the post-processing mechanism 63 contains a response message to the user (S8). Based on the output text, the sorting mechanism 34 determines whether the response message is output as a final response to the user's request in the output format "Final Answer" shown in Figure 8.

[0078] If the response message is not included in the output statement (NO in S8), the control unit 80 inputs a prompt to the LLM that includes the search results and / or control results in addition to the past input / output history, similar to steps S5 and S7 (S9). The control unit 80 repeats the processing in steps S4 to S9 according to the action selection result in the output statement obtained from the prompt by the inference processing, or the response message. If the action selection result in the output statement from the prompt input in step S5 is a data search, the control unit 80 may repeat the processing from step S4 onwards before executing step S6 to search for the equipment operation data again based on the output statement.

[0079] If the response message is included in the output text (YES in S8), the control unit 80 inputs the response message to the agent unit 30 as the final response, for example, using the distribution mechanism 34, and transmits the response message to the terminal device 7 via the communication unit 82 (S10).

[0080] According to the above process, a prompt containing the user's request message is input to the LLM (S3), and based on the output statement generated by the LLM's inference process, the equipment operation data of the air conditioner 10 related to the request message is searched (S4). Then, a prompt containing the past input / output history and the search results of the equipment operation data is input to the LLM (S5), and based on the output statement from the inference process, a control command generated according to the request message and the search results is sent to the air conditioner 10 (S6). In this way, a control command can be generated by referencing the operating state of the air conditioner 10 through a search in response to the user's request indicated by the request message, and even if the request is ambiguous, specific control can be performed according to the operating state of the air conditioner 10.

[0081] As described above, the operation of the server 8 allows for the provision of a control system 1 that combines inference processing by LLM with the retrieval of equipment operation data to determine the control of the air conditioner 10 according to the relevant operating conditions, even for ambiguous user requests. For example, with this system 1, based on a user's request of "it's hot," the system can retrieve the operating conditions of the air conditioner 10, such as power on / off, operating mode, and set temperature, and perform control such as lowering the set temperature based on the search results. This reduces the user's mental burden in controlling the operating conditions of the air conditioner 10, as they only need to input a request corresponding to their own condition into the system 1, without having to think about how to change the settings for air conditioning operation.

[0082] Furthermore, according to this system 1, the instructions a user gives to control the air conditioner 10 are not limited to predetermined control commands such as pre-registered words, but can be input using various expressions in natural language. In addition, control can be performed according to the operating data of the air conditioner 10, which is expected to improve user comfort. Moreover, by utilizing functions or settings of the air conditioner 10 that may be unfamiliar to the user in the control, for example, it is expected that efficient operation of the air conditioner 10 can be achieved and energy consumption by the air conditioner 10 can be optimized. For example, the prompt D1 in Figure 8 may include the operation manual for the air conditioner 10, and control information F1 may be described from an energy-saving perspective.

[0083] In this embodiment, until a response to the user's request is obtained (S8), the LLM is made to select an action in the output sentence through inference processing, and the result of the action is input to the LLM as feedback, so that the next output sentence is generated (S4 to S9). This makes it possible to determine the processing to be executed as an action in stages according to the user's request, for example. Such processing as actions is not limited to the order exemplified in steps S4 to S7, but for example, depending on the action selection result in the output sentence from the LLM, the processing of data retrieval and control of the air conditioner 10 can be selectively executed as actions.

[0084] 2-3. Multi-Agent In the above, an example was described in which the control unit 80 of the server 8 acts as an agent unit 30 and executes functions other than the distribution mechanism 34 in the language processing unit 83 based on one type of prompt D1 as illustrated in Figure 8 (see Figure 7). The various functions of the language processing unit 83 may be realized, for example, by the coordination of processes in which multiple LLM agents perform processing according to each prompt based on multiple types of prompts. Modifications of Embodiment 1, which is configured with such a multi-agent, will be described using Figures 9 to 13.

[0085] 2-3-1. Diagram 9 of the Language Processing Unit and Prompt Configuration is a diagram illustrating the Language Processing Unit 83A of the Server 8 in a modified example of Embodiment 1. In this modified example, the Language Processing Unit 83A includes an Agent Unit 30A, a Search Processing Unit 31, and a Control Processing Unit 32 that function in cooperation as a multi-agent, instead of the Agent Unit 30 shown in Figure 7. The Agent Unit 30A and each of the Processing Units 31 and 32 each include an LLM execution unit 61 and a post-processing mechanism 63, similar to the Agent Unit 30 in Figure 7. In this modification, the Agent Unit 30A includes an Agent Prompt Acquisition Unit 62a instead of the Prompt Acquisition Unit 62, the Search Processing Unit 31 includes a Search Prompt Acquisition Unit 62b, and the Control Processing Unit 32 includes a Control Prompt Acquisition Unit 62c.

[0086] For example, each prompt acquisition unit 62a, 62b, and 62c acquires a prompt to be input to the LLM based on a different type of prompt read from the storage unit 81. The prompts read by each of these prompt acquisition units 62a to 62c are illustrated in Figures 10 to 12. Figures 10, 11, and 12 are diagrams illustrating the agent prompt D10, search prompt D11, and control prompt D12 in this modified example, respectively.

[0087] The agent prompt D10 shown in Figure 10 includes an explanatory text E3 for a subordinate agent that the LLM instructs to act as an agent unit 30A, instead of the explanatory texts E1 and E2 for each process performed by the search execution unit 84 and the control execution unit 85 in prompt D1 (Figure 8). For example, explanatory text E3 teaches the LLM the roles of each subordinate agent, which correspond to the search processing unit 31 specialized for the processing of the search execution unit 84, and the control processing unit 32 specialized for the processing of the control execution unit 85, respectively, as well as how to instruct each subordinate agent. In addition, the agent prompt D10 includes an explanatory text E4 in the output format from the LLM that specifies the use of a subordinate agent, instead of the description of the action in prompt D1 in Figure 8.

[0088] The search prompt D11 shown in Figure 11 includes a description for enabling LLM to function as a subordinate agent corresponding to the search processing unit 31. For example, the search prompt D11 includes an item of equipment operation data that can be obtained from the equipment DB 87 in the explanatory text E11 which shows the data structure of the equipment DB. The items in the explanatory text E11 can be changed as appropriate to match the search target in the equipment operation data. The search prompt D11 includes an input field C3 into which instructions output from the agent unit 30A are input as a higher-level agent.

[0089] The control prompt D12 shown in Figure 12 includes a description for enabling LLM to function as a subordinate agent corresponding to the control processing unit 32. For example, the control prompt D12 includes an explanatory text E12 for the API that controls the operating state of the air conditioner 10. Items in the explanatory text E12 that indicate the operating state to be changed by the API can be changed as appropriate to match the operating state of the controlled object. The control prompt D12 includes an input field C4 into which instructions from the agent unit 30A are input as a higher-level agent.

[0090] 2-3-2. Server Operation Diagram 13 is a flowchart illustrating the operation of server 8 in this modified example. In this modified example, instead of the processing corresponding to prompt D1 in Figure 8 (S3, S4, S6 in Figure 6), the processing corresponding to prompts D10 to D12 in Figures 10 to 12 (S3A, S4A to S4B, S6A to S6B) is executed.

[0091] After the control unit 80 of the server 8 obtains a request message from the user and device information corresponding to that user (S1, S2), it inputs an agent prompt D10 containing the request message and device information to the LLM as the agent unit 30A (S3A). For example, in the agent unit 30A shown in Figure 9, the control unit 80, as the agent prompt acquisition unit 62a, acquires an agent prompt D10 (see Figure 10) with the request message and device information embedded in it.

[0092] The control unit 80, acting as the agent unit 30A, instructs the search processing unit 31 to generate a query to search the device DB 87, for example, in response to the output statement generated by the LLM from the input in step S3A, where the action selection result is a data search (S4A). For example, based on the agent prompt D10 input to the LLM in step S3A, an output statement including instructions to the search processing unit 31 in natural language is generated. These instructions, along with device information, are input to the search processing unit 31 from the agent unit 30A via the distribution mechanism 34, as shown in Figure 9, for example.

[0093] The control unit 80, acting as a search processing unit 31, generates a query based on the instructions generated in step S4A, in addition to the equipment information. In the search processing unit 31, the control unit 80, for example, as a search prompt acquisition unit 62b shown in Figure 9, acquires a search prompt D11 (see Figure 11) into which the instructions are embedded, and as an LLM execution unit 61, causes the LLM to generate a query corresponding to the instructions based on the search prompt D11. For example, the control unit 80, acting as a search processing unit 31, causes the search execution unit 84 to search for equipment operation data from the equipment DB 87 based on the generated query (S4B), and inputs the search results from the search execution unit 84 to the agent unit 30A.

[0094] The control unit 80, acting as, for example, the agent unit 30A, inputs a prompt to the LLM containing past input / output history and search results, similar to step S5 in Figure 6, and obtains the generated output statement. The control unit 80, acting as the agent unit 30A, instructs the control processing unit 32 to generate control parameters (S6A) depending on whether the action selection result in the output statement is control of the air conditioner 10. For example, in step S5, an output statement containing natural language instructions to the control processing unit 32 is generated from the agent prompt D10 included in the input / output history. These instructions, along with equipment information, are input to the control processing unit 32 from the agent unit 30A via the distribution mechanism 34, as shown in Figure 9, for example.

[0095] The control unit 80, acting as a control processing unit 32, generates control parameters for the air conditioner 10 based on the instructions input in step S6A as well as the equipment information. The control unit 80, acting as a control processing unit 32, acquires a control prompt D12 (see Figure 12) with the instructions embedded in it, acting as a control prompt acquisition unit 62c as illustrated in Figure 9, and causes the LLM execution unit 61 to generate a control command corresponding to the instructions based on the control prompt D12. For example, the control unit 80, acting as a control processing unit 32, transmits the generated control command to the air conditioner 10 via the control execution unit 85 (S6B), and inputs the response from the air conditioner 10 to the agent unit 30A.

[0096] Through the processing of server 8 as described above, the LLM inference process allows for the retrieval of equipment operation data for the air conditioner 10 as requested, and control can be performed according to the operating status of the air conditioner 10 (S3A to S6B).

[0097] Furthermore, the language processing unit 83A of this modified version has a multi-agent configuration as shown in Figure 9. In this modification, for example, instead of the search and control descriptions E1 and E2 in prompt D1 in Figure 8, the agent prompt D10 (Figure 10) contains a description E3 of the subordinate agent. For example, the descriptions E1 and E2 in prompt D1 are about the same length as the search prompt D11 (Figure 11) and the control prompt D12 (Figure 12), respectively. In this modified version, by providing the search prompt D11 and the control prompt D12, the agent prompt D10, which is repeatedly input to the LLM, can be shortened (see S3A, S5, S7). This can be expected to reduce the computational cost in the inference processing of the server 8 and the LLM, and to reduce noise in the output sentences caused by the lengthening of prompts.

[0098] Furthermore, with the multi-agent configuration described above, multiple LLMs optimized for each task, such as data retrieval and control, can be prepared, and one or more of them can be selectively used to realize the functions of the search processing unit 31 and the control processing unit 32. In this way, the inference processing in the agent unit 30A, the search processing unit 31, and the control processing unit 32 may be executed by different LLMs.

[0099] Furthermore, by providing a search prompt D11 and a control prompt D12 separately from the agent prompt D10, the length of the prompts can be suppressed even when including information that shows detailed specifications or schemas for the search and control processes, making it easier to extend each process. For example, even when the length of the prompt that can be entered into the LLM is limited by the number of characters, it can be easily extended. Moreover, the LLM into which each prompt D10 to D12 is entered does not have to be common; for example, by using an LLM specialized for the process corresponding to each prompt D10 to D12, an improvement in the overall processing performance of the language processing unit 83A can be expected. In this case, it is not necessary to use an LLM with general-purpose processing capabilities for each individual process, and various costs associated with arithmetic processing in the LLM can be reduced.

[0100] 3. Summary As described above, in this embodiment, the server 8 is an example of a control device for an air conditioner. The server 8 includes a communication unit 82 as an example of a communication interface for data communication, and a control unit 80 that controls an inference process that inputs a prompt to a large-scale language model (LLM) and generates an output sentence using the LLM. The LLM is trained by machine learning to generate an output sentence based on the input prompt. The control unit 80 acquires a request message via the communication unit 82 as an example of text indicating a user's request to the air conditioner 10, which is included in the prompt and input to the LLM (S1). Based on the output sentence generated by the inference process from the first prompt including the acquired request message, the control unit 80 acquires equipment operation data as an example of operation status data indicating the operating status of the air conditioner (S3-S4, S3A-S4B). Based on the output statement generated by inference processing from the second prompt which includes acquired equipment operation data, the control unit 80 transmits a control command to the air conditioner 10 via the communication unit 82 as an example of changing the operating state of the air conditioner 10 (S5-S6, S5-S6B).

[0101] According to the server 8 described above, the inference processing by LLM and the equipment operation data are combined to acquire equipment operation data according to the user's request, and the air conditioner 10 can be controlled according to the operating state indicated by the acquired data. This makes it easier to control the air conditioner 10 according to the user's request and operating state.

[0102] In this embodiment, as an example of the first and second prompts, each prompt D1, D10 (see Figures 8 and 10) includes an instruction (for example, output format information F2 for prompt D1) that causes the LLM to output an instruction statement (for example, a query or control command) that causes the control unit 80 to execute either data retrieval as an example of acquiring equipment operation data, or control of the air conditioner 10 as an example of changing the operating state. This allows the LLM's inference processing to generate an instruction statement corresponding to the process to be executed by the control unit 80 in its output statement.

[0103] In a modified example of this embodiment, agent prompts D10 (Figure 10), which are examples of first and second prompts, are configured to cause the LLM to output a command statement based on a lower-level prompt. The lower-level prompts include a search prompt D11 (Figure 11) as an example of a prompt related to acquiring equipment operation data, and a control prompt D12 (Figure 12) as an example of a prompt related to changing the operating state of the air conditioner 10. The control unit 80 selectively inputs the search prompt D11 and the control prompt D12 to the LLM based on the output statements generated from each agent prompt D10 as the first and second prompts (S4A, S6A). In such a multi-agent configuration, for example, it is possible to suppress the length of each prompt D10 to D12 and improve expandability.

[0104] In this embodiment, each of the first and second prompt examples, D1 and D10, includes control information F1 as shown in Figure 8, as an example of information that associates the user's request indicated by the request message with control to change the operating state of the air conditioner 10. According to this control information F1, for example, the inference processing of the LLM can determine control to change the operating state of the air conditioner 10 in response to the user's request.

[0105] In this embodiment, the control unit 80 retrieves equipment operation data in response to a request message by searching for equipment operation data stored in the equipment DB 87 as an example of data received from the air conditioner 10 via the communication unit 82 regarding the operating status (see S3-S4, S3A-S4B, Figures 4, 7, and 9). In this way, a search can be performed to retrieve equipment operation data related to the request message from the equipment DB 87 based on an output statement generated by the LLM's inference processing from a prompt including a request message. This allows, for example, the search results for the operating status according to the user's request to be included as input to the LLM (S5).

[0106] In this embodiment, the control unit 80 further includes the past input / output history in steps S3 and S3A, etc., as an example of the output statement generated from the first prompt in the second prompt and inputs it to the LLM (S5). This allows the LLM to perform inference processing in a way that simulates tracing the thought process according to the input / output history.

[0107] In this embodiment, the control unit 80 inputs a second prompt, an output statement generated from the second prompt, and a third prompt including an example of a control result resulting from changing the operating state of the air conditioner 10 in accordance with the output statement to the LLM (S7). Based on the output statement generated by the LLM's inference processing from the third prompt, the control unit 80 outputs a response message as an example of a response to the user's request indicated by the request message, or performs either acquiring equipment operation data or changing the operating state (S8, S4-S9, S4A-S9). As a result, while no response message is output to the user (NO in S8), selective execution of, for example, acquiring equipment operation data in accordance with the request message and controlling the air conditioner 10 in accordance with the current operating state indicated by the acquired equipment operation data is repeated.

[0108] The control system 1 in this embodiment includes an air conditioner 10, an input interface for inputting data (for example, the communication unit 72 and operation unit 73 of the terminal device 7, and the communication unit of the server 8), and a control unit 80 of the server 8 that controls the inference process of inputting prompts to a large-scale language model (LLM) and generating output sentences using the LLM. The LLM is trained by machine learning to generate output sentences based on the input prompts. In this system 1, the control unit 80 obtains a request message via the input interface, which is an example of text indicating a user's request to the air conditioner 10 and is included in the prompts and input to the LLM (S1). Based on the output sentence generated by the inference process from the first prompt including the acquired request message, the control unit 80 obtains equipment operation data, which is an example of operating status data indicating the operating status of the air conditioner 10 (S3-S4, S3A-S4B). The control unit 80 controls the air conditioner 10 to change its operating state based on an output statement generated by inference processing from a second prompt that includes acquired equipment operation data (S5-S6, S5-S6B).

[0109] In this embodiment, for example, the control system 1 described above provides a method for controlling the air conditioner 10. This method involves a control unit 80 that controls an inference process in which a prompt is input to a large-scale language model (LLM) and the LLM generates an output statement, acquiring a request message, which is an example of text indicating a user's request to the air conditioner 10 and is included in the prompt and input to the LLM (S1); acquiring equipment operation data, which is an example of operating state data indicating the operating state of the air conditioner 10, based on an output statement generated by the inference process from a first prompt including the acquired request message (S3-S4, S3A-S4B); and controlling the air conditioner to change the operating state of the air conditioner 10 based on an output statement generated by the inference process from a second prompt including the acquired equipment operation data (S5-S6, S5-S6B).

[0110] According to the control system 1 and control method for the air conditioner 10 described above, by combining inference processing by LLM and equipment operation data, it is possible to easily control the air conditioner 10 according to the user's requests regarding the air conditioner 10 and the operating status in the equipment operation data. The control system and control method disclosed herein are not limited to the control unit 80 of the server 8, but may also be executed by the control unit 12 of the air conditioner 10 or the control unit 70 of the terminal device 7, and may be executed by two or more combinations of each control unit 12, 70, 80.

[0111] (Embodiment 2) In Embodiment 1, a control system 1 was described in which a server 8 acquires equipment operation data of the air conditioner 10 in response to a request message from the user and controls the air conditioner 10 according to its operating state. In Embodiment 2, a control system 1 is described that further uses information about the external environment of the air conditioner 10 (for example, indoor air temperature and / or humidity) to control the air conditioner 10 in response to a request message.

[0112] Figure 14 is a diagram illustrating the server 8A in the control system 1 of Embodiment 2. In the server 8A of this embodiment, the storage unit 81 stores an environment DB 88 in addition to a user DB 86 and an equipment DB 87, similar to those in Figure 4. The environment DB 88 stores environment data that is sequentially received from the air conditioner 10 via, for example, the communication unit 82. The environment data includes information about the external environment of the air conditioner 10, such as the temperature and humidity of the indoor air detected by the indoor temperature sensor 14 and the indoor humidity sensor 15 of the air conditioner 10, respectively. The environment data is not limited to such indoor temperature and humidity, but may include various types of weather information.

[0113] In this embodiment, for example, the data search description E1 included in the prompt, similar to that in Figure 8, further describes the search for environmental data from the environmental DB 88 as an action option, in addition to the search for equipment operation data from the equipment DB 87.

[0114] In the server 8A of this embodiment, the control unit 80, as the language processing unit 83, inputs a prompt containing the user's request message to the LLM, for example, in the same manner as shown in Figure 6 (S3). Depending on the result of selecting an action in the output statement from the LLM, for example, if it is a data search of environmental data, the control unit 80, as the search execution unit 84, obtains the search result for environmental data from the environmental DB 88 based on the query included in the output statement, in the same manner as in step S4. The control unit 80 inputs a prompt containing the input / output history and the search result for environmental data to the LLM, in the same manner as in step S5, and transmits a control command generated based on the output statement from the LLM to the air conditioner 10 (S6).

[0115] Through the above process, in response to user requests, in addition to the operating status of the air conditioner 10 indicated by the equipment operation data, environmental data of the external environment can also be acquired, allowing, for example, the conditions inside and outside the air conditioner 10 to be reflected in the control.

[0116] The prompts in this embodiment are not limited to the above example, but may be configured similarly to the prompts D10 to D12 shown in Figures 10 to 12. For example, a prompt similar to the search prompt D11 in Figure 11 may include a description related to data retrieval from the environmental DB 88. Furthermore, the environmental data is not limited to information detected by the sensors 14 and 15 of the air conditioner 10, but may also include information detected by external devices. For example, the server 8A may acquire such information from external devices via a communication network such as the Internet.

[0117] As described above, in this embodiment, the control unit 80 of the server 8A further acquires environmental data, which is an example of data observed in the environment outside the air conditioner 10, based on the output statement generated by inference processing from a first prompt containing a user request message for the air conditioner 10 (see Figure 14). This makes it possible to acquire environmental data obtained in the external environment in addition to the equipment operation data of the air conditioner 10, in response to the request message. For example, based on the data acquired in this way, it becomes easier to perform control that is adapted not only to the operating state of the air conditioner 10 but also to the state of the external environment.

[0118] (Embodiment 3) In Embodiment 2, a control system 1 was described that uses environmental data of the external environment of the air conditioner 10 in addition to equipment operation data to control the air conditioner 10 in response to a user request message. In Embodiment 3, a control system 1 is described that further controls the air conditioner 10 according to the user's activity status (for example, physical activity level and / or heart rate).

[0119] Figure 15 is a diagram illustrating the server 8B in the control system 1 of Embodiment 3. In the server 8B of this embodiment, the storage unit 81 stores user status data indicating the user's activity status in the user DB 86, which is one of the DBs 86 to 88 similar to those in Figure 15. The user status data is acquired from the user by the terminal device 7, for example, and includes the user's physical activity level and heart rate.

[0120] In the control system 1 of this embodiment, the terminal device 7 acquires user status data measured by a wearable device attached to the user or by exercise equipment used by the user, for example, via a communication unit 72, and records it in a storage unit 71. The user DB 86 of this embodiment stores user status data that is sequentially received from the terminal device 7 via the communication unit 82.

[0121] In this embodiment, for example, the data search description included in the prompt, similar to that in Embodiment 2 (see E1 in Figure 8), further includes a description of searching for user status data from the user DB 86 as an action option, in addition to the equipment DB 87 and environment DB 88.

[0122] In the server 8B of this embodiment, the control unit 80, acting as the language processing unit 83, inputs a prompt containing the user's request message to the LLM, for example, similar to step S3 in Figure 6. Depending on whether the action selection result in the output statement from the LLM is a data search of user status data, the control unit 80, acting as the search execution unit 84, obtains the search result for user status data from the user DB 86 based on the query in the output statement, similar to step S4. The control unit 80 inputs a prompt containing the search result for user status data and input / output history to the LLM, similar to step S5, and transmits a control command generated based on the output statement from the LLM to the air conditioner 10 (S6).

[0123] Through the above processing, in response to user requests, user status data can be acquired in addition to the internal and external conditions of the air conditioner 10 indicated by, for example, equipment operation data and environmental data, so that the user's own condition, such as physical activity level and / or heart rate, can be reflected in the control of the air conditioner 10.

[0124] The prompts in this embodiment are not limited to the above example; for example, a prompt similar to the search prompt D11 in Figure 11 may include a description of data retrieval for user status data in the user DB 86. Furthermore, the server 8B does not need to store the environment DB 88 in the storage unit 81, nor does it need to acquire environment data from the air conditioner 10.

[0125] As described above, in this embodiment, the control unit 80 of the server 8B further acquires user status data, which is an example of data indicating the user's activity status, based on the output statement generated by inference processing from a first prompt containing a user request message to the air conditioner 10 (see Figure 15). This makes it possible to acquire, for example, the user's user status data in addition to the equipment operation data of the air conditioner 10 in response to the user's request message. For example, based on the data acquired in this way, it becomes easier to perform control that is adapted not only to the operating status of the air conditioner 10 but also to the user's own status.

[0126] (Other Embodiments) Although the present disclosure has been described above with reference to Embodiments 1 to 3, the present disclosure is not limited to the embodiments described above. The technology in the present disclosure is not limited thereto and can be applied to embodiments that have been modified, replaced, added, or omitted as appropriate.

[0127] In each of the embodiments described above, an example was explained in which a request message to the air conditioner 10 is input as text data by the terminal device 7 of the control system 1. Such request messages from the user are not limited to text input; they may also be acquired based on, for example, voice data representing the voice spoken by the user. For example, the request message may be converted from voice data to text data by a trained model or algorithm using various machine learning methods for speech-to-text recognition in the terminal device 7 and then sent to the server 8.

[0128] In each of the embodiments described above, a control system 1 was described in which a request message to the air conditioner 10 is input by a terminal device 7. The request message is not limited to the terminal device 7, but may also be input by the air conditioner 10, for example. A control system 1A relating to such further modifications will be described with reference to Figure 16.

[0129] Figure 16 is a block diagram showing the configuration of a control system 1A according to a further modification of Embodiment 1. In addition to having the same configuration as the control system 1 shown in Figure 1, the control system 1A of this modification is equipped with a microphone 16 in the indoor unit 20 of the air conditioner 10. The microphone 16 is configured similarly to, for example, the microphone 75 of the terminal device 7. The microphone 16 of this system 1A is an example of an input interface for inputting voice data. In this modification, the air conditioner 10 acquires a request message based on voice data when voice data including a predetermined wake word (for example, "air conditioner") is input.

[0130] In the above modified example, for instance, the control unit 12 of the air conditioner 10 transmits the audio data input from the microphone 16 to the terminal device 7 via the communication unit 13. The audio data may be converted to text data by the terminal device 7 and then transmitted to the server 8, or it may be transmitted to the server 8 and then converted to text data by the server 8. Alternatively, the control unit 12 of the air conditioner 10 may convert the input audio data to text data. The air conditioner 10 may then transmit the request message indicated by the text data to the terminal device 7 or the server 8.

[0131] As described above, in this modified example, the air conditioner 10 in the control system 1A is equipped with a microphone 16 as an example of an input interface, and the microphone 16 inputs voice data indicating a request message from the user to the air conditioner 10. With this control system 1A, the request message can be obtained based on the user's voice data input into the microphone 16 of the air conditioner 10.

[0132] In each of the embodiments described above, an example was explained in which control information F1 relating to the control of the air conditioner 10 is included in the prompt, as illustrated in prompt D1 (Figure 8). In this embodiment, information relating to the control of the air conditioner 10, such as control information F1, may be obtained in response to a request message from the user, for example, by document search technology such as search extension generation (RAG). For example, a prompt containing the obtained information is input to the LLM in place of or in addition to the control information F1. This allows information that is not initially included in the prompt to be obtained by RAG and added to the prompt, and to be reflected in the control of the air conditioner 10 using the LLM's inference processing.

[0133] In the embodiments described above, control systems 1 and 1A were described in which a common prompt D1, etc., is used regardless of the equipment information corresponding to the user of the air conditioner 10. In this embodiment, different prompts may be used depending on the equipment information. For example, in the control system of this embodiment, a prompt including functions, usage, and / or specifications specific to the model number of the air conditioner 10 indicated by the equipment information associated with the user in the user DB 86 may be input to the LLM. This allows for adjustment of the control of the air conditioner 10 and the response message to the user according to the model number, even if the user's request message is the same.

[0134] In the above embodiment, the control system may, for example, switch the prompt to be read from the prompts stored in the server's storage unit 81 for each part number, according to the part number of the air conditioner 10. The control system may, for example, obtain information associated with the part number from the equipment information corresponding to the user each time a request message is entered from the user, and add it to each prompt.

[0135] In the embodiments described above, the prompts input to the LLM in the control systems 1 and 1A were shown to include past input / output history to and from the LLM (S5, S7, S9). Such prompts input to the LLM may be acquired in a way that excludes input / output history prior to a predetermined period from the time of input. This makes it possible to suppress excessive length of the input prompts in each step S5, S7, and S9, reduce noise in the output sentences from the LLM, and accommodate limitations on the number of characters that can be input to the LLM.

[0136] In the embodiments described above, an example was given in which the control method of the Disclosure is executed in the server 8 of the control systems 1 and 1A. The control method of the Disclosure may also be executed, for example, in a terminal device 7 or an air conditioner 10.

[0137] While this disclosure is adequately described in relation to preferred embodiments with reference to the accompanying drawings, various variations and modifications will be obvious to those skilled in the art. Such variations and modifications should be understood to be included within the scope of this disclosure as defined by the attached claims.

[0138] Furthermore, general and specific aspects of this disclosure may be implemented by systems, methods, computer programs, computer-readable storage media, and combinations thereof.

[0139] Furthermore, in this specification, terms such as "first," "second," etc., are used solely for illustrative purposes and should not be understood as expressing or implying relative importance or ranking of technical features. Features designated as "first" and "second" express or imply that they include one or more such features.

[0140] (Examples of embodiments) The embodiments of this disclosure are described below as examples.

[0141] A first aspect of this disclosure is a control device for an air conditioner. The control device comprises a communication interface for data communication and a control unit that controls an inference process that inputs a prompt to a large-scale language model and generates an output sentence by the large-scale language model. The large-scale language model is trained by machine learning to generate an output sentence based on an input prompt. The control unit obtains text indicating a user's request to the air conditioner, which is included in the prompt and input to the large-scale language model, via the communication interface; obtains operating state data indicating the operating state of the air conditioner based on an output sentence generated by the inference process from a first prompt containing the obtained text; and changes the operating state of the air conditioner via the communication interface based on an output sentence generated by the inference process from a second prompt containing the obtained operating state data.

[0142] A second embodiment is the control device described in the first embodiment. In the second embodiment, the first and second prompts include instructions to cause the large-scale language model to output a command statement that causes the control unit to perform either acquiring operating state data or changing the operating state.

[0143] A third embodiment is the control device described in the second embodiment. In the third embodiment, the first and second prompts are configured to cause a large language model to output an instruction statement based on a lower-level prompt. The lower-level prompts include a prompt for acquiring operating status data and a prompt for changing the operating status. The control unit selectively inputs the lower-level prompts to the large language model based on the output statements generated from the first and second prompts, respectively.

[0144] A fourth embodiment is a control device according to any of the first to third embodiments. In the fourth embodiment, the first and second prompts include information relating a user request indicated by text to a control that changes the operating state of the air conditioner.

[0145] The fifth embodiment is a control device described in any of the first to fourth embodiments. In the fifth embodiment, the control unit acquires operating status data according to the acquired text by searching for data received from the air conditioner regarding the operating status via a communication interface.

[0146] The sixth embodiment is a control device according to any of the first to fifth embodiments. In the sixth embodiment, the control unit further includes the first prompt and the output sentence generated from the first prompt in the second prompt and inputs them to the large language model.

[0147] The seventh embodiment is the control device described in the sixth embodiment. In the seventh embodiment, the control unit inputs a second prompt, an output statement generated from the second prompt, and a third prompt including the result of changing the operating state of the air conditioner in response to the output statement from the second prompt to a large language model, and outputs a response statement to the user's request indicated by the text, or performs either acquiring operating state data or changing the operating state, based on the output statement generated by inference processing from the third prompt.

[0148] The eighth embodiment is a control device according to any of the first to seventh embodiments. In the eighth embodiment, the control unit further acquires data observed in the environment outside the air conditioner based on an output statement generated by inference processing from a first prompt containing acquired text.

[0149] The ninth embodiment is a control device according to any of the first to eighth embodiments. In the ninth embodiment, the control unit further acquires data indicating the user's activity state based on an output statement generated by inference processing from a first prompt containing acquired text.

[0150] A tenth aspect is a control system for an air conditioner. The control system comprises an air conditioner, an input interface for inputting data, and a control unit that controls an inference process that inputs prompts to a large-scale language model and generates output sentences using the large-scale language model. The large-scale language model is trained by machine learning to generate output sentences based on input prompts. The control unit obtains text from the input interface that indicates a user's request to the air conditioner and is included in the prompt to be input to the large-scale language model, obtains operating state data indicating the operating state of the air conditioner based on an output sentence generated by the inference process from a first prompt containing the obtained text, and controls the air conditioner to change its operating state based on an output sentence generated by the inference process from a second prompt containing the obtained operating state data.

[0151] The eleventh aspect is an air conditioner in the control system described in the tenth aspect. The air conditioner is equipped with an input interface, through which data representing text is input.

[0152] A twelfth aspect is a method for controlling an air conditioner. The control method includes a control unit that controls an inference process that inputs a prompt to a large language model and generates an output sentence by the large language model, the steps of: acquiring text indicating a user's request to the air conditioner, which is included in the prompt and input to the large language model; acquiring operating state data indicating the operating state of the air conditioner based on an output sentence generated by the inference process from a first prompt including the acquired text; and controlling the air conditioner to change the operating state of the air conditioner based on an output sentence generated by the inference process from a second prompt including the acquired operating state data.

[0153] This disclosure is applicable to various control devices, control systems, and control methods for controlling air conditioners using large-scale language models.

Claims

1. A control device for an air conditioner, comprising: a communication interface for data communication; and a control unit for controlling an inference process that inputs a prompt to a large-scale language model and generates an output sentence by the large-scale language model, wherein the large-scale language model is trained by machine learning to generate an output sentence based on an input prompt; the control unit, via the communication interface, acquires text indicating a user's request to the air conditioner, which is included in the prompt and input to the large-scale language model; acquires operating state data indicating the operating state of the air conditioner based on an output sentence generated by the inference process from a first prompt including the acquired text; and changes the operating state of the air conditioner via the communication interface based on an output sentence generated by the inference process from a second prompt including the acquired operating state data.

2. The control device according to claim 1, wherein the first and second prompts include instructions to cause the large-scale language model to output a command statement that causes the control unit to perform either the acquisition of the operating state data or the modification of the operating state.

3. The control device according to claim 2, wherein the first and second prompts are configured to cause the large language model to output the command statement based on the lower prompt, the lower prompts include a prompt for acquiring the operating state data and a prompt for changing the operating state, and the control unit selectively inputs the lower prompts to the large language model based on the output statements generated from the first and second prompts, respectively.

4. The control device according to claim 1, wherein the first and second prompts include information relating the user's request indicated by the text to a control that changes the operating state of the air conditioner.

5. The control device according to claim 1, wherein the control unit retrieves the operating status data in accordance with the acquired text by searching for data received from the air conditioner via the communication interface regarding the operating status.

6. The control device according to claim 1, wherein the control unit further includes the first prompt and the output sentence generated from the first prompt in the second prompt and inputs them to the large language model.

7. The control device according to claim 6, wherein the control unit inputs the second prompt, an output statement generated from the second prompt, and a third prompt including the result of changing the operating state of the air conditioner in accordance with the output statement from the second prompt to the large-scale language model, and outputs a response statement to the user's request indicated by the text, or performs either the acquisition of the operating state data or the change of the operating state, based on the output statement generated by the inference process from the third prompt, 8. The control device according to claim 1, wherein the control unit further acquires data observed in the environment outside the air conditioner based on the output statement generated by the inference process from the first prompt including the acquired text.

9. The control device according to claim 1, wherein the control unit further acquires data indicating the user's activity status based on an output statement generated by the inference process from a first prompt including acquired text.

10. A control system comprising: an air conditioner; an input interface for inputting data; and a control unit for controlling an inference process that inputs a prompt to a large-scale language model and generates an output sentence by the large-scale language model, wherein the large-scale language model is trained by machine learning to generate an output sentence based on an input prompt; the control unit acquires text indicating a user's request to the air conditioner via the input interface, which is included in a prompt and input to the large-scale language model; acquires operating state data indicating the operating state of the air conditioner based on an output sentence generated by the inference process from a first prompt including the acquired text; and controls the air conditioner to change its operating state based on an output sentence generated by the inference process from a second prompt including the acquired operating state data.

11. An air conditioner in the control system according to claim 10, comprising the input interface, wherein data representing the text is input via the input interface.

12. A control method for an air conditioner, comprising: a control unit that controls an inference process that inputs a prompt to a large-scale language model and generates an output sentence by the large-scale language model, the control unit acquiring text indicating a user's request to the air conditioner which is included in the prompt and input to the large-scale language model; acquiring operating state data indicating the operating state of the air conditioner based on an output sentence generated by the inference process from a first prompt including the acquired text; and controlling the air conditioner to change its operating state based on an output sentence generated by the inference process from a second prompt including the acquired operating state data.

Images