Air conditioner

Abstract

This air conditioner comprises an adapter that communicates with a server device. The adapter has a storage unit, a communication unit, and a control unit. The storage unit stores an initial training model that has a learned parameter relating to a prescribed prediction process executed by the air conditioner. The communication unit collects sensor data detected by a sensor included in the air conditioner, transmits the sensor data to the server device, and receives a first training model generated by the server device on the basis of the sensor data. The control unit executes the prescribed prediction process by using the initial training model until the first training model is received, and after reception of the first learning model, executes the prescribed prediction process by using the first training model.

Description

Air Conditioner

[0001] The present invention relates to an air conditioner capable of implementing various prediction functions through learning processing.

[0002] Conventionally, there has been known an air conditioning system that causes a learning unit such as AI (Artificial Intelligence) to learn operation history data such as a set temperature to generate a learning model, and controls air conditioning operation using the generated learning model (hereinafter also referred to as AI control). The learning model generated by the learning unit includes parameters obtained through learning (also referred to as learned parameters).

[0003] For example, the following Patent Document 1 describes an air conditioning system including a cloud-side learning unit and an adapter-side learning unit. Specifically, an adapter downloads a first learning model generated by the cloud-side learning unit to obtain a second learning model, and the adapter-side learning unit performs learning on the second learning model based on latest operation history data (the operation preference of a user in an air-conditioned space and changes in the actual environment) to generate a new second learning model.

[0004] Japanese Patent Application Laid-Open No. 2021-63611

[0005] However, even in the case of an air conditioner that combines rapid learning on the air conditioner side and high-accuracy learning on the cloud side as described in Patent Document 1, high-accuracy learning requires collecting data indicating the user's preference and the user's presence / absence tendency in the air-conditioned space over a predetermined period after the air conditioner is installed. Therefore, the user of the air conditioner cannot early receive the benefits of the learning model until such data is collected.

[0006] In view of the above circumstances, an object of the present invention is to provide an air conditioner that allows a user thereof to early receive the benefits of a learning model.

[0007] An air conditioner according to one embodiment of the present invention is an air conditioner equipped with an adapter that communicates with a server device. The adapter has a storage unit, a communication unit, and a control unit. The storage unit stores an initial learning model having learned parameters related to a predetermined prediction process performed by the air conditioner. The communication unit collects sensor data detected by sensors of the air conditioner and transmits it to the server device, and receives a first learning model generated by the server device based on the sensor data. The control unit executes the predetermined prediction process using the initial learning model until it receives the first learning model, and after receiving the first learning model, it executes the predetermined prediction process using the first learning model.

[0008] This configuration allows users of air conditioners to benefit from the learning model early on by performing prediction processing using the initial learning model until the first learning model is received.

[0009] The control unit may update the learned parameters of the initial learning model based on the sensor data until it receives the first learning model, and after receiving the first learning model, it may update the learned parameters of the first learning model based on the sensor data.

[0010] This configuration allows the learning model to be further trained using sensor data from the air conditioner, regardless of whether the first learning model has received the data or not.

[0011] The control unit may use the initial learning model and the first learning model to predict whether the user of the air conditioner is present or absent in the air-conditioned space.

[0012] This configuration allows for the prediction of the user's presence or absence even before the first learning model receives the data.

[0013] The initial learning model may have multiple learned parameters to accommodate multiple presence / absence trends of the user in the air-conditioned space.

[0014] This configuration allows the presence or absence of a user to be predicted using multiple presence / absence patterns even before the first learning model receives data.

[0015] The plurality of learned parameters may include a first learned parameter corresponding to an presence / absence pattern having a first absence time as the absence time in the air-conditioned space, and a second learned parameter corresponding to an presence / absence pattern having a second absence time shorter than the first absence time as the absence time in the air-conditioned space.

[0016] This configuration allows for the prediction of a user's presence or absence even before receiving data from the first learning model, by corresponding to patterns of presence or absence with long periods of absence and patterns with short periods of absence, such as weekdays and holidays.

[0017] The control unit may start collecting the sensor data when external power is supplied to the air conditioner, and may continue collecting the sensor data regardless of whether the air conditioner's air conditioning operation is started or stopped.

[0018] This configuration allows for the collection of sensor data even when the air conditioner is not operating, as long as external power is supplied to the air conditioner.

[0019] The control unit may, each time it receives the latest first learning model from the server device, compare the received first learning model with the previously received first learning model. If the two models do not match, it may update the previously received first learning model and the second learning model (which is the previously received first learning model further trained based on the sensor data) using the latest first learning model, store them in the storage unit, and execute the predetermined prediction process using the updated second learning model.

[0020] This configuration allows the system to store the previously received first learning model for comparison, and by determining whether there is a difference between the latest first learning model and the previously received first learning model, it is possible to always use the latest second learning model while avoiding unnecessary update processes.

[0021] The sensor data may be set to be collected at predetermined time intervals, have a timestamp indicating the collection time, and be stored in the order of the timestamps. In this case, the control unit may determine whether the timestamp of a predetermined number of sensor data, which has a predetermined timestamp, falls within a predetermined time range before and after the timestamp that the sensor data to be acquired at the predetermined number according to the setting should have, and if it determines that the timestamp is outside the predetermined time range before and after, it may determine that the sensor data is missing.

[0022] This configuration ensures reliable detection of deficiencies (missing data) in sensor data, even if there are errors in the sensor data timestamps. As a result, in this embodiment, additional training using datasets suspected to contain missing sensor data can be stopped. If there are no missing data in subsequently collected datasets (for example, datasets from the following day onwards), additional training will be performed on those datasets.

[0023] As described above, the present invention provides an air conditioner that allows users to benefit from the learning model at an early stage. However, this effect is not limited to the present invention.

[0024] This is a diagram showing the configuration of an air conditioning system according to one embodiment of the present invention. This is a block diagram showing the configuration of an adapter for an air conditioner according to one embodiment of the present invention. This is a block diagram showing the configuration of a server device for an air conditioning system according to one embodiment of the present invention. This is a flowchart showing the flow of learning processing by the server device in one embodiment of the present invention. This is a flowchart showing the flow of occupancy / absence prediction processing by the adapter for an air conditioner according to one embodiment of the present invention. This is a diagram explaining the preprocessing of sensor data by the adapter for an air conditioner according to one embodiment of the present invention. This is a diagram showing the update formula for the occupancy / absence pattern of the learning model by the adapter for an air conditioner according to one embodiment of the present invention. This is a flowchart showing the flow of the learning model update processing by the adapter for an air conditioner according to one embodiment of the present invention. This is a diagram explaining the learning model update processing by the adapter for an air conditioner according to one embodiment of the present invention. This is a diagram schematically showing the occupancy / absence pattern of the learning model used by the air conditioner according to one embodiment of the present invention. This is a diagram schematically showing the initial occupancy / absence pattern of the initial learning model used by the air conditioner according to one embodiment of the present invention.

[0025] Embodiments of the present invention will be described below with reference to the drawings.

[0026] [System Configuration] Figure 1 is a diagram showing the configuration of the air conditioning system according to this embodiment.

[0027] As shown in the figure, the air conditioning system according to this embodiment includes an air conditioner 100, an access point 3, a server device 4, a relay device 5, and a communication terminal 6, which are connected by a network 7. The air conditioner 100 includes an indoor unit 1 and an adapter 2.

[0028] The indoor unit 1 is, for example, part of an air conditioner 100 that is placed indoors and heats or cools the indoor air. The user of the indoor unit 1 can remotely operate the indoor unit 1 by operating the remote control 8. The indoor unit 1 is connected to an outdoor unit (not shown) by refrigerant piping, and the outdoor unit is equipped with an outdoor fan, compressor, etc.

[0029] The indoor unit 1 comprises a main body 1A, a control unit 1B that controls the main body 1A, and a sensor 1C. The main body 1A houses various devices, such as an indoor fan, expansion valve, indoor heat exchanger, and air deflector (not shown), within its casing.

[0030] The control unit 1B controls the indoor fan, expansion valve, and air deflector mentioned above. The control unit 1B adjusts the opening of the expansion valve to supply the necessary amount of refrigerant to the indoor heat exchanger to achieve the air conditioning capacity required by the user. The control unit 1B also drives the indoor fan to blow out the indoor air, which has undergone heat exchange with the refrigerant in the indoor heat exchanger, into the air-conditioned space by deflecting it with the air deflector. This enables heating, cooling, and dehumidification of the air-conditioned space.

[0031] Sensor 1C, for example, begins detecting the presence or absence of people in the air-conditioned space where the indoor unit 1 is installed, once the indoor unit 1 is connected to the commercial power supply and power is supplied. From this point onward, unless the power supply is interrupted, it will continue to detect the presence or absence of people in the air-conditioned space regardless of whether the indoor unit 1 is running or stopped.

[0032] In addition, sensors 1C may include a radiation sensor for measuring the radiant temperature of the floor surface in the air-conditioned space, a room temperature sensor for measuring the indoor temperature in the air-conditioned space, a humidity sensor for measuring the indoor humidity in the air-conditioned space, and so on.

[0033] The adapter 2 has a communication function that connects the indoor unit 1 and the access point 3 wirelessly, and a control function that AI-controls the indoor unit 1. An adapter 2 is provided for each indoor unit 1.

[0034] Access point 3 is a device that connects adapter 2 and network 7 wirelessly, for example, using a WLAN (Wireless Local Area Network). Network 7 is a communication network such as the Internet.

[0035] The server device 4 has functions such as generating a learning model used by the adapter 2 when it AI-controls the indoor unit 1, and a database for storing sensor data, etc. The server device 4 is located, for example, in a data center.

[0036] The relay device 5 is connected to the network 7 by wired communication and also to the server device 4 by wired communication. The relay device 5 receives sensor data and the like used for generating or updating the learning model in the server device 4 from the indoor unit 1 via the adapter 2 and the network 7 and transmits it to the server device 4. The relay device 5 also transmits the learning model generated or updated in the server device 4 to the adapter 2 via the network 7. The relay device 5 is located, for example, in a data center. The relay device 5 has a first relay unit 5A, a second relay unit 5B, and a third relay unit 5C.

[0037] The first relay unit 5A relays various data related to AI control between the adapter 2 and the server device 4. Specifically, the first relay unit 5A transmits sensor data and the like used to generate or update the learning model, received from the adapter 2 via the access point 3 and the network 7, to the server device 4, and transmits the learning model generated or updated by the server device 4 to the adapter 2 via the network 7 and the access point 3.

[0038] The second relay unit 5B receives the operating conditions of the indoor unit 1 (such as the operating mode, like cooling or heating, and the set temperature) set by the user using the communication terminal 6 while away from home, via the access point 3 and network 7, and transmits this to the indoor unit 1 via the network 7, access point 3, and adapter 2.

[0039] The third relay unit 5C receives external data, such as weather forecasts, via the network 7, and transmits the received external data to the server device 4. The third relay unit 5C also transmits the received external data to the adapter 2 via the network 7 and access point 3.

[0040] Alternatively, the relay device 5 may be omitted, and the adapter 2 and the server device 4 may communicate directly.

[0041] Communication terminal 6 is a mobile device such as a smartphone owned by the user.

[0042] [Configuration of Adapter] FIG. 2 is a block diagram showing the configuration of the adapter 2. As shown in the figure, the adapter 2 includes a first communication unit 21, a second communication unit 22, a CPU (Central Processing Unit) 23, and a storage unit 24.

[0043] The first communication unit 21 is a communication interface (IF) such as a UART (Universal Asynchronous Receiver Transmitter) that is connected to the control unit 1B in the indoor unit 1 via wired communication. The second communication unit 22 is a communication IF such as a WLAN that is connected to the access point 3 via wireless communication.

[0044] The CPU 23 controls the entire adapter 2. The storage unit 24 includes, for example, a ROM (Read Only Memory), a RAM (Random Access Memory), and the like, and stores various types of information such as data and programs.

[0045] The CPU 23 includes an acquisition unit 23A, a transmission unit 23B, a reception unit 23C, an adapter-side learning unit 23D, a setting unit 23E, and a prediction unit 23F.

[0046] The acquisition unit 23A acquires sensor data from the indoor unit 1 at predetermined time intervals, for example, acquisition timing every 5 minutes. The acquisition unit 23A stores the sensor data acquired at 5-minute cycles in a sensor data memory 24A. Details of the sensor data will be described later.

[0047] The transmission unit 23B transmits the sensor data stored in the sensor data memory 24A to the access point 3, the access point 3 transmits the received sensor data to the relay device 5 via the network 7, and a first relay unit 5A of the relay device 5 transmits the received sensor data to the server device 4.

[0048] The receiving unit 23C receives, from the server device 4, a learning model generated by the server device 4 (hereinafter also referred to as an AI learning model). In this embodiment, as the AI learning model, for example, an absence / presence prediction model for predicting whether a user is present or absent is received, and the received AI learning model is stored in the model memory 24C. The absence / presence prediction model has, for example, an absence / presence pattern indicating the tendency of a user's presence or absence in an air-conditioned space. For example, the absence / presence pattern includes a numerical group (trained parameters) indicating the user's existence probability at predetermined time intervals in the air-conditioned space.

[0049] The adapter-side learning unit 23D updates the learning model used by the adapter 2. Specifically, the adapter-side learning unit 23D updates the AI learning model (absence / presence prediction model) stored in the model memory 24C through additional learning based on sensor data.

[0050] Generating an AI learning model by the server device 4 requires a large amount of sensor data (for example, sensor data for 30 days). The adapter-side learning unit 23D determines the initial learning model pre-stored in the adapter 2 as the learning model to be used on the adapter 2 side until the adapter 2 receives the AI learning model generated by the server device 4 from the server device 4. After receiving the AI learning model from the server device 4, the adapter-side learning unit 23D determines the AI learning model as the learning model to be used on the adapter 2 side.

[0051] The adapter-side learning unit 23D updates the trained parameters of the initial learning model based on the sensor data until the AI learning model is received from the server device 4. After receiving the AI learning model, the adapter-side learning unit 23D updates the trained parameters of the AI learning model based on the sensor data until an updated AI learning model is next received from the server device 4. Such a learning process in which the learning model is additionally learned based on sensor data and the trained parameters are updated on the air conditioner 100 side instead of the server device 4 side is referred to as edge learning process.

[0052] The setting unit 23E applies the learning model stored in the model memory 24C to the prediction unit 23F. Based on the learning model, the prediction unit 23F predicts whether or not there are people in the air-conditioned space of the indoor unit 1.

[0053] Based on the prediction results of presence or absence, the control unit 1B of the indoor unit 1 transmits to the user's communication terminal 6 the content of the air conditioning operation recommended to the user, such as setting the start time or stop time of the air conditioning operation. The content of the recommended air conditioning operation includes, for example, setting a timer time for starting the air conditioning operation according to the wake-up time, return-home time, departure time, and bedtime.

[0054] Furthermore, based on the predicted presence or absence, the control unit 1B sends a signal to the user's communication terminal 6 recommending an additional function operation different from the air conditioning operation. For example, the recommendation for an additional function operation might be a suggestion for a timer time to start the heating and sterilization operation at the time when absence is predicted.

[0055] Furthermore, based on the prediction of presence or absence, the control unit 1B pauses the operation of the air conditioner 100 during the time when absence is predicted.

[0056] Specifically, the control unit 1B extracts time periods (for example, periods of absence of 5 hours or more) in which additional function operation, which differs from normal air conditioning operation, can be performed, based on the occupancy prediction results for 24 hours in the air-conditioned space. The control unit 1B calculates the available operating time from the absence time, compares the calculated available operating time with the operating time of the additional function operation stored in advance, and extracts the time period in which the additional function operation can be performed from the absence time if the available operating time is greater than or equal to the operating time of the additional function operation stored in advance.

[0057] The storage unit 24 includes a sensor data memory 24A, a presence / absence pattern memory 24B, a model memory 24C, a flash memory 24D, an external memory 24E, and a prediction result memory 24F.

[0058] The sensor data memory 24A temporarily stores sensor data acquired from the indoor unit 1. The presence / absence pattern memory 24B stores the presence / absence patterns of the AI ​​learning model received from the server device 4 and the presence / absence patterns of the initial learning model. The presence / absence pattern is data that indicates the tendency of users to be present or absent in the air-conditioned space. Specifically, it is an array of data with 1 row and 144 columns (in 10-minute increments from 0:00 to 24:00 and 23:50) where a decimal number from 0 indicating absence to 1 indicating presence is taken at predetermined time intervals.

[0059] The presence / absence patterns of the AI ​​learning model downloaded from the server device 4 are patterns generated by the server device 4 using, for example, sensor data of presence / absence from sensor 1C for the past 30 days, and day of the week and holiday information obtained by the server device 4 from an external server (not shown). In this embodiment, up to five types of presence / absence patterns are generated. Figure 10 schematically shows the presence / absence patterns of this embodiment. In Figure 10 and Figure 11, which will be described later, the presence / absence patterns generated from the sensor data of the air conditioner 100 are schematically shown, for example, by dividing the probability of a user's presence in an air-conditioned space (e.g., living room) into five stages, from highest to lowest probability of presence, as presence probability 5, presence probability 4, presence probability 3, presence probability 2, and presence probability 1. The presence / absence patterns are associated with each day of the week so that the CPU 23 can determine which presence / absence pattern a user tends to behave in on each day of the week. The association between the day of the week and the presence / absence pattern is performed, for example, by the server device 4.

[0060] For example, as shown in Figure 10A, Monday and Tuesday are associated with presence / absence pattern 1; as shown in Figure 10B, Wednesday and Thursday are associated with presence / absence pattern 2; as shown in Figure 10C, Friday is associated with presence / absence pattern 3; as shown in Figure 10D, Saturday is associated with presence / absence pattern 4; and as shown in Figure 10E, Sunday is associated with presence / absence pattern 5.

[0061] Furthermore, the day of the week information is the information for Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, and Sunday. The holiday information is information that identifies holidays among the days of the week Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, and Sunday. The day of the week information and holiday information are obtained from an external server (not shown) via the second communication unit 22.

[0062] On the other hand, the presence / absence patterns of the initial model are pre-stored independently of the sensor data from sensor 1C, and consist of two patterns: a presence / absence pattern that shows a tendency for longer periods of absence (initial presence / absence pattern 1), and a presence / absence pattern that shows a tendency for shorter periods of absence (initial presence / absence pattern 2). Figure 11 schematically shows the initial presence / absence patterns of this embodiment.

[0063] Initial presence / absence pattern 1 is a pattern (weekday pattern) in which the total or continuous absence time during a day (for example, the time when the probability of being absent is above a predetermined value (or the probability of being present is below a predetermined value)) is equal to or greater than a predetermined time (for example, 8 hours; hereinafter also referred to as the first absence time). For example, initial presence / absence pattern 1 is the presence / absence pattern shown in Figure 11A. Initial presence / absence pattern 2 is a pattern (holiday pattern) in which the total or continuous absence time is less than the first absence time (for example, 5 hours; hereinafter also referred to as the second absence time). For example, initial presence / absence pattern 2 is the presence / absence pattern shown in Figure 11B. These presence / absence patterns are, for example, average data of presence / absence pattern trends of a large number of users acquired in advance.

[0064] The model memory 24C stores the AI ​​learning model (first learning model) downloaded from the server device 4, the previous learning model used in the previous presence / absence prediction process, and the online learning model (second learning model), which is used in the user presence / absence prediction process and undergoes additional learning based on sensor data.

[0065] Before the AI ​​learning model is downloaded, the online learning model is either the initial learning model itself, or a learning model in which the initial learning model or its learned parameters have been further trained by the adapter-side learning unit 23D based on the sensor data. On the other hand, if the AI ​​learning model has already been downloaded, the online learning model is either the downloaded AI learning model itself, or a learning model in which the AI ​​learning model or its learned parameters have been further trained by the adapter-side learning unit 23D based on the sensor data. As will be described later, the AI ​​learning model before further training is kept separately for comparison (determination of whether or not it matches) with the newly downloaded AI learning model.

[0066] The flash memory 24D stores the initial learning model used for predicting the presence or absence of the user until it receives the AI ​​learning model from the server device 4. When the initialization of the learning models in the model memory 24C is performed, they revert to the initial learning model.

[0067] The initial learning model has a first learned parameter corresponding to initial presence / absence pattern 1 and a second learned parameter corresponding to initial presence / absence pattern 2. The learned parameters are, for example, a 1x144 array of data (present / absence data) that shows the probability of a user's presence at a predetermined time interval.

[0068] External memory 24E stores external data, such as weather forecast data. Prediction result memory 24F stores prediction results of the presence or absence of users in the air-conditioned space, which are predicted using the AI ​​learning model or initial learning model described above.

[0069] [Server Device Configuration] Figure 3 is a block diagram showing the configuration of the server device 4. As shown in the figure, the server device 4 includes a communication unit 41, a CPU 42, and a storage unit 43.

[0070] The communication unit 41 is a communication interface that connects to the relay device 5 via wired communication and communicates with the adapters 2 of the multiple air conditioners 100 via the relay device 5. The CPU 42 controls the entire server device 4. The storage unit 43 has, for example, an HDD (Hard Disk Drive), ROM, RAM, etc., and stores various information such as data and programs.

[0071] The CPU 42 includes a cloud-side learning unit 42A, a receiving unit 42B, and a transmitting unit 42C.

[0072] The cloud-side learning unit 42A receives sensor data from each air conditioner 100 via the adapter 2, access point 3, network 7, and relay device 5. The cloud-side learning unit 42A then generates an AI learning model using a predetermined period of sensor data (for example, 6 days' worth) stored in the sensor data memory 43A from the sensor data received from each adapter 2. Because the cloud-side learning unit 42A uses all of the sensor data for the predetermined period to generate the AI ​​learning model, it can generate a learning model with high prediction accuracy even if some irregular sensor data that should not be learned is included. This type of learning that uses all of the sensor data is sometimes called batch learning.

[0073] The cloud-side learning unit 42A generates or updates an AI learning model for the air conditioner 100 based on the learning results. The cloud-side learning unit 42A then stores the newly generated AI learning model or the updated learning model of an existing AI learning model in the cloud-side learning storage unit 43B. The cloud-side learning unit 42A can also receive sensor data from multiple air conditioners 100 and generate or update an AI learning model for each air conditioner 100.

[0074] The receiving unit 42B receives sensor data and the like from each adapter 2 via the access point 3, network 7, and relay device 5. The transmitting unit 42C transmits to the adapter 2 of each air conditioner 100 via the relay device 5, network 7, and access point 3 an AI learning model newly generated for each air conditioner 100 by the cloud-side learning unit 42A, or a learning model that is an updated version of an existing AI learning model.

[0075] The memory unit 43 includes a sensor data memory 43A and a cloud-side learning memory unit 43B. The sensor data memory 43A stores sensor data received from the adapter 2 of the air conditioner 100 for each air conditioner 100.

[0076] Each air conditioner 100 has an air conditioner ID for identification, and when the adapter 2 transmits sensor data, it also transmits this air conditioner ID. The server device 4, which receives the sensor data from the adapter 2, stores the sensor data in the sensor data memory 43A for each air conditioner ID included in the received signal. The cloud-side learning memory unit 43B stores the AI ​​learning model generated or updated by the server device 4.

[0077] [System Operation] Next, we will explain the operation of the system configured as described above.

[0078] (Learning process by the server device) First, the learning process performed by the server device 4 will be explained. The operation of the server device 4 is performed by the cooperation of the CPU 42 of the server device 4 and the software stored in the memory unit 43. For convenience, in the following explanation, the CPU 42 will be considered the main operator. Figure 4 is a flowchart showing the flow of the learning process performed by the server device 4.

[0079] As shown in the figure, the CPU 42 (cloud-side learning unit 42A) of the server device 4 first determines whether or not it has received sensor data from the adapter 2 (step 41). The sensor data transmitted from the adapter 2 is, for example, six days' worth of data.

[0080] If the CPU 42 determines that sensor data has been received (Yes in step 41), it stores the sensor data in the sensor data memory 43A (step 42). If the CPU 42 determines that sensor data has not been received (No in step 41), it terminates processing.

[0081] Next, the CPU 42 determines whether the sensor data stored in the sensor data memory 43A has existed for a predetermined period of time from the most recent time (step 43). In the presence / absence prediction model of this embodiment, the predetermined period is, for example, about one to two months. If it is determined that the sensor data has not existed for the predetermined period (No. in step 43), the CPU 42 terminates the process.

[0082] If the CPU determines that the above sensor data exists for a predetermined period of time or longer (Yes in step 43), the CPU 42 generates an AI learning model to predict the presence or absence of people based on the sensor data (step 44).

[0083] Next, the CPU 42 stores and updates the generated AI learning model in the cloud-side learning memory unit 43B (step 45).

[0084] The CPU 42 then transmits the AI ​​learning model to the adapter 2 (step 46). The CPU 42 performs the above process each time it receives sensor data from the adapter 2.

[0085] (Operation of the air conditioner) Next, the operation of the air conditioner 100 will be explained. The operation of the air conditioner 100 is performed by the cooperation of the hardware, such as the CPU 23 of the adapter 2 of the air conditioner 100, and the software stored in the storage unit 24. For convenience, in the following explanation, the CPU 23 will be considered the main operator.

[0086] Figure 5 is a flowchart showing the flow of the presence / absence prediction and learning processes by adapter 2. The learning process is the process in which CPU 23 updates (additional learning) the learned parameters of the online learning model currently in use.

[0087] As shown in the figure, the CPU 23 of the adapter 2 first determines whether or not the AI ​​learning model received from the server device 4 exists in the model memory 24C (step 51).

[0088] If the CPU 23 determines that the AI ​​learning model received from the server device 4 exists in the model memory 24C (Yes in step 51), the CPU 23 executes presence / absence prediction processing on the AI ​​learning model as an online learning model (step 53).

[0089] On the other hand, if the CPU 23 determines that the AI ​​learning model received from the server device 4 does not exist in the model memory 24C (No. in step 51), the CPU 23 deploys the initial learning model stored in the flash memory 24D to the model memory 24C as an online learning model and performs presence / absence prediction processing (step 52).

[0090] Next, the CPU 23 determines whether or not it has received sensor data for a predetermined period of time from the indoor unit 1 (step 54). The predetermined period is, for example, 30 hours, but is not limited to this.

[0091] If the CPU determines that it has received sensor data for a predetermined time (Yes in step 54), it determines whether the sensor data is valid as data to be used for training the online learning model (step 55).

[0092] Here, the presence / absence sensor data (raw data) is, for example, the amount of human activity over a 5-minute period in an air-conditioned space, shown in five stages: 1: no activity (absent), 2: low activity, 3: medium activity, 4: high activity, and 5: undetermined activity, and has a timestamp indicating the time the data was collected. The CPU 23 determines that the sensor data is invalid if it determines that this 5-minute activity level has not been updated for a predetermined time during the day, or throughout the day.

[0093] Specifically, for example, if the sensor data shows that the activity level is low, medium, or high for a predetermined period of time within a day or throughout the day, and there are no instances of no activity, or if there is no activity for a predetermined period of time within a day or throughout the day, and there are no instances of low, medium, or high activity, then the sensor data is determined to be invalid.

[0094] If the CPU determines that the sensor data is invalid (No in step 55), it stops the learning process and returns to step 51. If the CPU determines that the sensor data is valid (Yes in step 55), it preprocesses the sensor data for learning (step 56).

[0095] Specifically, CPU 23 uses sensor data from 3:00 AM on the day of the learning session up to 30 hours prior (up to 9:00 PM on the day before the day before the learning session) as input. From the input sensor data, CPU 23 outputs presence / absence data from 0:00 AM to 12:00 AM on the day before the learning session at 10-minute intervals.

[0096] More specifically, the CPU 23 first converts the above five-level activity data into presence / absence information represented by two integers: present (1) / absent (0). For example, if the activity level over a five-minute period is low, medium, or high, it is considered present (1), and if there is no activity or the activity level is undetermined, it is considered absent (0) (see, for example, the upper table in Figure 6). However, if the sensor data obtained by converting the five-minute activity data into presence / absence information is used directly to quantify presence / absence, there is a risk that the tendency of the user's presence or absence in the air-conditioned space may not be accurately captured. For example, even if the user is present in the air-conditioned space where the indoor unit 1 is installed (e.g., living room), the sensor 1C may not detect the user and incorrectly determine that the user is absent. Alternatively, if the user repeatedly moves between the air-conditioned space and other rooms (e.g., a child's room or bedroom) for short periods, it may be incorrectly assumed that the air-conditioned space is often unoccupied. In such cases, if the interval between present and present in the presence / absence information is within a predetermined time (e.g., within 180 minutes), the absence during that period may be supplemented to present. In this embodiment, in order to complement the presence / absence pattern of the day before the learning day, sensor data from two days prior to the learning day (up to 21:00) and sensor data from the day of the learning day (up to 3:00) are acquired.

[0097] In this way, a group is created using 144 sensor data points, divided into 10-minute intervals within the range of 0:00 to 24:00 the day before the learning day, and the presence / absence value (1 / 0) corresponding to that time is entered (for example, the table below in Figure 6).

[0098] In the preprocessing, the CPU 23 performs processing that takes into account the timestamp discrepancy of the input sensor data. Figure 6 is a diagram illustrating this process.

[0099] The input sensor data is data obtained by converting activity level data for a predetermined time (e.g., 5 minutes) into presence / absence information. As mentioned above, it is set to be collected at predetermined time intervals (e.g., every 5 minutes), has a timestamp indicating the collection time, and is stored in the order of the timestamps. However, errors can occur in the timestamps. Since these errors accumulate, if the time is used as the basis for inputting the presence / absence values, in the worst case, an error of about 10 minutes may occur in the timestamp of the sensor data within the range of 0:00 to 24:00 the day before the learning day. Therefore, when the CPU 23 inputs presence / absence values ​​into a group created from 144 sensor data points divided into 10-minute intervals over 24 hours, it does not refer to the time, but instead focuses on the number of data points and inputs the 144 data points in the order in which the data was collected.

[0100] On the other hand, if we focus only on the number of data points, we can eliminate errors in the timestamp, but we cannot detect data loss that may occur due to power outages or other reasons (because the input uses sensor data from 3:00 on the learning day to 21:00 two days prior to the learning day, so even if several hours of data are missing, the system will still count the data from before that time). Therefore, the CPU 23 sets a range that can tolerate a time shift in the entire data set, and if the error exceeds this range, it judges the data to be defective (missing) and interrupts the learning process.

[0101] In other words, as shown in Figure 6, the CPU 23 determines the data closest to 21:00 two days prior to the learning day as the first data, and sets the 360th data as the last data. The CPU 23 determines whether the timestamp of the last data is within 3:00 ± a predetermined range on the learning day (i.e., whether the timestamp is within a predetermined range around 3:00, which is the timestamp that the 360th sensor data acquired according to the above setting (5-minute intervals) should have). If it exceeds the predetermined range, the CPU 23 interrupts the learning process, indicating that there is missing data. In this embodiment, the predetermined range is ±10 minutes (2:50 to 3:10).

[0102] This configuration ensures reliable detection of deficiencies (missing data) in sensor data, even if the timestamp of the sensor data deviates from the set time. As a result, in this embodiment, additional training using a dataset suspected to have missing sensor data can be stopped. If there are no missing data in the dataset collected thereafter (for example, presence / absence data from the next day), additional training will be performed on that dataset.

[0103] Next, the CPU 23 performs an update (additional learning) process for the presence / absence pattern of the online learning model based on the sensor data (presence / absence data) after the above preprocessing (step 57).

[0104] Specifically, the CPU 23 first extracts presence / absence patterns from the current online learning model. As mentioned above, the initial learning model has two presence / absence patterns, but after downloading the AI ​​learning model from the server device 4, it will have up to one to five presence / absence patterns.

[0105] As described above, since there are multiple (two or more) presence / absence patterns, the CPU 23 selects the presence / absence pattern to be updated. Specifically, the CPU 23 calculates the cosine similarity between the multiple presence / absence patterns stored and the pre-processed sensor data (present / absence data), each as an array of 144 columns (10-minute intervals from 0:00 to 24:00 and 23:50). The presence / absence pattern with the highest cosine similarity is selected for update. The CPU 23 then updates the presence / absence pattern in the presence / absence pattern memory 24B for the combination of the presence / absence pattern (a) and presence / absence data (b) with the highest similarity, using the update formula shown in Figure 7 (the difference between the presence / absence pattern and the presence / absence data multiplied by a weight parameter).

[0106] In other words, the CPU 23 applies the update formula to each of the 144 elements of the presence / absence pattern and overwrites the original presence / absence pattern (a) with the calculation result (a') calculated by the update formula. The CPU 23 then overwrites the presence / absence pattern of the online learning model with the presence / absence pattern updated by this overwrite. The above is the update (additional learning) process for the presence / absence pattern of the online learning model. Note that the value of the weight parameter w in the update formula in Figure 7 is, for example, 0.1, but is not limited to this.

[0107] Next, the CPU 23 determines whether the sensor data has been stored in the sensor data memory 24A for a predetermined period (step 58). The predetermined period is, for example, 6 days, but is not limited to this.

[0108] If the CPU 23 determines that sensor data has been accumulated for a predetermined period (Yes in step 58), it transmits the sensor data for that predetermined period to the server device 4 via the network 7 and the first relay unit 5A (step 59). The sensor data transmitted here is the presence / absence data before the above preprocessing, but it may also be the presence / absence data after the above preprocessing.

[0109] The CPU 23 repeatedly performs the above processing while power is supplied to the indoor unit 1 and the adapter 2. After the CPU 23 transmits the sensor data for the predetermined period to the server device 4, the AI ​​learning model is generated by the cloud-side learning unit 42A as described above in the flowchart shown in Figure 4.

[0110] Next, each time the latest AI learning model is generated in the cloud-side learning unit 42A of the server device 4, the CPU 23 receives the latest AI learning model from the server device 4. Now, we will explain the learning model update process performed on the air conditioner 100 side when there is both the latest AI learning model received by the CPU 23 and the online learning model updated by edge learning. Figure 8 is a flowchart showing the flow of the online learning model update process by the adapter 2. Figure 9 is a diagram illustrating the said update process.

[0111] As shown in Figure 8, the CPU 23 first determines whether or not it has received an AI learning model from the server device 4 (step 81). In this embodiment, a presence / absence prediction model is used as an example of the AI ​​learning model received from the server device 4. However, the present invention is not limited to this. For example, it may be a prediction model other than the presence / absence prediction model, such as a perceived temperature prediction model that predicts the user's perceived temperature. Therefore, if the AI ​​learning model downloaded from the server device 4 includes multiple prediction models (for example, a presence / absence prediction model and various other prediction models), the CPU 23 expands the multiple prediction models downloaded from the server device 4 into the model memory 24C, as shown in Figure 9.

[0112] In this embodiment, the CPU 23 extracts the presence / absence prediction model from the AI ​​learning models (multiple various prediction models) deployed in the model memory 24C.

[0113] If the CPU 23 determines that it has received an AI learning model (Yes in step 81), it compares the received AI learning model (presence / absence prediction model) with the previously learned model (previous presence / absence prediction model) stored in the model memory 24C (step 82, Figure 9).

[0114] Next, the CPU 23 determines from the above comparison results whether the received AI learning model and the previously learned model are inconsistent (step 83, Figure 9). Inconsistency between the two models is determined, for example, by whether the learned parameters (present / absent data) of both models are exactly the same. However, it may also be determined by whether the two parameters are substantially the same, for example, if the difference between the two parameters is within a predetermined threshold (e.g., a few percent).

[0115] If the presence / absence prediction model does not match the previous presence / absence prediction model (Yes in step 83), the CPU 23 determines whether the received presence / absence prediction model is available (step 84, Figure 9).

[0116] Then, if the presence / absence prediction model does not match the previous presence / absence prediction model and is deemed usable (Yes in step 84), the CPU 23 overwrites and updates the previous presence / absence prediction model and the online learning model in the model memory 24C with the received presence / absence prediction model (step 85, Figure 9). Subsequently, the CPU 23 executes presence / absence prediction processing using the overwritten online learning model and performs processes such as setting the timer time for starting air conditioning operation, and recommending additional function operation or operation suspension during the predicted absence period, based on the results.

[0117] On the other hand, if the CPU 23 determines that the learned parameters of the presence / absence prediction model match those of the previous presence / absence prediction model (No. in step 83), or if it determines that the AI ​​learning model is unavailable (No. in step 84), the CPU 23 does not overwrite the previous presence / absence prediction model and the online learning model in the model memory 24C with the received presence / absence prediction model. In other words, the CPU 23 maintains the previous presence / absence prediction model and the online learning model in the model memory 24C (step 86). Subsequently, the CPU 23 continues the presence / absence prediction processing using the maintained online learning model and performs tasks such as setting the timer time for starting air conditioning operation based on the results, and recommending additional function operation or operation suspension during the absence prediction period.

[0118] The CPU 23 repeatedly performs the above processing each time it downloads an AI learning model (e.g., a presence / absence prediction model) from the server device 4. This configuration allows the CPU 23 to store the previous learning model (e.g., the previous presence / absence prediction model) for comparison and determine whether the latest AI learning model matches the previous learning model (whether there is a difference or not). This avoids unnecessary update processing when there is no difference between the latest AI learning model and the previous learning model. Furthermore, it prevents the loss of additionally learned online learning models if the online learning model has been further trained by edge learning processing.

[0119] [Summary] As explained above, according to this embodiment, the air conditioner 100 can perform presence / absence prediction processing using the initial learning model until it receives the AI ​​learning model from the server device 4. This makes it possible for the user of the air conditioner 100 to receive the benefits of the learning model early on (for example, receiving recommendations for setting the timer time for starting air conditioning operation based on presence / absence prediction results, or receiving recommendations for operating additional functions or stopping operation during the predicted absence time). Furthermore, even if the initial learning model cannot make an appropriate prediction, edge learning processing is performed using sensor data from the air conditioner 100. As a result, the learned parameters of the initial learning model are updated to appropriate values, and the online learning model becomes able to make appropriate predictions. In addition, edge learning processing is performed using sensor data from the air conditioner 100 even after the AI ​​learning model has been received. Therefore, even if the period between receiving one AI learning model and receiving the next becomes longer, the learned parameters of the AI ​​learning model are further learned through edge learning processing, so the online learning model is updated as needed, and the online learning model can make appropriate predictions (for example, predicting the presence or absence of a user).

[0120] [Modifications] The present invention is not limited to the embodiments described above, and can be modified in various ways without departing from the spirit of the invention.

[0121] The data processing and generation intervals, such as the sensor data acquisition interval, the sensor data transmission interval to the server device 4, and the learning model training interval, as shown in the above-described embodiment are not limited to those described above and can be changed as appropriate. Furthermore, the various thresholds shown in the above-described embodiment can also be changed as appropriate.

[0122] In the embodiments described above, a presence / absence prediction model that predicts the presence or absence of people in an air-conditioned space was given as an example of a learning model. However, the prediction targets of the learning model are not limited to presence or absence. For example, the learning model may predict perceived temperature, temperature unevenness, building heat load, etc., and the initial learning model and AI learning model described above are generated and stored accordingly.

[0123] For example, a perceived temperature prediction model predicts the perceived temperature of a user in an air-conditioned space after a predetermined time (e.g., 5 minutes, 10 minutes, etc.), predicts the user's operating preferences based on the predicted perceived temperature, and is a learning model for controlling the air conditioner (setting temperature, etc.) based on those operating preferences. In this case, sensor data and external data such as temperature distribution, indoor temperature, setting temperature, outdoor temperature, timestamp, and cloud cover are used, and learned parameters are generated from these.

[0124] Furthermore, the building heat load prediction model is a learning model that predicts the heat load of a building after a predetermined time (e.g., one day) has elapsed, and plans and executes an operation plan so that the operation of the air conditioner becomes more efficient (energy-saving) according to the predicted heat load. In this case, sensor data and external data such as the maximum temperature, minimum temperature, average temperature, maximum humidity, minimum humidity, average humidity, rainfall, solar radiation, and average temperature of the air-conditioned space on the day before the prediction target day are used, and the predicted maximum temperature, predicted minimum temperature, rainfall, and predicted room temperature on the prediction target day are used, and learned parameters are generated from these.

[0125] Of the inventions described in the claims of this application, the invention described as "information processing method" is one in which each step is performed automatically by at least one device such as a computer through information processing by software, and not by a human using a computer or other device. In other words, the "information processing method" is an information processing method using computer software, and not a method in which a human operates a computer as a calculating tool.

[0126] 1...Indoor unit 2...Adapter 4...Server device 7...Network 21...First communication unit 22...Second communication unit 23...CPU 24...Memory unit 24C...Model memory (RAM) 24D...Flash memory 41...Communication unit 42...CPU 43...Memory unit 100...Air conditioner

Claims

1. An air conditioner equipped with an adapter for communicating with a server device, wherein the adapter includes: a storage unit for storing an initial learning model having learned parameters relating to a predetermined prediction process performed by the air conditioner; a communication unit for collecting sensor data detected by sensors of the air conditioner and transmitting it to the server device, and receiving a first learning model generated by the server device based on the sensor data; and a control unit for performing the predetermined prediction process using the initial learning model until the first learning model is received, and for performing the predetermined prediction process using the first learning model after the first learning model is received.

2. An air conditioner according to claim 1, wherein the control unit updates the learned parameters of the initial learning model based on the sensor data until the first learning model is received, and updates the learned parameters of the first learning model based on the sensor data after the first learning model is received.

3. An air conditioner according to claim 1 or 2, wherein the control unit predicts the presence or absence of a user of the air conditioner in the air-conditioned space using the initial learning model and the first learning model.

4. An air conditioner according to claim 3, wherein the initial learning model has a plurality of learned parameters to correspond to a plurality of presence or absence trends of the user in the air-conditioned space.

5. An air conditioner according to claim 4, wherein the plurality of learned parameters include a first learned parameter corresponding to an presence / absence pattern having a first absence time as the absence time in the air-conditioned space, and a second learned parameter corresponding to an presence / absence pattern having a second absence time shorter than the first absence time as the absence time in the air-conditioned space.

6. An air conditioner according to claim 1 or 2, wherein the control unit starts collecting the sensor data when external power is supplied to the air conditioner, and continues collecting the sensor data regardless of whether the air conditioning operation of the air conditioner is started or stopped.

7. An air conditioner according to claim 1 or 2, wherein the control unit, each time it receives the latest first learning model from the server device, compares the received first learning model with the previously received first learning model, and if the two models do not match, updates the previously received first learning model and the previously received first learning model with the latest first learning model, stores them in the storage unit, and performs the predetermined prediction process with the updated second learning model.

8. An air conditioner according to claim 2, wherein the sensor data is set to be collected at predetermined time intervals, has a timestamp indicating the time of collection, is stored in the order of the timestamps, and the control unit determines whether the timestamp of a predetermined number of sensor data from the sensor data having the predetermined timestamp falls within a predetermined time range before and after the timestamp that the sensor data to be acquired at the predetermined number according to the setting should have, and determines that the sensor data is missing if it is determined that the timestamp is outside the predetermined time range before and after.

Images