Model generation device, model generation program, and model generation method

Abstract

To allow a user to appropriately grasp the state of a controlled object controlled by a control system.SOLUTION: A model generation device 50 includes a teacher data acquisition unit 51B and a model generation unit 51C. The model generation device 50 acquires multiple teacher data using: room temperature T1 or T2, supply air temperature Tsa, supply air quantity Q1 or Q2, that are measured values related to which of states of "cold", "hot" and "other" air in an air conditioning zone controlled by an air conditioning system is in, as explanatory variables; and the states at the measured values related to "cold", "hot" and "other" as objective variables. The model generation unit 51C generates models M1 and M2 that derive, through statistical machine learning based on the multiple teacher data, respective probabilities of "cold", "hot", and "other" that the air in the air conditioning zone at input values using new measured values as the input values can take.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a model generation device, a model generation program, and a model generation method. [Background technology]

[0002] Among air conditioning systems, there is a VAV (Variable Air Volume) type air conditioning system that conditions the air in each air conditioning zone by controlling the supply air temperature and the total amount of air supplied to all air conditioning zones on the air conditioner side and controlling the amount of air supplied to each air conditioning zone using a VAV unit (Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2016-125688 Summary of the Invention [Problem to be solved by the invention]

[0004] The above-described air conditioning system is a type of control system that controls the air in an air-conditioning zone, and users such as control system administrators can deal with abnormalities in the control system by understanding which of multiple states the controlled object is in. Therefore, it is desirable for users to appropriately understand which of multiple states the controlled object is in.

[0005] As an example, a user such as a manager of the above-mentioned air conditioning system monitors the environmental (air) state of an air-conditioning zone by monitoring measurements of the indoor environment (room temperature, airflow, etc.). However, simply monitoring the measurements makes it difficult for the user to properly understand the environmental state of the air-conditioning zone as perceived by the occupants of the air-conditioning zone (the state as perceived subjectively by the occupants). As a result, the user such as a manager is unable to discover an abnormality in the environmental state of the air-conditioning zone, or the discovery is delayed, until a complaint is made by the occupants of the air-conditioning zone. As such, in conventional air conditioning systems, the user's inability to properly understand the environmental state of the air-conditioning zone can lead to inconveniences such as the user being unable to take action to address the abnormality.

[0006] The present invention has been made in view of the above points, and an object of the present invention is to allow a user to appropriately understand the state of a control target of a control system. [Means for solving the problem]

[0007] In order to solve the above problem, the model generation device of the present invention comprises a teacher data acquisition unit configured to acquire a plurality of teacher data sets in which a measurement value related to which of a plurality of states a controlled object controlled by a control system is in is used as an explanatory variable, and in which the state at the time of that measurement value is used as a target variable among the plurality of states; and a model generation unit configured to generate, by statistical machine learning based on the plurality of teacher data sets, a model that uses a new measurement value as an input value and derives the probability of each of the plurality of states that the controlled object can take at the input value,

[0008] The model generation device may further include a probability acquisition unit configured to input new measurement values into the model and acquire the probability of each of the plurality of states derived by the model, and a notification unit configured to notify at least the state with the highest probability among the plurality of states based on the probability of each of the plurality of states acquired by the probability acquisition unit.

[0009] The model generation device may further include a probability acquisition unit configured to input new measurement values into the model and acquire the probability of each of the plurality of states derived by the model, and an alarm unit configured to notify at least a portion of the probability of each of the plurality of states acquired by the probability acquisition unit.

[0010] The model generation device may further include a probability acquisition unit configured to input new measurement values into the model and acquire the probability of each of the plurality of states derived by the model, and an alarm unit that notifies the user when any of the probabilities of each of the plurality of states acquired by the probability acquisition unit exceeds a predetermined value.

[0011] The model may be a model using the following formula (A) of the Naive Bayes method. P(y|x)=P(x|y)*P(y) / P(x)...(A) x is the measured value and y is one of the plurality of states.

[0012] The P(y) may be set by the user.

[0013] The P(x|y) may be determined using a histogram showing the distribution of P(x|y) for each measurement value.

[0014] The teacher data acquisition unit may be configured to identify the state of the control object as the objective variable based on information input by a user to a control system that controls the control object.

[0015] The controlled object may be the air in an air conditioning zone, the control system may be an air conditioning system that controls the air, and the teacher data acquisition unit may be configured to set the objective variable of the teacher data to "cold" when a user increases the set temperature of the air conditioning system, and to set the objective variable of the teacher data to "hot" when a user decreases the set temperature of the air conditioning system.

[0016] The model generation method of the present invention includes a training data acquisition step configured to acquire a plurality of training data sets in which a measurement value related to which of a plurality of states a controlled object controlled by a control system is in is used as an explanatory variable and the state at the time of that measurement value is used as a target variable among the plurality of states; and a model generation step configured to generate, by statistical machine learning based on the plurality of training data sets, a model that uses the new measurement value as an input value and derives the probability of each of the plurality of states that the controlled object can take at the input value,

[0017] The model generation program of the present invention causes a computer to execute a training data acquisition step configured to acquire multiple training data sets using a measurement value related to which of multiple states a controlled object controlled by a control system is in as an explanatory variable and the state at the time of that measurement value as a target variable from among the multiple states; and a model generation step configured to generate, by statistical machine learning based on the multiple training data sets, a model that uses a new measurement value as an input value and derives the probability of each of the multiple states that the controlled object can take at the input value, using the new measurement value as an input value. [Effects of the Invention]

[0018] According to the present invention, it is possible to allow a user to appropriately understand the state of a control target of a control system. [Brief explanation of the drawings]

[0019] [Figure 1] FIG. 1 is a hardware configuration diagram of a model generating device and an air conditioning system according to an embodiment of the present invention. [Figure 2] FIG. 2 is a configuration diagram of the model generating device. [Figure 3] FIG. 3 is a diagram illustrating an example of the configuration of time-series data and teacher data. [Figure 4] FIG. 4 is a diagram illustrating an example of the configuration of time-series data and teacher data. [Figure 5] Figure 5 is a graph of the probability distribution of P(x|y) versus room temperature. [Figure 6] FIG. 6 is a graph of the probability distribution of P(x|y) versus supply air temperature. [Figure 7] FIG. 7 is a graph of the probability distribution of P(x|y) versus the air supply rate. [Figure 8] FIG. 8 is a table showing the values of P(x|y) derived by the model. [Figure 9] FIG. 9 is a diagram showing a notification image displayed on the display device. [Figure 10] FIG. 10 is a diagram showing a notification image displayed on the display device. DETAILED DESCRIPTION OF THE INVENTION

[0020] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0021] 1, a model generating device 50 of this embodiment is configured as a monitoring device that monitors an air conditioning system 10. Below, the air conditioning system 10 will be explained first, and then the model generating device 50 will be explained.

[0022] The air conditioning system 10 is configured as a VAV air conditioning system that controls the air in air conditioning zones Z1 and Z2 (more specifically, room temperatures T1 and T2, described below) individually.

[0023] The air conditioning system 10 includes an air conditioner 11, an air supply duct 12, VAV units 13A and 13B, an air supply temperature sensor 14, air flow meters 15A and 15B, room temperature sensors 16A and 16B, air conditioning operation panels 17A and 17B, VAV controllers 18A and 18B, and an air conditioner controller 19.

[0024] The air conditioner 11 is configured to supply conditioned air, which is cooled or heated air. The supply air duct 12 forms a flow path for the conditioned air that is supplied from the air conditioner 11 to air conditioning zones Z1 and Z2 via VAV units 13A and 13B, respectively. The VAV units 13A and 13B control the amount of conditioned air supplied to the air conditioning zones Z1 and Z, respectively. Each of the VAV units 13A and 13B is equipped with a damper and an actuator that controls the opening degree of the damper, and the amount of conditioned air supplied is controlled by the opening degree of the damper.

[0025] The supply air temperature sensor 14 measures the temperature of the conditioned air from the air conditioner 11 as the supply air temperature Tsa. The air flow meters 15A and 15B measure the flow rates of the conditioned air to the air conditioning zones Z1 and Z2, respectively, as supply air volumes Q1 and Q2. The supply air volumes Q1 and Q2 are expressed as a ratio to the maximum supply air volume of the air conditioning zone Z1 or Z2. The supply air volumes Q1 and Q2 are expressed as a flow rate (m 3 / h). The room temperature sensors 16A and 16B measure the room temperature T1 of the air conditioning zone Z1 and the room temperature T2 of the air conditioning zone Z2, respectively.

[0026] The air conditioning operation panel 17A is operated by a user U1 in the air conditioning zone Z1. A set temperature SP1, which is the target temperature for the room temperature T1 in the air conditioning zone Z1, is input into the air conditioning operation panel 17A by operation by the user U1 (for example, a tenant resident of the air conditioning zone Z1). The air conditioning operation panel 17B is operated by a user U2 in the air conditioning zone Z2 (for example, a tenant resident of the air conditioning zone Z2). A set temperature SP2, which is the target temperature for the room temperature T2 in the air conditioning zone Z2, is input into the air conditioning operation panel 17B by operation by the user U2.

[0027] Each of the VAV controllers 18A and 18B includes various computers such as a PLC (Programmable Logic Controller).

[0028] A set temperature SP1 input to air conditioning operation panel 17A is set in VAV controller 18A. Based on the deviation between set temperature SP1 and room temperature T1 measured by room temperature sensor 16A, VAV controller 18A calculates a required air supply volume such that room temperature T1 becomes set temperature SP1. VAV controller 18A transmits the required air supply volume to air conditioner controller 19, while controlling the opening of the damper of VAV unit 13A to ensure the required air supply volume. VAV controller 18A controls the opening of the damper based on the deviation between air supply volume Q1 measured by air flow meter 15A and the required air supply volume. VAV controller 18A periodically (e.g., every 10 minutes) supplies the current set temperature SP1, air supply volume Q1, and room temperature T1 to model generation device 50 via air conditioner controller 19.

[0029] Similarly, VAV controller 18B transmits the required air supply volume calculated based on the deviation between set temperature SP2 and room temperature T2 to air conditioner controller 19. Furthermore, VAV controller 18B controls the opening of the damper of VAV unit 13B based on the deviation between air supply volume Q2 and the required air supply volume so as to ensure the required air supply volume. VAV controller 18B periodically (for example, every 10 minutes) supplies the current set temperature SP2, air supply volume Q2, and room temperature T2 to model generation device 50 via air conditioner controller 19.

[0030] The air conditioner controller 19 is configured to include various computers such as a PLC and a personal computer. The air conditioner controller 19 controls the air conditioner 11 so that the temperature of the conditioned air supplied from the air conditioner 11, in this case the supply air temperature Tsa measured by the supply air temperature sensor 14, matches a preset supply air temperature set value SPsa. The supply air temperature set value SPsa is changed periodically according to a predetermined schedule. Furthermore, the air conditioner controller 19 controls the air conditioner 11 so that the conditioned air is blown at a supply air volume that ensures the required air supply volumes from each of the VAV controllers 18A and 18B. The air conditioner controller 19 periodically (for example, every 10 minutes) supplies the current supply air temperature Tsa to the model generation device 50.

[0031] The model generation device 50 is configured as a monitoring device that monitors the air conditioning system 10. The model generation device 50 generates models M1 and M2, which will be described later. The model generation device 50 is configured using various types of computers, such as a personal computer, a server computer, or a PLC. As shown in FIG. 2 , the model generation device 50 includes a processor 51, such as a central processing unit (CPU), a random access memory (RAM) 52 that functions as the main memory of the processor 51, and a storage device 53 that stores a model generation program executed by the processor 51 and training data, which will be described later. The storage device 53 is a non-volatile storage device, such as a hard disk or solid-state drive (SSD). The model generation device 50 further includes an operation device 54, which includes a keyboard, a mouse, a touch sensor, and the like, and a display device 55 that displays various images. The model generation device 50 also includes a communication module (not shown) that communicates with the air conditioner controller 19 via wired or wireless communication.

[0032] By executing the model generation program stored in the storage device 53, the processor 51 operates as an information collection unit 51A, a teacher data acquisition unit 51B, a model generation unit 51C, a probability acquisition unit 51D, and an alarm unit 51E shown in Figure 2.

[0033] The information collection unit 51A collects the room temperatures T1 and T2, supply air temperature Tsa, supply air rates Q1 and Q2, and set temperatures SP1 and SP2 that are periodically supplied from the air conditioner controller 19 and from the VAV controllers 18A and 18B via the air conditioner controller 19, and sequentially stores the collected information in the storage device 53. The information collection unit 51A associates this information with time information indicating the time of supply and sequentially stores the information in the storage device 53. At this time, the information collection unit 51A stores the information in the storage device 53 separately for air conditioning zone Z1 and air conditioning zone Z2. Therefore, this information is stored in the storage device 53 as time-series data for air conditioning zones Z1 and Z2 as shown in FIGS. 3 and 4. In the time-series data for air conditioning zone Z1, the room temperature T1, supply air temperature Tsa, supply air rate Q1, and set temperature SP1 are associated with each time. In the time series data for air conditioning zone Z2, room temperature T2, supply air temperature Tsa, supply air flow rate Q2, and set temperature SP2 are associated with each time.

[0034] The training data acquisition unit 51B shown in FIG. 2 acquires training data D1 and D2 based on the time-series data of air conditioning zones Z1 and Z2 stored in the storage device 53. As shown on the right side of FIG. 3, the training data D1 for air conditioning zone Z1 is data in which the measured values of room temperature T1, supply air temperature Tsa, and supply air volume Q1 are explanatory variables, and the state of the air in air conditioning zone Z1 (the object of control of air conditioning system 10) affected by these measured values is the objective variable (label). The air state as the objective variable is the subjective state that user U1 feels about this air (particularly room temperature T1). Examples of this state include "cold," "hot," and "other." Similarly, as shown on the right side of FIG. 4, the training data D2 for air conditioning zone Z2 is data in which the measured values of room temperature T2, supply air temperature Tsa, and supply air volume Q2 are explanatory variables, and the state of the air in air conditioning zone Z2 (the object of control of air conditioning system 10) affected by these measured values is the objective variable (label). Each record in one row shown in FIGS. 3 and 4 is one piece of training data D1 or D2.

[0035] When user U1 feels that the air in air-conditioning zone Z1 is "cold," user U1 is likely to operate air conditioning operation panel 17A to increase set temperature SP1. When user U1 feels that the air in air-conditioning zone Z1 is "hot," user U1 is likely to operate air conditioning operation panel 17A to decrease set temperature SP1. The same applies to set temperature SP2.

[0036] For the reasons described above, as shown in FIG. 3, the teacher data acquiring unit 51B labels the teacher data immediately before the set temperature SP1 or SP2 increases as "cold" as the objective variable. The teacher data acquiring unit 51B labels the teacher data immediately before the set temperature SP1 or SP2 decreases as "hot" as the objective variable. The teacher data acquiring unit 51B labels the other teacher data, that is, the teacher data in which the set temperature SP1 or SP2 does not change, as "other" as the objective variable. This is because it is difficult to understand how the user U1 or U2 is feeling.

[0037] The model generation unit 51C uses statistical machine learning with teacher data D1 to generate a model M1 for air conditioning zone Z1, which takes new measurement values (room temperature T1, supply air temperature Tsa, supply air volume Q1) as input values and derives and outputs the probability of each of the states that the air can take at these input values ("cold," "hot," and "other" as perceived by user U1). The model generation unit 51C uses statistical machine learning with teacher data D2 to generate a model M2 for air conditioning zone Z2, which takes new measurement values (room temperature T2, supply air temperature Tsa, supply air volume Q2) as input values and derives and outputs the probability of each of the states that the air can take at these input values ("cold," "hot," and "other" as perceived by user U1). Note that models M1 and M2 are collectively referred to as model M.

[0038] The generated model M calculates each probability using the naive Bayes method using the following equation (1). P(y|x)=P(x|y)*P(y) / P(x)...(1) Here, x is the measured value, and y indicates one of "cold," "hot," or "other." P(y) is the probability that the state will be y, and is also called the prior probability. P(x) is the probability that the measured value will be x, and is also called the marginal likelihood. P(x|y) is the conditional probability that the measured value will be x when the state is y, and is the likelihood. P(y|x) is the probability that the state will be y when the measured value is x, and is the posterior probability. The above formula (1) is prepared for each air conditioning zone.

[0039] Since P(x) is independent of the response variable, there is no need to calculate a specific value. Here, as will be described later, we normalize each P(x|y)*P(y) so that the sum of all y's, that is, P(y|x) for "cold," "hot," and "other," is 100%. This normalization takes P(x) into consideration.

[0040] P(y) is usually calculated using the component ratios of "cold," "hot," and "other." However, since "cold" and "hot" are only used when a change occurs in the set temperature SP1 or SP2, this component ratio is likely to deviate from the actual situation. Therefore, P(y) can be set by the user. The user sets an appropriate P(y) by operating the operation device 54. Here, P(y) is set equal for each state, that is, P(cold) = P(hot) = P(other) = 1 / 3.

[0041] P(x|y) is defined by statistical machine learning. Here, the model generation unit 51C uses statistical machine learning to approximate a set of training data with an appropriate probability distribution in the form of a histogram showing the probability (likelihood) for each measurement value, as shown in Figures 5 to 7. This approximated probability distribution represents P(x|y). The probability distribution is defined for each model and each measurement value. Figure 5 shows the probability distribution of the room temperature T1 used in model M1, Figure 6 shows the probability distribution of the supply air temperature Tsa used in model M1, and Figure 7 shows the probability distribution of the supply air volume Q1 used in model M1. The probability distributions are stored in and referenced by storage device 53.

[0042] After models M1 and M2 are generated, probability acquisition unit 51D in Fig. 2 acquires room temperature T1 measured by room temperature sensor 16A or the like, supply air temperature Tsa, and supply air volume Q1, which are supplied from air conditioner controller 19 or from VAV controller 18A via air conditioner controller 19. Probability acquisition unit 51D inputs these acquired values into model M1, and acquires the probabilities of "cold," "hot," and "other," which are the air conditions in air conditioning zone Z1 at the input values, output by model M1 in response to this input.

[0043] The operation of the model M1 will be described below assuming that the input values input to the model M1 are room temperature T1=26.0° C., Tsa=14.0° C., and a measured air volume of 40%.

[0044] Model M1 refers to the probability distribution shown in Figure 5 and obtains P(x|y) values closest to room temperature T1 = 26.0°C (including 26.0°C), i.e., P(x|cold), P(x|hot), P(x|other) (see dotted line area A). Model M1 refers to the probability distribution shown in Figure 6 and obtains P(x|y) values closest to Tsa = 14.0°C (including 14.0°C), i.e., P(x|cold), P(x|hot), P(x|other) (see dotted line area B). Model M1 refers to the probability distribution shown in Figure 7 and obtains P(x|y) values closest to Tsa = 40% (including 40%), i.e., P(x|cold), P(x|hot), P(x|other) (see dotted line area C). In this way, model M1 refers to the probability distribution that shows the distribution of probabilities (likelihoods) for each measurement value, and obtains P(x|y) corresponding to the measurement value in the probability distribution that is closest to the input value (including the measurement value with the same value as the input value).

[0045] Figure 8 shows the P(x|y) values obtained in this way for each input value and for each air condition in air conditioning zone Z1. Model M1 multiplies the three P(x|y) values for each measurement value for each air condition (cold, hot, or other) to calculate the P(x|y) values for "cold," "hot," or "other" for the entire model M1, to which the three measurement values are input. In other words, the P(x|cold), P(x|hot), and P(x|other) for the entire model M1 are calculated. Model M1 calculates P(x1|cold) (x1=[26.0°C, 14.0°C, 40%]; the same applies below) for the entire model M1 as follows: P(T1=26.0°C|cold)*P(Tsa=14.0°C|cold)*P(Q1=40%|cold)=0.011*0.274*0.073=0.00022. Similarly, model M1 calculates P(x1|hot) and P(x1|other) for the entire model M1.

[0046] Model M1 calculates P(y|x) for each state y, i.e., P(cold|x1), P(hot|x1), and P(other|x1), by substituting each of the calculated P(x1|cold), P(x1|hot), and P(x1|other) into P(x|y) in (1) above. Here, as mentioned above, P(y) is set to the same 1 / 3 for each state. P(x) is taken into account by normalizing P(x1|cold)*P(cold), P(x1|hot)*P(hot), and P(x1|other)*P(other) so that the sum of P(cold|x1), P(hot|x1), and P(other|x1) is 100%.

[0047] Model M1 calculates P(cold|x1), P(hot|x1), and P(other|x1) using the following formula: However, to normalize P(x1|y)*P(y), P(x1)=[P(x1|cold)*P(cold)+P(x1|hot)*P(hot)+P(x1|other)*P(other)] / 100. ·P(cold|x1)=P(x1|cold)*P(cold) / P(x1)=1% P(hot|x1)=P(x1|hot)*P(hot) / P(x1)=75% ·P(Other|x1)=P(x1|Other)*P(Other) / P(x1)=24%

[0048] Model M1 outputs each calculated probability to probability acquisition unit 51D. In this way, when newly measured values (in the above, room temperature T1=26.0°C, supply air temperature Tsa=14.0°C, supply air volume Q1=40%) are input, model M1 calculates each probability of "cold," "hot," or "other" that the air condition in air conditioning zone Z1 at this measurement value can take, and outputs these to probability acquisition unit 51D.

[0049] 2 acquires the room temperature T2 measured by the room temperature sensor 16B or the like, the supply air temperature Tsa, and the supply air volume Q2, which are supplied from the air conditioner controller 19 or from the VAV controller 18A via the air conditioner controller 19. The probability acquisition unit 51D inputs these acquired values into the model M2, and acquires the probabilities of the air state of "cold," "hot," and "other" in the air conditioning zone Z2 at the input values, which the model M2 outputs in response to this input. The operation of the model M2 is similar to the operation of the model M1 described above.

[0050] The notification unit 51E notifies the user of the most probable state among "cold," "hot," and "other" based on the probabilities of each of "cold," "hot," and "other" acquired by the probability acquisition unit 51D. For example, the notification unit 51E generates an image, such as that shown in FIG. 9, representing each state, such that the most probable state among the probabilities output from the model M1 is the air state of the air conditioning zone Z1, and the most probable state among the probabilities output from the model M2 is the air state of the air conditioning zone Z2, and displays the image on the display device 55. The room temperature and set temperature in FIG. 9 are, for example, room temperatures T1 and T2 and set temperatures SP1 and SP2 supplied from the VAV controllers 18A and 18B via the air conditioner controller 19.

[0051] The probability acquisition unit 51D may acquire the probabilities of the states of the air conditioning zones Z1 and Z2 using the models M1 and M2 each time the probability acquisition unit 51D periodically acquires the room temperatures T1 and T2, the supply air temperature Tsa, and the supply air volumes Q1 and Q2. The notification unit 51E may then update the air states of the air conditioning zones Z1 and Z2 displayed on the display device 55 each time the probability acquisition unit 51D acquires the probabilities. As a result, the latest air states are displayed on the display device 55.

[0052] The notification unit 51E may display on the display device 55 an image showing the change in each probability periodically acquired by the probability acquisition unit 51D as described above. Specifically, as shown in FIG. 10 , the notification unit 51E may display on the display device 55 a graph showing the time-series data of the room temperature T1 and the set temperature SP1 in the air-conditioning zone Z1 and a graph showing the time-series data of the probabilities of "cold" and "hot" output from the model M1, with the time axes associated with each other. As shown in FIG. 10 , the probability of "cold" is high before the set temperature SP1 increases, that is, before the user U1 feels cold and increases the set temperature SP1. Furthermore, the probability of "hot" is high before the set temperature SP1 decreases, that is, before the user U1 feels hot and decreases the set temperature SP1. Therefore, the probabilities output by the model M1 appropriately reflect the "cold" and "hot" sensations perceived by the user U1, i.e., the air condition in the air-conditioning zone Z1. The notification unit 51E may also display an image similar to that shown in FIG. 10 for the air conditioning zone Z2.

[0053] The notification unit 51E may issue a notification when any one of the probabilities of "cold," "hot," and "other" acquired by the probability acquisition unit 51D (particularly, any one of the probabilities of "cold" and "hot") exceeds a predetermined value. For example, the notification unit 51E may monitor the probabilities acquired by the probability acquisition unit 51D, and when the probability of "cold" for air conditioning zone Z1 exceeds a predetermined value, the notification unit 51E may display a message image such as "Air conditioning zone Z1 may be cold" on the display device 55.

[0054] The notification unit 51E may make the above notification by outputting a warning sound from a speaker or the like, or by turning on a warning lamp.

[0055] The manager of the air conditioning system 10 performs a predetermined task in response to the notification. For example, if the set temperature of an air conditioning zone with a high probability of being "cold" or "hot" is set to "automatic," the manager changes the set temperature. The setting change is performed by operating the operation device 54 of the model generation device 50. The changed set temperature input to the model generation device 50 is set in the VAV controller 18A or 18B via the air conditioner controller 19. For example, if there is an air conditioning zone with a high probability of being "cold" or "hot," the manager may change various settings of the air conditioning system 10 (such as the supply air temperature set value SPsa and the damper opening of the VAV unit 13A or 13B). This change is also performed by operating the operation device 54. If a state with a high probability of being "cold" or "hot" occurs frequently, the manager of the air conditioning system 10 may perform maintenance on the entire air conditioning system 10.

[0056] As described above, in this embodiment, models M1 and M2 are generated by statistical machine learning using training data D1 or D2, which uses measured values such as room temperature T1 as explanatory variables and the air condition in air-conditioning zone Z1 as a target variable. These models M1 and M2 derive the probability of each state ("cold," "hot," "other") that the air in each of air-conditioning zones Z1 and Z2 can assume for the measured values used as input values. By checking these probabilities or information output from the model generation device based on these probabilities, the user as the administrator of the air-conditioning system 10 can grasp the probability of the air condition in each of air-conditioning zones Z1 and Z2, and appropriately grasp the air condition.

[0057] The above example illustrates a case with two air conditioning zones, but in a VAV air conditioning system applied to a large building, the number of air conditioning zones becomes enormous. In this case, it is difficult for the manager to grasp the air condition of each air conditioning zone on an ongoing basis. As in this embodiment, a model is generated for each air conditioning zone, the probability of the state of each air conditioning zone is calculated, and notification is performed by the notification unit 51E, thereby reducing the monitoring burden on the manager.

[0058] Because users U1 and U2's perception of "cold" or "hot" is influenced by their personal temperature sensitivity, it is difficult for an administrator to address "cold" or "hot" simply by monitoring measurements such as room temperature. In this embodiment, training data in which the objective variable (label) for the measurement value when users U1 and U2 change the set temperature is "cold" or "hot" is used for machine learning. This allows the probabilities output by models M1 and M2 to reflect users U1 and U2's personal temperature sensitivity, i.e., circumstances that are difficult to grasp from measurement values, allowing the administrator to properly grasp the air condition in air-conditioning zones Z1 and Z2, in this case, the user's subjective state.

[0059] Conventionally, a manager would only become aware that the air in air-conditioning zone Z1 or Z2 was not of a good quality when a complaint or the like was received from user U1 or U2 in air-conditioning zone Z1 or Z2. In this embodiment, models M1 and M2 allow the manager to recognize and deal with air problems in air-conditioning zone Z1 or Z2 before a complaint or the like is received. In addition, the manager can also deal with undesirable conditions (conditions that user U1 or U2 is putting up with) that have not yet resulted in a complaint.

[0060] When a complaint is made by user U1 or U2, training data may be used for the machine learning, with the content of the complaint ("cold" or "hot") as the objective variable and the measurement value at the time of the complaint as the explanatory variable. However, since complaints do not occur frequently, the amount of training data based on the complaint is usually small. Even if machine learning is performed using only the training data, it is difficult to secure a sufficient amount of training data for the machine learning. In this embodiment, the objective variable is set by focusing on the fact that when the set temperature SP1 or SP2 is changed, it can be estimated that user U1 or U2 feels "cold" or "hot," and such changes occur more frequently than the complaints. Therefore, the amount of training data required for machine learning can be easily obtained.

[0061] In this embodiment, models M1 and M2 are models using the above formula (1) of the Naive Bayes method, "P(y|x) = P(x|y) * P(y) / P(x)," and therefore the accuracy of each output probability is high. Furthermore, by allowing a user, such as an administrator, to set P(y) in the above formula (1), a model tailored to the circumstances of each individual air conditioning system 10 can be obtained. Note that P(y) may be different between models M1 and M2. Increasing P(cold) increases the probability that model M1 or M2 will output "cold" (the same applies to "hot" and "other"). In this embodiment, P(x|y) is determined using a histogram (see FIGS. 5 to 7) showing the distribution of P(x|y) for each measured value. In this embodiment, since the distribution of P(x|y) for each measured value is not a normal distribution, the use of the histogram improves the accuracy of determining P(x|y).

[0062] (Variation) The above-described embodiment can be modified in various ways. Modifications are listed below, and at least some of the modifications can be combined as appropriate.

[0063] (Variation 1) The measured values included in the training data D1 and D2 and input to the models M1 and M2 are selected arbitrarily. The measured values may be any measured value related to which of a plurality of states ("cold," "hot," "other") the air conditioning zone Z1 or Z2 is in. For example, humidity may be used as the measured value. Measured values include physical quantities fed back in feedback control, but manipulated variables, which are other physical quantities used in this feedback control, are also treated here as a type of measured value. In other words, the measured value is a concept that includes manipulated variables.

[0064] (Variation 2) The model generation device 50 may be a device that generates a model for a control system other than the air conditioning system 10. The training data acquisition unit 51B may acquire multiple training data sets using measurement values (including values obtained by taking a moving average) related to which of multiple states the controlled object controlled by the control system is in as explanatory variables and the state at the time of the measurement value as the objective variable. The model generation unit 51C may generate a model using statistical machine learning based on the multiple training data sets, in which a new measurement value is used as an input value and the probability of each of the multiple states the controlled object can take when the input value is the new measurement value. The other control system may be a control system used in various plants. The other control system may be a heating system that heats steel in a heating furnace. In this case, the controlled object is the temperature inside the heating furnace. The measurement value is process data such as the flame intensity of a burner that heats the heating furnace. The controlled object can be in one of the following states: "high," "low," or "optimum temperature," and these may be labeled by an administrator when generating the training data.

[0065] The teacher data acquisition unit 51B may identify the state of the controlled object as the objective variable of the teacher data (e.g., the above-mentioned "cold," "hot," or "other," or the above-mentioned "high," "low," or "optimum temperature") based on information input by the user to the control system that controls the controlled object (e.g., the set temperature SP1 or SP2 input by the user, or information input by the user to distinguish between "high," "low," and "optimum temperature" in the heating system). This generates a model that outputs the above-mentioned probability reflecting the user's perception.

[0066] (Variation 3) The hardware configuration of the model generating device 50 is arbitrary. At least a portion of each of the units 51A to 51E may be configured with one or more logic circuits. Examples of logic circuits include an ASIC (Application Specific Integrated Circuit) and an FPGA (Field Programmable Gate Array). The various data stored in the storage device 53 may be stored in another storage device installed outside the model generating device 50. The model generating program may be stored in a non-transitory storage medium other than the storage device 53. The device and system in this specification may be a single device in which various components are housed in a single housing, or may be a combination of multiple devices in which various components are housed in multiple separate housings.

[0067] (Scope of the present invention) Although the present invention has been described above with reference to the embodiments and modifications, the present invention is not limited to the above embodiments and modifications. For example, the present invention includes various modifications to the above embodiments and modifications that can be understood by a person skilled in the art within the scope of the technical concept of the present invention. The configurations listed in the above embodiments and modifications can be combined as appropriate within a range that does not contradict each other. [Explanation of symbols]

[0068] 10...Air conditioning system, 11...Air conditioner, 12...Air supply duct, 13A, 13B...VAV unit, 14...Air supply temperature sensor, 15A, 15B...Air volume meter, 16A, 16B...Room temperature sensor, 17A, 17B...Air conditioning operation panel, 18A, 18B...VAV controller, 19...Air conditioner controller, 50...Model generation device, 51...Processor, 51A...Information collection unit, 51B...Teaching data acquisition unit, 51C...Model generation unit, 51D...Probability acquisition unit, 51E...Alarm unit, 53...Memory device, 54...Operation device, 55...Display device, M1, M2...Model, U1, U2...User, Z1, Z2...Air conditioning zone.

Claims

1. A teacher data acquisition unit configured to acquire a plurality of teacher data sets using a measurement value relating to which of a plurality of states the air in an air conditioning zone controlled by an air conditioning system is in as an explanatory variable, and a state of the plurality of states at the time of measurement of the measurement value as a target variable; a model generation unit configured to generate a model by statistical machine learning based on the plurality of teacher data, the model deriving a probability of each of the plurality of states that the control object can take when a new measurement value is an input value, the measured values of each of the plurality of teacher data are measured in chronological order to form time-series data; The teacher data acquisition unit Among the plurality of teacher data, the objective variable of the teacher data in which the measurement value measured immediately before the user inputs an increase in the set temperature to the air conditioning system is set as “cold” is set as an explanatory variable, Among the plurality of teacher data, the objective variable of teacher data in which the measurement value measured immediately before the user inputs a value to the air conditioning system to lower the set temperature is used as an explanatory variable is set to "hot." Model generation device.

2. a probability acquisition unit configured to input new measurement values to the model and acquire probabilities of each of the plurality of states derived by the model; a notification unit configured to notify at least the state with the highest probability among the plurality of states based on the probabilities of the respective states acquired by the probability acquisition unit; and The model generating device of claim 1 further comprising:

3. a probability acquisition unit configured to input new measurement values to the model and acquire probabilities of each of the plurality of states derived by the model; a notification unit configured to notify at least a part of the probabilities of the respective states acquired by the probability acquisition unit; The model generating device according to claim 1 or 2, further comprising:

4. a probability acquisition unit configured to input new measurement values to the model and acquire probabilities of each of the plurality of states derived by the model; a notification unit that notifies a user when any of the probabilities of the plurality of states acquired by the probability acquisition unit exceeds a predetermined value; The model generating device according to claim 1 , further comprising:

5. The model is a model using the following formula (A) of the Naive Bayes method: The model generating device according to any one of claims 1 to 4. P(y|x)=P(x|y)*P(y) / P(x)...(A) x is the measured value and y is one of the plurality of states.

6. The P(y) can be set by the user. The model generating device according to claim 5 .

7. The P(x|y) is determined using a histogram showing the distribution of P(x|y) for each measurement value.

7. The model generating device according to claim 5 or 6.

8. A teacher data acquisition step configured to acquire multiple teacher data sets using a measurement value relating to which of multiple states the air in an air conditioning zone controlled by an air conditioning system is in as an explanatory variable, and using the state of the multiple states at the time the measurement value is measured as a target variable; a model generation step configured to generate a model by statistical machine learning based on the plurality of teacher data, the model deriving a probability of each of the plurality of states that the control object can take when the new measurement value is an input value; 1. A model generation method comprising: the measured values of each of the plurality of teacher data are measured in chronological order to form time-series data; The teacher data acquisition step includes: Among the plurality of teacher data, the objective variable of the teacher data in which the measurement value measured immediately before the user inputs an increase in the set temperature to the air conditioning system is set as “cold” is set as an explanatory variable, Among the plurality of teacher data, the objective variable of teacher data in which the measurement value measured immediately before the user inputs a value to the air conditioning system to lower the set temperature is used as an explanatory variable is set to "hot." Model generation method.

9. On the computer, a teacher data acquisition step configured to acquire a plurality of teacher data sets using a measurement value relating to which of a plurality of states the air in an air-conditioning zone controlled by the air-conditioning system is in as an explanatory variable, and using a state of the plurality of states at the measurement timing of the measurement value as a response variable; a model generation step configured to generate a model by statistical machine learning based on the plurality of teacher data, the model deriving a probability of each of the plurality of states that the control object can take when the new measurement value is an input value; A model generation program that executes the measured values of each of the plurality of teacher data are measured in chronological order to form time-series data; The teacher data acquisition step includes: Among the plurality of teacher data, the objective variable of the teacher data in which the measurement value measured immediately before the user inputs an increase in the set temperature to the air conditioning system is set as “cold” is set as an explanatory variable, Among the plurality of teacher data, the objective variable of teacher data in which the measurement value measured immediately before the user inputs a value to the air conditioning system to lower the set temperature is used as an explanatory variable is set to "hot." Model generator.

Images