JP2010055303A - Learning data management device, learning data management method and air-conditioner for vehicle, and control device of apparatus - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a learning data management device and learning data management method which select and store only the data which are significant for the learning of a machine learning system. <P>SOLUTION: The learning data management device includes: a storage part 51 for storing the set of state information related with control target equipment and prescribed setting, with respect to the equipment as learning data; a conviction calculating part 55 for, when new state information is acquired by a state information acquisition part 3, calculating conviction showing the certainty of the degree of recommendation of the prescribed setting of control target equipment, obtained by inputting the new state information to a machine learning system from the number of learning data relevant to the prescribed setting from among learning data stored in the storage part 51; and a data storage determining part 56, which when the conviction satisfies a prescribed reference showing that it is necessary to learn the machine learning system, stores the set of the new state information and the prescribed setting as learning data to the storage part 51, and which when the conviction does not satisfy the prescribed reference, abandons the new state information. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a learning data management device, a learning data management method, a vehicle air conditioner, and a device control device, and more particularly to learning of a machine learning system that determines the setting of a control target device using state information corresponding to a specific situation. The present invention relates to a learning data management apparatus that accumulates learning data to be used, a learning data management method, a vehicle air conditioner that uses them, and a device control apparatus.

  In recent years, research on automatic control devices using a machine learning system using a neural network or a Bayesian network has been actively conducted. In such a machine learning system, in order to appropriately control the target device, the correspondence between the input parameter values corresponding to various assumed situations and the various control values to be selected by the system. It is very important to learn a machine learning system using a data set that accurately represents the relationship. Moreover, it is desirable that the learning data included in the set is distributed as much as possible with respect to various assumed situations.

  In addition, an automatic control device using a machine learning system periodically acquires, as learning data, signals for operating a device to be controlled and detection signals from various sensors provided in association with the device. Some learn machine learning systems every time learning data is acquired. Since the machine learning system that learns dynamically as described above learns according to the operation of the user, there is an advantage that automatic control reflecting the user's preference is possible. However, in such a machine learning system, an appropriate set of learning data cannot be prepared in advance, and thus the learning data included in the set may be unevenly distributed. When a machine learning system is learned using such a set, an automatic control device equipped with the learned machine learning system may perform inappropriate control that does not match user preferences. In particular, when all operation signals and detection signals acquired periodically are used as learning data, a large number of learning data are stored in a steady state (that is, no operation is performed or only a routine operation is performed). The acquired learning data has a dominant influence on the learning of the machine learning system. Therefore, a machine learning system learned using such a collection of learning data may not be able to quickly respond to changes in the external environment or changes in the user's intention. Therefore, in order to cope with such a problem, a forgetting-type histogram calculation device has been proposed that learns by reducing the weight of acquired data with older time (see, for example, Patent Document 1).

JP 2007-18530 A

  The forgetting-type histogram calculation apparatus disclosed in Patent Document 1 multiplies all learning data by a weight for forgetting in the same manner. Therefore, the forgetting-type histogram calculation apparatus erases important learning data corresponding to a specific operation acquired at a certain time (for example, learning data corresponding to an operation that the user performs only occasionally) as time elapses. Resulting in. Therefore, there is a possibility that the learning data corresponding to the specific operation necessary for appropriately learning the machine learning system cannot be accumulated.

  Accordingly, an object of the present invention is to provide a learning data management apparatus and a learning data management method that can select and store only data important for learning of a machine learning system.

  Another object of the present invention is to provide a vehicle air conditioner and device that can automatically execute an appropriate operation for a specific situation by using such a learning data management device or learning data management method. It is to provide a control device.

  According to the description of claim 1, as one aspect of the present invention, in order to automatically control the device (2), at least one state information on the device (2) acquired by the state information acquisition unit (3) Is provided, a learning data management device for accumulating learning data used for learning of a machine learning system for obtaining a recommendation degree for setting the device (2) to a predetermined setting is provided. The learning data management apparatus includes a storage unit (51) that accumulates a set of state information and a predetermined setting as learning data, and a storage unit (51) when new state information is acquired by the state information acquisition unit (3). ) Represents the probability of the recommended degree of the predetermined setting obtained by inputting the new state information to the machine learning system from the number of learning data related to the predetermined setting When the certainty factor calculation unit (55) for calculating the certainty factor and a certain criterion indicating that the certainty factor needs to learn the machine learning system are satisfied, a set of new state information and a predetermined setting is used as learning data. A data accumulation determining unit (56) that accumulates in the storage unit (51) and rejects new state information when the certainty factor does not satisfy a predetermined criterion. By having such a configuration, the learning data management apparatus according to the present invention can select only the state information when the learning of the machine learning system related to the predetermined setting item is insufficient, and accumulate it as learning data. .

According to the second aspect of the present invention, the certainty factor calculation unit (55) uses a function that increases the certainty factor as the number of learning data corresponding to the state information and the predetermined setting or the recommendation factor increases. It is preferable to determine the degree.
In this case, according to the description of claim 3, the function receives the number of learning data related to the predetermined setting and the recommendation level as inputs, and outputs a belief for the predetermined value of the recommendation level as a beta distribution or It preferably includes a Dirichlet distribution.
As a result, the confidence level is high when the machine learning system is sufficiently learned, and the confidence level is low when learning of the machine learning system is insufficient. It is possible to appropriately calculate a certainty factor that is a criterion for determining whether or not to accumulate.

  According to the fourth aspect of the present invention, the data accumulation determining unit (56) determines the latest certainty factor obtained with respect to the recommended degree of predetermined setting and the previous certainty factor obtained with respect to the recommended degree of predetermined setting. When the absolute value of the difference from the degree is larger than the first threshold corresponding to the change in the recommendation degree of the predetermined setting by learning the machine learning system, it is preferable to determine that the predetermined criterion is satisfied. . By having such a configuration, the learning data management device can appropriately determine whether or not state information should be accumulated as learning data.

  Alternatively, according to the fifth aspect, the data accumulation determining unit (56) uses the second threshold value corresponding to the latest degree of certainty obtained with respect to the predetermined setting recommendation degree that the recommendation degree is appropriate. In the following cases, it is preferable to determine that a predetermined standard is satisfied. By having such a configuration, the learning data management device can appropriately determine whether or not state information should be accumulated as learning data.

  According to a sixth aspect of the present invention, a device control apparatus is provided as another embodiment of the present invention. The control device includes a state information acquisition unit (3) that acquires at least one state information regarding the device (2) to be controlled, and inputs at least one state information to the machine learning system, thereby the device (2). An inference unit (53) for calculating a recommendation degree for setting the predetermined value to the predetermined setting, a control unit (54) for controlling the device (2) according to the recommendation level, and a memory for storing a set of state information and a predetermined setting as learning data When new state information is acquired by the unit (51) and the state information acquisition unit (3), among the learning data stored in the storage unit (51), the number of learning data related to the predetermined setting , A certainty factor calculation unit (55) for calculating a certainty factor representing the certainty of the recommended degree of predetermined setting obtained by inputting the new state information into the machine learning system, and the certainty factor learning the machine learning system Must A set of new state information and a predetermined setting is stored in the storage unit (51) as learning data when the predetermined standard indicating that there is a certain state, and a new state is obtained when the certainty factor does not satisfy the predetermined standard A data accumulation determination unit (56) for rejecting information and a learning unit (57) for learning the machine learning system using the learning data accumulated in the storage unit (51). By having such a configuration, the control device according to the present invention can select only the state information when learning of the machine learning system related to the predetermined setting item is insufficient, and accumulate it as learning data. Therefore, since the control apparatus can learn a machine learning system using an appropriate learning data set, it can automatically execute an appropriate setting operation for a specific situation.

  Moreover, according to Claim 7, the vehicle air conditioner is provided as another form of this invention. The air conditioner includes an air conditioning unit (2) that supplies conditioned air into the vehicle, a state information acquisition unit (3) that acquires at least one state information about the vehicle, and inputs at least one state information to the machine learning system. By doing so, an inference unit (53) for calculating a recommendation level for setting the air conditioning unit (2) to a predetermined setting, an air conditioning control unit (54) for controlling the air conditioning unit (2) according to the recommendation level, state information and a predetermined level Among the learning data stored in the storage unit (51), when new state information is acquired by the storage unit (51) and the state information acquisition unit (3), A certainty factor calculation unit (55) that calculates a certainty factor representing the certainty of the recommended degree of the predetermined setting obtained by inputting the new state information to the machine learning system from the number of learning data related to the predetermined setting. ) And confidence When a predetermined criterion indicating that it is necessary to learn the machine learning system is satisfied, a set of the new state information and a predetermined setting is accumulated as learning data in the storage unit (51), and the certainty factor is the predetermined criterion. When not satisfying, it has a data accumulation judgment part (56) which rejects new state information, and a learning part (56) which learns a machine learning system using learning data accumulated in storage part (51). The vehicle air conditioner according to the present invention can select only the state information when the learning of the machine learning system related to the predetermined setting item is insufficient, and accumulate it as learning data. Therefore, since the vehicle air conditioner can learn the machine learning system using an appropriate learning data set, an appropriate setting operation can be automatically executed for a specific situation.

  Furthermore, according to the description of claim 8, as still another aspect of the present invention, in order to automatically control the device (2), at least one of the device (2) acquired by the state information acquisition unit (3) By inputting the state information, a learning data management method is provided in which learning data used for learning of a machine learning system for obtaining a recommendation degree for setting the device (2) to a predetermined setting is accumulated in the storage unit (51). In the learning data management method, when new state information is acquired by the state information acquisition unit (3), learning data related to a predetermined setting among learning data already stored in the storage unit (51) is stored. From the number, a step of calculating a certainty factor representing the certainty of the recommended degree of the predetermined setting obtained by inputting the new state information to the machine learning system, and the certainty factor needs to learn the machine learning system. A set of new state information and a predetermined setting is stored in the storage unit (51) as new learning data, and a new state is obtained when the certainty factor does not satisfy the predetermined standard. Rejecting the information. The learning data management method according to the present invention can select only the state information when the learning of the machine learning system related to the predetermined setting item is insufficient, and accumulate it as learning data.

  In addition, the code | symbol in the parenthesis attached | subjected to each said part is an example which shows a corresponding relationship with the specific means as described in embodiment mentioned later.

Hereinafter, an example in which the learning data management device according to the present invention is applied to automatic control of an in-vehicle air conditioner will be described.
An air conditioner according to an embodiment of the present invention automatically corrects a setting according to a surrounding situation using a machine learning system that obtains a recommendation degree related to a specific setting operation. When the air conditioner accumulates state information representing the surrounding situation as learning data used for learning of the machine learning system, a sufficient number of samples are obtained from the state information that can be accumulated as learning data. Reject items and store only those with a small number of samples.

  FIG. 1 is a schematic configuration diagram showing an overall configuration of an air conditioner 1 according to one embodiment of the present invention. As shown in FIG. 1, an air conditioner 1 includes an air conditioning unit 2 mainly composed of a mechanical configuration, an information acquisition unit 3 for acquiring state information about a vehicle, an operation panel 4 functioning as an operation unit, an air conditioner A control unit 5 that controls each unit of the apparatus 1 is provided.

The air conditioning unit 2 cools or warms the air in the vehicle or the air taken from outside the vehicle, and supplies the air to the vehicle. For this purpose, the air conditioning unit 2 includes a refrigeration cycle (for example, a compressor, a receiver, an expansion valve, etc.) for cooling the refrigerant, an intake port and a blower fan for taking in air from inside or outside the vehicle, To obtain conditioned air by adjusting an evaporator for exchanging heat between the heated air and the refrigerant, a heater core for heating the taken-in air, and a mixing ratio of air passing through the heater core and air bypassing the heater core And an air outlet for sending conditioned air into the vehicle.
In addition, since the well-known various structure used for a vehicle-mounted air conditioner can be employ | adopted as the air-conditioning part 2, detailed description of the structure of the air-conditioning part 2 is abbreviate | omitted here.

  The information acquisition unit 3 includes at least one sensor that acquires various state information about the vehicle. In the present embodiment, there are an inside air temperature sensor, an outside air temperature sensor, and a solar radiation sensor as typical sensors of the information acquisition unit 3. The inside air temperature sensor is installed together with an aspirator on an instrument panel or the like in the vicinity of the steering wheel in order to measure the temperature (inside air temperature) Tr in the passenger compartment. The outside air temperature sensor is installed on a radiator grill in front of the vehicle in order to measure a temperature outside the vehicle compartment (outside air temperature) Tam. Furthermore, in order to measure the intensity (intensity of solar radiation) S of the sunlight shining into the vehicle interior, a solar radiation sensor is attached in the vicinity of the windshield in the vehicle interior.

Furthermore, the information acquisition unit 3 acquires, in the vehicle interior, one or more in-vehicle cameras for photographing the face of the humidity sensor, the driver, and the passenger, an in-vehicle camera for photographing a state outside the vehicle, and biometric information of the occupant. You may have a body temperature sensor or an exhaust gas sensor for measuring the odor of exhaust gas. Furthermore, the information acquisition unit 3 may include an in-vehicle device such as a navigation system as a sensor. For example, the navigation system may acquire position information such as the current position of the vehicle, the traveling direction, the surrounding area information, and Gbook information as the state information. The in-vehicle device may acquire time information such as a day of the week and the current time as state information. Furthermore, the in-vehicle device acquires various operation information such as accelerator opening, steering wheel, brake, power window opening, wiper, turn lever or car audio ON / OFF, vehicle speed, vehicle behavior information, etc. as status information. Also good.
The information acquisition unit 3 acquires one or more pieces of state information periodically (for example, every second, every 4 seconds, every minute, etc.). Or the information acquisition part 3 acquires one or more state information according to the information acquisition request from the control part 5, when a user performs setting operation with respect to an air conditioner. Then, the information acquisition unit 3 passes the acquired state information to the control unit 5.

  The operation panel 4 is an operation unit for performing a setting operation for an air conditioner to be automatically controlled. For example, various switches for adjusting setting information of the air conditioner 1 and a display for displaying the setting information. It has a display part and the like. Then, the operation panel 4 performs an operation in which the user changes any setting to the air conditioner (for example, changes the set temperature, adjusts the air volume or the wind direction, sets the inside air circulation mode or the outside air introduction mode, etc.). When this is detected, a signal representing the operation content is transmitted to the control unit 5. For example, the setting information includes a set temperature Tset in the vehicle, an air volume W, an intake air setting mode (inside air circulation mode or outside air introduction mode), a wind direction setting, and the like.

FIG. 2 is a functional block diagram of the control unit 5 of the air conditioner 1.
The control unit 5 includes a storage unit 51 including one or a plurality of microcomputers (not shown) including a CPU, ROM, RAM, and the like (not shown) and their peripheral circuits, and an electrically rewritable nonvolatile memory. And the communication part 52 which communicates with each sensor of the information acquisition part 3, etc. according to vehicle-mounted communication standards like a control area network (CAN). The storage unit 51 stores a program executed by the control unit 5, various setting parameters used by the program, and the like. The storage unit 51 stores data representing the structure of the machine learning system and a learning data set used for learning of the machine learning system.

Further, the control unit 5 includes an inference unit 53, an air conditioning control unit 54, a certainty factor calculation unit 55, a data accumulation determination unit 56, and a learning unit 57 as functional modules realized by a computer program executed on the microcomputer. Have.
Hereinafter, the operation of the control unit 5 will be described in detail.

  The inference unit 53 calculates a recommendation degree that represents the probability of the setting that is considered to be optimum for the air conditioner 1 according to a specific situation, based on the machine learning system. In this embodiment, a probability model based on a Bayesian network is used as the machine learning system. A Bayesian network models a probabilistic causal relationship between a plurality of events, and is a network represented by an acyclic directed graph in which propagation between nodes is obtained with a conditional probability. For details on the Bayesian network, see Yoichi Motomura, Hirotoshi Iwasaki, “Bayesian Network Technology”, first edition, Denki University Press, July 2006, Kazuo Shigeki, et al., “Outline of Bayesian Network,” July 2006, or supervised by Morio Onoe, “Pattern Identification”, first edition, New Technology Communications, July 2001, etc.

FIG. 3 shows an example of the probability model used in this embodiment. The probability model 300 shown in FIG. 3 is used for automatic control of the air conditioner 1, and based on the position and time of the vehicle, the degree of recommendation for setting the air conditioner 1 to either the inside air circulation mode or the outside air introduction mode is used as a probability. Output. For this purpose, the probability model 300 has two input nodes 301 and 302 and one output node 303, and each input node is connected to the output node 303. Further, the time zone (x ₁ ) and the current position (x ₂ ) are given to the input nodes 301 and 302 as state information to be input. Then, the output node 303 outputs the probability P (x ₃ = True | x ₁ , x ₂ ) for setting the outside air introduction mode and the probability P (x ₃ = False | x ₁ , x ₂ ) for setting the inside air circulation mode. To do. Note that P (x ₃ = False | x ₁ , x ₂ ) is equal to (1-P (x ₃ = True | x ₁ , x ₂ )).

  Conditional probability tables (hereinafter referred to as CPT) 311 and 312 shown in FIG. 3 correspond to the input nodes 301 and 302, respectively, and define the prior probabilities output by the input nodes 301 and 302 for the input state information. To do. The CPT 313 defines the probability that the output node 303 outputs corresponding to the output node 303 as a conditional probability distribution assigned to each set of state information values of each input node.

Here, if the time zone is morning (x ₁ = 1), the current position is near the user's home (x ₂ = 1) and all the information given to each input node is known, the air conditioner is set to the outside air introduction mode The probability P (x ₃ = True | x ₁ = 1, x ₂ = 1) and the probability P (x ₃ = False | x ₁ = 1, x ₂ = 1) to set the inside air circulation mode are determined from the CPT 313, respectively. 0.95 and 0.05. On the other hand, when the time zone is night (x ₁ = 2) and the current position is on the national road (x ₂ = 3), the probability P (x ₃ = True | x ₁ = 2, x ₂ = 3) and the probability P (x ₃ = False | x ₁ = 2, x ₂ = 3) for setting the inside-air circulation mode are 0.33 and 0.67 from the CPT 313, respectively. When the time zone or the current position cannot be acquired for some reason, the inference unit 53 refers to the CPT 311 or the CPT 312 to obtain the prior probability of the state information that cannot be acquired, and uses the probability propagation method to determine the prior probability. The probabilities P (x ₃ = True | x ₁ , x ₂ ) and P (x ₃ = False | x ₁ , x ₂ ) can be calculated using the probabilities.

In the above example, the inference unit 53 uses a probability model having a two-layer network configuration for simplification. However, the inference unit 53 may use a probability model having a network configuration of three or more layers including an intermediate layer. Further, the number of input nodes, the type of state information given to the input nodes, and the classification of the value of the state information are not limited to the above example.
Further, the inference unit 53 may employ various other known machine learning systems such as a perceptron type neural network.

The inference unit 53 corrects the setting parameters of the air conditioning unit 2 such as the _set temperature T _set and the air volume W according to the obtained recommendation level. For example, if the recommendation degree obtained for the operation of setting the air conditioner to the inside air circulation mode is higher than a predetermined threshold, the inference unit 53 sets the setting parameter related to the intake air setting mode to the inside air circulation mode. Set. Alternatively, the inference unit 53 may include all setting parameter values included in a specific setting item (for example, the setting item is a set temperature T _set , and the set parameter value may be 25 ° C. or 26 ° C. that the set temperature T _set can take. Etc.), the setting parameter is corrected so that the corresponding recommendation level becomes the highest value. The inference unit 53 temporarily stores the setting parameters in the RAM of the control unit 5 so that the setting parameters can be used by the respective units of the control unit 5 when the setting parameters are corrected as necessary by the above processing. To do.

The air conditioning control unit 54 reads each setting information and sensing information acquired from each sensor from the RAM, and controls the air conditioning unit 2 based on those values. At this time, if the setting parameter corrected by the inference unit 53 is stored, the air conditioning control unit 54 uses the corrected setting parameter.
Specifically, the air-conditioning control unit 54 determines the required outlet temperature (air-conditioning temperature T _ao ) of the conditioned air sent from each outlet based on the set temperature T _set and the measurement signals of each temperature sensor and solar radiation sensor. decide. Thereafter, the air conditioning control unit 54 determines the opening of the air mix door so that the temperature of the conditioned air becomes the air conditioning temperature _Tao . And the air-conditioning control part 54 transmits a control signal to the temperature control servomotor for moving an air mix door so that an air mix door may become the opening degree.

The air-conditioning control unit 54, the air-conditioning temperature T _ao, and the like based on the set temperature T _set and the evaporator outlet temperature, and controls the ON / OFF of the compressor constituting the refrigeration cycle. The air conditioning control unit 54, as a rule, operates the compressor and operates the refrigeration cycle when cooling the interior of the vehicle or operating the defroster.

Further, the air conditioning control unit 54 obtains the air volume and the air volume ratio of the conditioned air sent from each outlet based on the air conditioning temperature T _ao , the set temperature T _set and the like. And the air-conditioning control part 54 adjusts the rotation speed of the blower fan of the air-conditioning part 2 so that it may correspond to the determined air volume. Moreover, the air-conditioning control part 54 determines the opening degree of each blower outlet so as to correspond to the air volume ratio. Furthermore, the air conditioning control unit 54 determines the ratio of the air that the air conditioner 1 takes in from the inside air inlet and the air that takes in from the outside air inlet based on the air conditioning temperature T _ao , the set temperature T _set , the inside air temperature _{Tr, and the} like. Set.

Air-conditioning control unit 54, in order to determine the air-conditioning temperature T _ao, e.g., the set temperature T _{The set,} inside temperature T _r, the outside temperature T _am and the amount of solar radiation S and temperature control expression indicating the relationship between the air-conditioning temperature T _ao Is used. In order to determine the air volume W, the air-conditioning control unit 54 uses, for example, a set temperature T _set , an inside air temperature T _r , an outside air temperature _Tam, and an air volume control expression that expresses the relationship between the solar radiation amount S and the air volume W. Alternatively, the air conditioning control unit 54 can use various known control methods in order to determine the air conditioning temperature _Tao and the air volume W. Similarly, the air-conditioning control unit 54 can use various well-known control methods for the determination of the airflow ratio, the ON / OFF control of the compressor, and the determination of the intake ratio. Therefore, detailed description of those control methods is omitted.

The certainty calculation unit 55 recommends performing a specific setting operation for the state information as a reference value for determining whether or not the state information acquired by the information acquisition unit 3 needs to be accumulated as learning data. A certainty factor r representing the certainty of the degree is calculated. Here, it is considered that the machine learning system for obtaining the recommendation degree needs to be learned in the following cases.
(1) The recommendation degree of the setting operation corresponding to a specific situation is low. For example, the setting of the air conditioner in a specific situation may not be constant. In such a case, when a machine learning system is learned using state information acquired in that particular situation, the machine learning system cannot properly determine which setting is optimal for that particular situation, A low recommendation level is output for any setting operation. Therefore, there is a high possibility that an appropriate setting operation is not recommended for the user. Therefore, when the recommendation degree of the setting operation corresponding to a specific situation is low, it is considered necessary to learn the machine learning system.
(2) The number of learning data corresponding to a specific situation is small. In such a case, it may not be possible to specify what setting the user prefers for the specific situation based on the learning data. Therefore, when the number of learning data corresponding to a specific situation is small, it is considered necessary to learn a machine learning system.

ここで、a_s、a_nは、それぞれ、所定の一つ以上の状態情報に対して、空調装置１の特定の設定項目について取り得る値s、nについて、学習データとして記憶部５１に蓄積されている数である（例えば、設定項目が吸気設定モードである場合、a_sは空調装置の外気導入モードに設定された回数であり、a_nは内気循環モードに設定された回数である）。またｆは、推論部５３で求められる推薦度に対応する変数である。またΒ(x,y)はベータ関数であり、Γ(x)はガンマ関数である。さらに、cmax、cminは、それぞれ、確信度rの算出対象とする推薦度fの上限及び下限を表す。例えば、特定の設定項目の値sに設定する操作が推薦される場合の確からしさを求めたい場合、cmaxを推薦度の最高値（例えば、1）に設定し、cminを推薦度の最高値と最低値の平均値（例えば、0.5）あるいはその値sに設定する操作が推薦される推薦度の下限値（例えば、0.6あるいは0.9など）に設定することができる。また、特定の設定項目の値sに設定する操作に対する所定の推薦度f₀の確からしさを求めたい場合、cmax、cminを、それぞれ、その推薦度f₀に所定のマージン値m（例えば、0.05）を加えた値及びその推薦度f₀からマージン値mを差し引いた値に設定することができる。 When the acquired state information corresponds to the above (1) or (2) regarding a specific setting item, the set of the state information is accumulated as learning data, and corresponds to both (1) and (2) If not, it is preferable to reject the set of state information. Therefore, the certainty factor calculation unit 55 determines the number of learning data related to a specific setting item among the learning data stored in the storage unit 51 as a criterion as to whether or not the above (1) or (2) is satisfied. Then, the certainty factor r is calculated based on the following equation.

Here, a _s, a _n, respectively, for a given one or more state information, a specific set possible for the item value s of the air conditioner 1, the n, stored in the storage unit 51 as the learning data and the number is (e.g., if the setting item is a suction setting mode, is a _s a number that is set to the outside air introduction mode of the air conditioner, the number of times a _n is set to the inside air recirculation mode). F is a variable corresponding to the recommendation degree obtained by the inference unit 53. Β (x, y) is a beta function, and Γ (x) is a gamma function. Further, cmax and cmin represent the upper limit and the lower limit of the recommendation degree f to be calculated for the certainty factor r, respectively. For example, if you want to determine the probability when an operation to set the value s of a specific setting item is recommended, set cmax to the highest recommendation level (for example, 1) and cmin to the highest recommendation level. It can be set to an average value (for example, 0.5) of the minimum value or a lower limit value (for example, 0.6 or 0.9) of the recommendation level at which the operation to set the value s is recommended. Also, if you want to find a certain likelihood of recommendation level f ₀ to the operation of setting the value s of a specific setting item, cmax, the cmin, respectively, the recommendation degree f ₀ to a predetermined margin value m (for example, 0.05 ) it can be set to a value obtained by subtracting a margin value m from the values and the recommended degree f ₀ plus.

As is clear from the equation (1), the certainty factor r is obtained as an integral value of a beta distribution having the learning data number and recommendation level corresponding to the setting operation recommended in a specific situation as inputs. Therefore, confidence r, the higher the recommendation level f, or, a _s, a higher value sum is large in a _n. Therefore, the certainty factor r is a relatively low value when the set of acquired state information corresponds to the above (1) or (2). On the other hand, when the set of acquired state information does not correspond to either (1) or (2), the certainty factor r is a relatively high value.

When there are three or more values that can be taken for a specific setting item of the air conditioner 1 (for example, when the setting temperature of the air conditioner is a setting item), the certainty factor calculation unit 55 is the function in (1) above. A Dirichlet distribution may be used as ρ (f). ρ (f) corresponds to the distribution of the recommendation degree f when the learning parameter is expressed as a distribution. In the two-layer Bayesian network, since the learning parameter corresponds to the CPT on a one-to-one basis, it is preferable to set a beta distribution (Dirichlet distribution) for ρ (f). Similarly, in a general machine learning system, it is preferable to set a distribution suitable for the learning parameter and set the distribution of the recommendation degree f to ρ (f).
Not limited to this, the certainty factor calculation unit 55 calculates another certainty factor r, so that the certainty factor r increases as the number of learning data related to the predetermined setting or the recommendation factor f increases. May be used.
The certainty factor calculation unit 55 calculates the certainty factor r every time the state information is acquired by the information acquisition unit 3. The certainty factor calculation unit 55 stores the calculated certainty factor r in the storage unit 51 in association with the acquired state information and setting items.

変化量Eが小さい場合、その特定の設定操作に対して、機械学習システムは十分に学習されていると考えられる。そこで、データ蓄積判定部５６は、変化量Eが、所定の閾値Th1よりも小さい場合、最新の状態情報を棄却する。なお、所定の閾値Th1は、これ以上機械学習システムを学習しても、得られる推薦度fの値はほとんど変化しない場合に対応する変化量Eの値（例えば、0.1）として設定される。
一方、変化量Eが大きい場合、その状態情報の値の組に対して、機械学習システムは十分に学習されていないと考えられる。そこで、データ蓄積判定部５６は、変化量Eが閾値Th1以上の場合、その状態情報に推薦度を算出した設定操作を関連付けて、記憶部５１に学習データとして蓄積する。 Whether the data accumulation determining unit 56 accumulates one or more latest state information acquired by the information acquiring unit 3 as learning data in the storage unit 51 based on the certainty factor r calculated by the certainty factor calculating unit 55. Or decide whether to reject. Therefore, each time the information acquisition unit 3 acquires the state information, the data accumulation determination unit 56 updates the state information and the latest certainty factor r ^t obtained for the specific setting operation and the same state information and The certainty factor r ^t−1 obtained once before the specific setting operation is read from the storage unit 51. Then, the data accumulation determination unit 56 calculates a certainty degree change amount E as a criterion for determining whether or not the acquired state information is accumulated as learning data. Note that the change amount E is calculated based on the following equation.

When the change amount E is small, it is considered that the machine learning system is sufficiently learned for the specific setting operation. Therefore, the data accumulation determination unit 56 rejects the latest state information when the change amount E is smaller than the predetermined threshold Th1. Note that the predetermined threshold Th1 is set as a value (for example, 0.1) of the amount of change E corresponding to the case where the value of the degree of recommendation f obtained hardly changes even if the machine learning system is further learned.
On the other hand, when the change amount E is large, it is considered that the machine learning system is not sufficiently learned for the set of state information values. Therefore, when the change amount E is equal to or greater than the threshold value Th1, the data accumulation determining unit 56 associates the setting operation for calculating the recommendation degree with the state information and accumulates it as learning data in the storage unit 51.

Alternatively, the data accumulation determination unit 56 may determine whether to accumulate one or more latest state information based on the value of the certainty factor r itself. If the latest confidence r ^t associated with a particular setting operation to the latest state information is higher than a predetermined threshold value Th2, the data storage determining unit 56 rejects the set of values of the condition information. Note that the threshold value Th2 is set to a value of the certainty factor r corresponding to the fact that the recommended level of the predetermined setting is appropriate for the setting operation for calculating the set of state information values and the recommended level f. On the other hand, when the latest certainty factor r is equal to or smaller than the threshold value Th2, the data accumulation determining unit 56 associates the setting operation for calculating the recommendation degree with the set of state information values, and accumulates them as learning data in the storage unit 51. Whether or not the recommendation level is appropriate is determined by a setting automatically corrected by the occupant through the operation panel 4 within a predetermined period (for example, 1 minute) after changing the setting of the air conditioner 1 according to the obtained recommendation level. Can be determined based on the number of times of further changing to a different setting. For example, when the ratio of the number of re-corrections by the occupant exceeds 50% of the number of automatic corrections, it is determined that the recommendation level is inappropriate and the reliability value and the recommendation level are determined to be inappropriate. The threshold Th2 can be determined by investigating the relationship between the times beforehand through experiments.
Further, the data accumulation determination unit 56 may check both of the above-described determination criteria, and if either one of the determination criteria is satisfied, the state information may be accumulated in the storage unit 51 as learning data.

  The learning unit 57 learns the machine learning system of the inference unit 53 using the learning data set accumulated in the storage unit 51. In this embodiment, a Bayesian network is used as the machine learning system. Therefore, the learning unit 57 updates the CPT corresponding to each node of the Bayesian network that outputs the recommendation degree for each setting included in the specific setting item, among the learning data included in the learning data set. Extract learning data related to setting items. Then, the learning unit 57 performs cross tabulation representing the frequency of appearance of each setting value of the specific setting item for each category of the value of the state information serving as the input parameter of the Bayesian network to be learned from the extracted learning data. A table (hereinafter referred to as CTT) is created. Then, the learning unit 57 becomes an input parameter by dividing the appearance frequency of each setting value of the specific setting item included in the value classification of each state information by the sum of the appearance frequencies of all the setting values. For each status information value category, the probability for each setting value of the specific setting item is obtained, and the CPT is updated.

  Note that the learning unit 57 may obtain a CPT for each of a plurality of probability models having a predetermined graph structure in the same manner as described above. Then, the learning unit 57 may calculate an information amount reference for the plurality of probability models, and select a probability model in which the value of the information amount reference is optimal. In this case, the learning unit 57 uses AIC (Akaike information criterion), Bayesian information criterion (BIC), Takeuchi information criterion (TIC), minimum description length (MDL) criterion, etc. as information criterion. it can.

Furthermore, the learning unit 57 may learn the machine learning system using other various known methods. For example, the learning unit 57 may search for the graph structure of the Bayesian network using a K2 algorithm or a genetic algorithm. For example, in the case of using a genetic algorithm, a plurality of genes having each element as to whether or not each node is connected is prepared. Then, the learning unit 57 calculates the fitness of each gene using the above information criterion. Thereafter, the learning unit 57 selects a gene having a fitness level equal to or higher than a predetermined level, and performs the operation such as crossover or mutation to create the next generation gene. The learning unit 57 repeats such an operation a plurality of times and selects the gene having the highest fitness. The learning unit 57 uses the graph structure described by the selected gene for the Bayesian network.
When a perceptron type neural network is employed as the machine learning system, the learning unit 57 can learn the machine learning system using a back propagation algorithm.
The learning unit 57 stores, in the storage unit 51, data related to the learned machine learning system (graph structure of the probability model, CPT corresponding to each node, and the like).

  Hereinafter, the learning process of the air conditioner 1 according to one embodiment of the present invention will be described with reference to the flowchart shown in FIG. This learning process is executed by the control unit 5.

  First, one or more pieces of state information are acquired by the information acquisition unit 3 and transferred to the control unit 5 (step S101). When receiving the state information, the inference unit 53 of the control unit 5 inputs the state information to an available machine learning system, and calculates a recommendation degree for each setting of setting items corresponding to the machine learning system (step) S102). Next, the certainty factor calculation unit 55 of the control unit 5 calculates the certainty factor from the number of learning data related to the setting corresponding to the recommendation degree calculated by the inference unit 53 (step S103).

  Next, the data accumulation determination unit 56 of the control unit 5 determines whether or not a criterion for accumulating the acquired state information as learning data is satisfied based on the calculated certainty factor (step S104). . For example, as described above, the data accumulation determination unit 56 determines that the determination criterion is satisfied when the certainty degree change amount E is equal to or greater than a predetermined threshold, and the certainty degree change amount E is less than the predetermined threshold value. Determine that the criteria are not met.

In step S104, when the data accumulation determination unit 56 determines that the above determination criterion is not satisfied, the control unit 5 ends the learning process. On the other hand, in step S104, when the data accumulation determination unit 56 determines that the above determination criterion is satisfied, the acquired state information is stored in the storage unit 51 as learning data in association with the setting for calculating the recommendation degree ( Step S105). Then, the learning unit 57 of the control unit 5 learns the machine learning system using the newly accumulated learning data and the learning data already stored in the storage unit 51 (step S106). Thereafter, the control unit 5 ends the learning process.
When there are a plurality of machine learning systems that can be used, the control unit 5 executes the processes of steps S102 to S106 described above for each machine learning system. When the criteria for accumulating as learning data are satisfied for a plurality of machine learning systems, the state information obtained in step S101 is accumulated for each of the machine learning systems. On the other hand, for any machine learning system, if the criterion for accumulation as learning data is not satisfied, the state information obtained in step S101 is rejected after being used for air conditioning processing. The control unit 5 executes this learning process every time new state information is acquired.

  Hereinafter, the air conditioning process of the air conditioner 1 according to one embodiment of the present invention will be described with reference to the flowchart shown in FIG. The air conditioning process is executed by the control unit 5 and operates independently of the learning process.

  As shown in FIG. 5, first, when the engine switch is turned on, the control unit 5 operates the air conditioner 1. And each status information is acquired from the information acquisition part 3 through the communication part 52 (step S201).

Next, the inference unit 53 of the control unit 5 calculates the recommendation degree for each setting of the setting item corresponding to the machine learning system by inputting the value of the acquired state information to an available machine learning system. (Step S202). Then, the inference unit 53 compares the obtained recommendation level with a predetermined threshold value (step S203). When the recommendation level is equal to or higher than the threshold (for example, 0.9), the inference unit 53 corrects the setting parameter of the corresponding air conditioner 1 so as to achieve the setting (step S204). And the air-conditioning control part 54 of the control part 5 controls the air-conditioning part 2 based on the setting parameter corrected as needed (step S205). Specifically, the air-conditioning control unit 54 adjusts the air mix door of the air-conditioning unit 2, the number of rotations of the blower fan, the opening degree of each outlet, and the like so as to obtain a desired air-conditioning temperature and air volume. .
On the other hand, if the recommendation level is less than the threshold value in step S203, the inference unit 53 does not correct the setting parameters of the air conditioner 1. In this case, the air conditioning control unit 54 performs the air conditioning unit 2 according to the normal air conditioning control process. Is controlled (step S206). When there are a plurality of machine learning systems that can be used, the control unit 5 executes the processes of steps S202 to S206 described above for each machine learning system.

  Thereafter, the control unit 5 repeats the processes of steps S201 to S206 at regular time intervals until the operation is stopped.

  As described above, the air conditioner to which the learning data management device according to the present invention is applied is recommended for a predetermined setting obtained by the machine learning system each time state information is newly acquired, and the setting. On the basis of the beta distribution or the Dirichlet distribution with the number of learning data related to the input, a certainty factor representing the certainty of the recommendation level is calculated. And the air conditioner which concerns determines whether it accumulate | stores the acquired status information as learning data for learning a machine learning system based on the certainty factor. Therefore, the air conditioner can store only state information corresponding to a case where the machine learning system is not sufficiently learned as learning data. Therefore, the air conditioner can prevent the learning data obtained in the steady state from having a dominant influence on the learning of the machine learning system. It is possible to change the setting to recommend quickly.

  In addition, this invention is not limited to said embodiment. For example, the air conditioner may erase a part or all of the learning data corresponding to the set of state information and setting information to be ignored from the already accumulated learning data set.

Furthermore, in said embodiment, the inference part 53 made the parameter corrected based on a machine learning system the setting parameter which a passenger | crew can set directly through the operation part 4, such as setting temperature and an air volume. However, the inference unit 53 is configured so that each part of the air-conditioning unit 2 includes the air-conditioning temperature _Tao calculated using the temperature control control formula or the blower fan speed calculated using the airflow control formula, the opening degree of the air mix door, and the like. The control parameters directly related to the operation of may be modified.

  In addition, the air conditioner 1 may separately generate and use a machine learning system for calculating the probability of performing a predetermined setting operation for each user registered in advance. Similarly, the air conditioner 1 may store a learning data set separately for each user. In this case, the air conditioner 1 separately has a mechanism for identifying the occupant, and uses only the machine learning system and the learning data set associated with the registered user identified as the occupant by the mechanism. As a mechanism for identifying an occupant, for example, a camera that acquires an occupant's face image and an operation on a control unit that compares and collates the occupant's face image with a registered user's face image by pattern matching or the like. Software module.

  The present invention is not limited to an air conditioner, and automatically corrects any setting based on a machine learning system when a specific situation occurs in another device or a device to be controlled. The machine learning system can be applied to a controller that sequentially learns. For example, the present invention can be applied to car audio. In this case, it is assumed that every time an occupant approaches a point where an AM broadcast that informs traffic jam information can be received, an operation for selecting the AM broadcast is performed. At this time, the combination of the position information of the vehicle corresponding to the point and the setting operation for selecting the AM broadcast is accumulated as learning data. When the learning data is sufficiently accumulated, the machine learning system is learned based on the position information of the vehicle so that the recommendation level for the setting operation for selecting the AM broadcast is increased. When approaching a point where reception is possible, the AM broadcast is automatically selected. However, after the machine learning system is sufficiently learned, even if the position information of the vehicle corresponding to the point is acquired according to the present invention, the position information is not accumulated as learning data. For this reason, the number of learning data of the combination of the position information of the vehicle corresponding to the point and the setting operation for selecting the AM broadcast that has already been accumulated does not increase unnecessarily.

On the other hand, when an FM broadcast with a passenger is selected every time the vehicle passes the same point as described above, the car audio to which the present invention is applied selects the position information of the vehicle corresponding to the point and the FM broadcast. A combination of setting operations to be performed is accumulated as learning data. Therefore, the machine learning system is quickly relearned based on the position information of the vehicle so that the recommendation level for the setting operation for selecting the FM broadcast is increased. When the car audio approaches the point, the FM broadcast is automatically selected.
Further, the setting operation performed by the automatic control may be turning on / off the car audio, operating the CD, adjusting the volume, and the like when the car audio is a control target device.

  Further, the present invention provides vehicle body control, such as power windows, key locks, headlights, hazard lamps, mirrors, fillers, sunroofs, wipers, automatic inter-vehicle control systems (ACC), electronically controlled suspensions (AVS), shifts, etc. Can be used to control. Further, when the device to be controlled is the vehicle itself, the setting operation performed by the automatic control is a driving operation of the vehicle, for example, depressing the brake pedal, accelerating / decelerating, moving the wiper, opening / closing the power window. The operation may be as follows.

  Furthermore, the present invention is not limited to in-vehicle devices, and can be applied to other devices whose settings are modified according to user preferences using a machine learning system.

Furthermore, the learning data management device according to the present invention is provided independently of a device to be automatically controlled or a control device that controls such a device, and is necessary between the learning data management device and the control device. Such information may be transmitted / received to / from each other. In this case, the learning data management apparatus may include only the storage unit, the certainty factor calculation unit, and the data accumulation determination unit among the units in the above-described embodiments. Alternatively, the learning data management apparatus may include only a storage unit, a certainty factor calculation unit, a data accumulation determination unit, and a learning unit.
As described above, those skilled in the art can make various modifications within the scope of the present invention.

It is a schematic block diagram of the air conditioner to which the learning data management apparatus of this invention is applied. It is a functional block diagram of the control part of the air conditioner shown in FIG. It is a figure which shows an example of the probability model which is a machine learning system used in this embodiment. It is a flowchart of the learning process of the air conditioner which concerns on embodiment of this invention. It is a flowchart of the air-conditioning process of the air conditioner which concerns on embodiment of this invention.

Explanation of symbols

1 Air conditioner (learning data management device)
2 Air Conditioning Unit 3 Information Acquisition Unit 4 Operation Panel 5 Control Unit 51 Storage Unit 52 Communication Unit 53 Inference Unit 54 Air Conditioning Control Unit 55 Certainty Calculation Unit 56 Data Accumulation Determination Unit 57 Learning Unit

Claims (8)

In order to automatically control the device (2), the device (2) is set to a predetermined setting by inputting at least one state information on the device (2) acquired by the state information acquisition unit (3). A learning data management device for accumulating learning data used for learning of a machine learning system for obtaining a recommendation degree,
A storage unit (51) for storing a set of the state information and the predetermined setting as learning data;
When new state information is acquired by the state information acquisition unit (3), among the learning data stored in the storage unit (51), the new information is calculated based on the number of learning data related to the predetermined setting. A certainty factor calculation unit (55) for calculating a certainty factor representing the certainty of the recommended degree of the predetermined setting obtained by inputting correct state information to the machine learning system;
When the certainty degree satisfies a predetermined criterion indicating that it is necessary to learn the machine learning system, the set of the new state information and the predetermined setting is accumulated as learning data in the storage unit (51), A data accumulation determination unit (56) for rejecting the new state information when the certainty factor does not satisfy the predetermined criterion;
A learning data management device characterized by comprising:   The certainty factor calculation unit (55) obtains the certainty factor using a function that increases the certainty factor as the number of learning data corresponding to the state information and the predetermined setting or the recommendation factor increases. The learning data management apparatus according to claim 1.   3. The function according to claim 2, wherein the function includes a beta distribution or a Dirichlet distribution in which the number of learning data related to the predetermined setting and the recommendation level are input, and the certainty factor for the predetermined value of the recommendation level is output. The learning data management device described.   The data accumulation determination unit (56) is configured to calculate the absolute value of the difference between the latest certainty factor obtained with respect to the predetermined setting recommendation degree and the previous certainty factor obtained with respect to the predetermined setting recommendation degree. 4 is determined to satisfy the predetermined criterion when it is larger than a first threshold corresponding to a change in the recommendation degree of the predetermined setting by learning the machine learning system. The learning data management apparatus according to any one of the above.   The data accumulation determining unit (56), when the latest certainty factor obtained with respect to the recommendation level of the predetermined setting is equal to or less than a second threshold corresponding to the recommendation level of the predetermined setting being appropriate, The learning data management apparatus according to claim 1, wherein the learning data management apparatus determines that the predetermined criterion is satisfied. A state information acquisition unit (3) for acquiring at least one state information regarding the device (2) to be controlled;
An inference unit (53) for calculating a recommendation degree for setting the device (2) to a predetermined setting by inputting the at least one state information to a machine learning system;
A control unit (54) for controlling the device (2) according to the recommendation degree;
A storage unit (51) for storing a set of the state information and the predetermined setting as learning data;
When new state information is acquired by the state information acquisition unit (3), among the learning data stored in the storage unit (51), the new information is calculated based on the number of learning data related to the predetermined setting. A certainty factor calculation unit (55) for calculating a certainty factor representing the certainty of the recommended degree of the predetermined setting obtained by inputting correct state information to the machine learning system;
When the certainty degree satisfies a predetermined criterion indicating that it is necessary to learn the machine learning system, the set of the new state information and the predetermined setting is accumulated as learning data in the storage unit (51), A data accumulation determination unit (56) for rejecting the new state information when the certainty factor does not satisfy the predetermined criterion;
A learning unit (57) for learning the machine learning system using learning data stored in the storage unit (51);
A control device comprising: A vehicle air conditioner,
An air conditioning unit (2) for supplying conditioned air into the vehicle;
A state information acquisition unit (3) for acquiring at least one state information about the vehicle;
An inference unit (53) for calculating a recommendation degree for setting the air conditioning unit (2) to a predetermined setting by inputting the at least one state information to a machine learning system;
An air conditioning control unit (54) for controlling the air conditioning unit (2) according to the recommendation degree;
A storage unit (51) for storing a set of the state information and the predetermined setting as learning data;
When new state information is acquired by the state information acquisition unit (3), among the learning data stored in the storage unit (51), the new information is calculated based on the number of learning data related to the predetermined setting. A certainty factor calculation unit (55) for calculating a certainty factor representing the certainty of the recommended degree of the predetermined setting obtained by inputting correct state information to the machine learning system;
When the certainty degree satisfies a predetermined criterion indicating that it is necessary to learn the machine learning system, the set of the new state information and the predetermined setting is accumulated as learning data in the storage unit (51), A data accumulation determination unit (56) for rejecting the new state information when the certainty factor does not satisfy the predetermined criterion;
A learning unit (56) for learning the machine learning system using learning data accumulated in the storage unit (51);
A vehicle air conditioner comprising: In order to automatically control the device (2), the device (2) is set to a predetermined setting by inputting at least one state information on the device (2) acquired by the state information acquisition unit (3). A learning data management method for storing learning data used for learning of a machine learning system for obtaining a recommendation degree in a storage unit (51),
When new state information is acquired by the state information acquisition unit (3), among the learning data already accumulated in the storage unit (51), the number of learning data related to the predetermined setting Calculating a certainty factor representing the certainty of the recommended degree of the predetermined setting obtained by inputting new state information to the machine learning system;
When the certainty factor satisfies a predetermined criterion indicating that it is necessary to learn the machine learning system, the set of the new state information and the predetermined setting is stored in the storage unit (51) as new learning data. And, when the certainty factor does not satisfy the predetermined criterion, rejecting the new state information;
The learning data management method characterized by including.

Images