JP3339810B2 - Multi-input multi-output control apparatus, method, and medium - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a multi-input multi-output control apparatus, method, and medium, and more particularly, to a multi-input multi-output control apparatus having a non-linear correspondence relation to a large number of types of input data. The present invention relates to a multi-input multi-output control device, method, and medium for a vehicle air conditioner or the like that outputs output data of any type.

[0002]

2. Description of the Related Art A neural network system in which a neural circuit is modeled is known as a system that can easily find the correspondence between a large number of linear and nonlinear inputs and outputs. The neural network is a model of a neural circuit. The input layer, the intermediate layer, and the output layer composed of a plurality of neurons are weighted and connected, and the connection relationship is learned based on known input / output data. .
As a result, an output corresponding to the input can be obtained even if the input / output relationship is nonlinear.

[0003] However, a neural network system having a single input / output correspondence has a simple structure, but requires a large number of connections and a large number of input / output data for learning the neural network. Also, the time required for learning is enormous.

In order to solve such a problem, input data having an arbitrary distribution is classified into categories, and when the same category cannot be learned by a single neural network, learning is performed by each of a plurality of neural networks. (Japanese Patent Laid-Open No. Hei 3-2959, Japanese Patent Laid-Open No. 5-13)
No. 5000).

[0005]

However, since a plurality of neural networks are used for learning with each of the plurality of neural networks, the size of the entire neural network system becomes large. Therefore, the calculation speed is reduced and the cost is increased.

[0006] Further, even when a device for category classification is required and a neural network is constructed for each category, there is no correlation between the categories and continuous data is required as a whole output. The continuity of the output data of the neural network may be lost.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a multi-input multi-output control device, method and medium capable of improving the operation efficiency in consideration of the above fact.

[0008]

In order to achieve the above object, according to the present invention, a plurality of types of input data having a non-linear correspondence with a plurality of types of input data are provided. A multi-input multi-output control device for outputting the output data, input means for inputting the input data, output means for outputting the output data, and intermediate means for transmitting data from the input means to the output means And comprising between the input means and the output means a neural network model having a multilayer structure in which a connection relationship between a large number of neural circuit elements is modeled, and a predetermined predetermined circuit included in the intermediate means. Highly linear data relation for each of multiple types of intermediate data as output of neural circuit elements
And classifying the input means with an input group consisting of a type of input data having a high dependency determined by a data relationship having a high non-linearity, and providing the dependency between the neural circuit element of the intermediate means corresponding to the classified input group. The respective connection relations of the neural circuit elements are determined so as not to be affected by low-type input data, and the corresponding means that defines the connection relations of the neural circuit elements between the intermediate means and the output means. Correspondence between the multiple types of input data and the multiple types of output data of the corresponding means using the relationship data in which the correspondence between the multiple types of input data and the multiple types of output data is known in advance. And a construction means for establishing a relationship.

According to a second aspect of the present invention, in the multiple-input multiple-output control apparatus according to the first aspect, the construction means is configured to determine whether the input data belonging to the classified input group and the multiple types of output data are included. Using the specific data for each input group whose correspondence is known in advance, the correspondence between the input data belonging to the classified input group and many types of output data is constructed, and based on the constructed correspondence for each input group. And adjusting the connection relationship of the neural circuit element between the intermediate means and the output means.

The third aspect of the present invention is the first or second aspect.
The multi-input / multi-output control device according to the above, further comprising a correction means for correcting the output data of the output means to a predetermined tendency. The invention as set forth in claim 4 is
From the input data, you can determine the season or vehicle type difference or occupant preference.
Estimated by fuzzy inference and the estimated season
Or, the correspondence between the difference in vehicle type or the occupant preference and the correction amount
Calculate the correction amount by referring to the mapped map
The output data of the output means is calculated based on the obtained correction amount.
It further comprises a fuzzy inference device for correcting
I do. According to a fifth aspect of the present invention, there is provided
Calculate the correction amount corresponding to the difference of the knot or vehicle type or the occupant's preference.
The output data of the output means is calculated based on the obtained correction amount.
That we also have a neural network to compensate
Features.

[0011] The invention described in claim 6 is claim 1 to claim 1.
5. In the multi-input multi-output control device according to any one of 5, the input data is environmental information around the vehicle and predetermined setting information, and the output data is device information of a vehicle air conditioner. There is a feature.

A seventh aspect of the present invention is a multi-input multi-output control method for outputting a large number of types of output data having a correspondence including a nonlinear correspondence with a large number of types of input data. In addition, a portion between the input and output of the input data is constituted by a neural network model having a multilayer structure in which a connection relationship between a large number of neural circuit elements is modeled, and a predetermined portion included between the input and the output is included. A plurality of types of intermediate data as outputs of the predetermined neural circuit element, and an input group consisting of a highly dependent input data type determined by a highly linear data relationship and a highly nonlinear data relationship. The input data is classified, and the connection of the neural circuit element is prevented so that the input data is not affected by the low-dependency type input data until the neural circuit element corresponding to the classified input group. Each of the relationships is determined, and the connection relationship of the neural circuit element from the output of the intermediate data to the output of the output data is determined, and the correspondence between the multiple types of input data and the multiple types of output data Is the input of the input data using the known relation data in advance .
Characterized by constructing the multiple types of correspondence between the input data and the number of types of output data up al output.

The invention according to claim 8 is the multi-input / multi-output control method according to claim 7, wherein the construction of the correspondence relation is such that the input data belonging to the classified input group and the multiple types of The correspondence between the output data and the input data belonging to the classified input group and the various types of output data are constructed using the specific data for each input group that is known in advance. Based on the correspondence
And adjusting the connections of the neural elements between the output of the data to the output of the output data.

The invention according to claim 9 is the invention according to claim 7 or 8.
3. The multi-input multi-output control method according to claim 1, wherein the output data is further corrected to a predetermined tendency. According to a tenth aspect of the present invention, a seasonal or vehicle type difference or occupant preference is estimated from the input data by fuzzy inference, and a corresponding relationship between the estimated season or vehicle type difference or occupant preference and a correction amount. There obtains a correction amount by referring to the mapped map, and corrects the output data by the correction amount determined said. Claim 1
The invention described in 1 obtains a correction amount corresponding to the difference or the passenger's preference seasons or vehicle type from the input data by a neural network, and corrects the output data by the correction amount determined said.

According to a twelfth aspect of the present invention, there is provided the multi-input multi-output control method according to any one of the seventh to eleventh aspects, wherein the input data includes environmental information around the vehicle and predetermined information. In addition to the setting information, the output data is device information of a vehicle air conditioner.

According to a thirteenth aspect of the present invention, there is provided a multi-input system for outputting, by a computer, a plurality of types of output data in a correspondence including a non-linear correspondence with respect to a plurality of types of input data inputted. A recording medium on which a multi-output control program is recorded, wherein a period from the input to the output of the input data is constituted by a multi-layered neural network model that models a connection relationship of a large number of neural circuit elements, and Is highly dependent on each of a plurality of types of intermediate data as an output of a predetermined predetermined neural circuit element included in a period from to the output, which is determined by a highly linear data relationship and a highly nonlinear data relationship. The input data is classified by an input group consisting of the type of input data, and the type having a low dependency between the neural network elements corresponding to the classified input group. While determining the connection relationship of the neural circuit elements so as not to be affected by the input data, and determining the connection relationship of the neural circuit element from the output of the intermediate data to the output of the output data, correspondence between the number of types the many types and the input data of the output data in advance using a known relationship data, entering-
The present invention is characterized in that a correspondence between the many types of input data and the many types of output data from input to output of force data is constructed.

According to a fourteenth aspect of the present invention, there is provided a recording medium on which the multi-input / multi-output control program according to the thirteenth aspect is recorded, wherein the correspondence is constructed by input data belonging to the classified input group. The correspondence relationship between the input data belonging to the classified input group and the many types of output data is constructed by using the specific data for each input group whose correspondence relationship with the many types of output data is known in advance. The connection relationship of the neural circuit element from the output of the intermediate data to the output of the output data is adjusted based on the correspondence relationship for each input group.

According to a fifteenth aspect of the present invention, there is provided a recording medium storing the multi-input / multi-output control program according to the thirteenth or fourteenth aspect, wherein the output data is further corrected to a predetermined tendency. The invention according to claim 16 estimates the season or vehicle type difference or occupant preference from the input data by fuzzy inference, and the correspondence between the estimated season or vehicle type difference or occupant preference and the correction amount. There obtains a correction amount by referring to the mapped map, and corrects the output data by the correction amount determined said. The invention according to claim 17 obtains a correction amount corresponding to a season or a difference in a vehicle type or a preference of an occupant from the input data by a neural network,
And corrects the output data by the correction amount determined said.

According to an eighteenth aspect of the present invention, there is provided a recording medium on which the multi-input / multi-output control program according to any one of the thirteenth to seventeenth aspects is recorded, wherein the input data is an environment around a vehicle. The output data is device information of a vehicle air conditioner, in addition to the information and predetermined setting information.

In the multi-input multi-output control device according to the first aspect, as shown in FIG.
The correspondence unit 10 outputs a large number of types of output data 20 that have a correspondence relationship including a non-linear correspondence relationship with 2. The correspondence unit 10 includes an input unit 14, an intermediate unit 16, and an output unit 18. Input data 12 is input means 1
4 and output data 20 is output from the output means 18. Between these input means 14 and output means 18, there is an intermediate means 16 for propagating data. Each of the input means 14, the intermediate means 16, and the output means 18 has a large number of neural circuit elements (so-called neurons) 22, and these neural circuit elements 22 are connected.

That is, from the input means 14 to the output means 18
Up to this point, a neural network model having a multilayer structure in which the connection relation of a large number of neural circuit elements 22 is modeled. The predetermined neural circuit element 2 included in the intermediate means 16
The input unit 14 is classified into an input group including input data types that have a high dependency on each of a plurality of types of intermediate data as the second output. For Figure 1, the predetermined neural element 22 a predetermined, the device 22 _1, 22 _2, 22 _3, as an element 22 _4, 22 _5, 22 _6, as dependent high input data, data X _1, X ₂ , X ₃ and data X ₄ ,
X _5, are divided into two input groups by X _6.

Intermediate means corresponding to the classified input group
Types of inputs with low dependence on up to 16 neural network elements
To avoid being affected by the data,
The connection relationship is determined. That is, data X₁, X_Two,
X_ThreeThe neural circuit element of the intermediate means 16 on the side
X_Four, X_Five, X₆Data to avoid being affected by
X _Four, X_Five, X₆In the neural circuit element of the intermediate means 16 on the side
Do not connect. Thereby, the input group of the input means and the intermediate means
16 data X₁, X_Two, X_ThreeSide area 24 and the input hand
The input group of the stage and the data X of the intermediate means 16_Four, X_Five, X₆~ side
And the region 26 of FIG. Then, it comes out with the intermediate means 16.
All the connection relations of the neural circuit element 22 with the force means 18
Determine. Thereby, the final neural circuit element of the intermediate means 16 is obtained.
An area 28 from the child 22 to the output means 18 is defined.

The structuring means 30 uses the relationship data in which the correspondence between the many types of input data 12 and the many types of output data 20 is known in advance, and Is established with the type of output data 20.

As described above, since the connection relation of the neural circuit elements is determined so as not to be affected by the input data of the type having low dependence, the connection relation of the neural circuit elements can be reduced, and the calculation efficiency can be improved. Can be done.

In the case where the input data 12 interacts with a plurality of data classified at the time of classification as described above, the input data 12 may be duplicated. For example, when each of the data X ₃ and X ₄ is required in the example of FIG. 1, each of the data X ₃ and X ₄ may be input as shown in FIG.

According to a second aspect, the construction means uses the specific data for each input group in which the correspondence between the input data belonging to the classified input group and a large number of types of output data is known in advance.
A correspondence relationship between input data belonging to the classified input group and many types of output data is constructed, and a connection relationship of the neural circuit elements between the intermediate means and the output means based on the constructed correspondence relation for each input group. To adjust.

That is, in the same manner as in the first aspect, the neural circuit element is controlled so as not to be affected by the input data of a low dependency type until the neural circuit element of the intermediate means 16 corresponding to the classified input group. 22 connection relationships are defined. Then, an area including the output means 18 connected to the final neural circuit element of the intermediate means 16 is determined . As a result, FIG.
As shown in the figure, the input group of the input means, the area on the side of the data X ₁ , X ₂ , X ₃ of the intermediate means 16 and the area 25 composed of the output means 18, the input group of the input means and the data X of the intermediate means 16
_4, X _5, defines a region 27 of the X ₆ side. These settings can be determined by a data relationship having high linearity and a data relationship having high nonlinearity. And the intermediate means 16
A region 28 from the last neural circuit element 22 to the output means 18 is determined.

[0028] The construction means uses input data belonging to the classified input group and uses input data belonging to the classified input group, using specific data for each input group in which the correspondence between the input data belonging to the classified input group and many types of output data is known in advance. And many types of output data. That is, in FIG.
7 are independently learned and constructed. The connection relation of the neural circuit elements between the intermediate means and the output means, which is the area 28, is adjusted based on the correspondence relation for each of the constructed input groups.

Thus, for example, when input data and output data have a linear or non-linear relationship, they can be separately learned. For this reason, it is possible to suppress insufficient learning when the linearity is weak, that is, when the nonlinearity is strong, and to suppress over-learning when the linearity is weak, thereby improving the calculation efficiency.

Here, to obtain the relationship between the input data and the output data, there is an environment control device that adjusts to an optimum environment state according to the current environment state. As the environment control device, an air conditioner such as an air conditioner, a cooling / heating device,
There is a control device for a living space in a vehicle such as a seat position or a suspension according to a vehicle speed. These devices do not change the main relationship, such as user preference, time adjustment such as morning and afternoon, and seasonal adjustment, but may require fine adjustment. Corresponding all of these fine adjustments to the corresponding means may cause the computational efficiency to deteriorate.

In view of the above, the present invention can further comprise a correcting means for correcting the output data of the output means to a predetermined tendency. Thereby, fine adjustment can be performed without deteriorating the calculation efficiency.

As described in claim 6 , the input data is used as environmental information around the vehicle and predetermined setting information, and the output data is used as device information such as the set temperature and air volume of the vehicle air conditioner. By doing so, the multi-input multi-output control device can be easily used for an air conditioner.

In order to output a large number of types of output data having a correspondence including a non-linear correspondence with a large number of types of input data inputted, a multi-input multiple-output control method according to claim 7 is used. By the control, the correspondence can be easily specified regardless of the linear correspondence or the non-linear correspondence, and output data optimal for the inputted input data can be obtained. That is, between the input and the output of the input data, a neural network model having a multilayer structure in which the connection relation of a large number of neural circuit elements is modeled, and the predetermined between the input and the output is included. A plurality of types of intermediate data as an output of a predetermined neural circuit element, each of which is an input group consisting of a type of input data having a high dependence determined by a highly linear data relationship and a highly nonlinear data relationship; Classifying data, and determining the connection relationship of the neural circuit elements so that the neural network elements corresponding to the classified input group are not affected by the low-dependence type input data until the neural circuit elements; The connection relationship of the neural circuit element from the output of the intermediate data to the output of the output data is determined, and the multiple types of input data and the multiple types of output data are determined. Data and correspondence with the previously known relationship data, or the output from the input of the input data
Constructing the correspondence between the many types of input data and a number of types of output data in.

In the construction of the correspondence, the correspondence between the input data belonging to the classified input group and the plurality of types of output data is specified for each input group in which the correspondence is known in advance. A correspondence between input data belonging to the classified input group and a large number of types of output data is constructed using data, and the output data of the intermediate data is output from the output of the intermediate data based on the constructed correspondence for each input group. The connection relationship of the neural circuit element up to the output can be adjusted.

[0035] The output data, as described in claim 9 can be further corrected to a predetermined trend.

As described in claim 12 , the input data is used as environment information around the vehicle and predetermined setting information, and the output data is used as device information of a vehicle air conditioner. The present invention can be applied to a vehicle air conditioner.

According to the present invention, there is provided a multi-input multi-output control method.
The present invention can be realized by a multi-input / multi-output control program of a recording medium in which the multi-input / multi-output control program described in (1) is recorded. That is, on a recording medium on which a multi-input multi-output control program for outputting a large number of types of output data in a correspondence including a nonlinear correspondence to a large number of types of input data inputted by a computer is provided. In addition, between the input and output of the input data, a multi-layered neural network model that models the connection relationship of a large number of neural circuit elements is configured, and a predetermined value included between the input and the output is included. A plurality of types of intermediate data as outputs of the predetermined neural circuit element, and an input group consisting of a highly dependent input data type determined by a highly linear data relationship and a highly nonlinear data relationship. The input data is classified, and the input data is affected by the low-dependence type input data until the neural circuit element corresponding to the classified input group. And the connection relationship of the neural circuit elements between the output of the intermediate data and the output of the output data is determined, and the plurality of types of input data and The input data is output from the input of the input data using the relation data in which the correspondence relation with the many types of output data is known in advance.
The correspondence between the multiple types of input data up to the force and the multiple types of output data is constructed.

In the construction of the correspondence, the correspondence between the input data belonging to the classified input group and the plurality of types of output data is known for each input group in advance. The specific data is used to construct a correspondence between the input data belonging to the classified input group and a large number of types of output data, and the output data of the intermediate data is output based on the constructed correspondence for each input group. The connection relation of the neural circuit element up to the output of (i) can be adjusted.

Further, as described in claim 15, the output data, by further correcting the predetermined trend, it is possible to easily fine-tune.

Further, as described in claim 18 , the input data is used as environmental information around the vehicle and predetermined setting information, and the output data is used as device information of a vehicle air conditioner. The present invention can be applied to a vehicle air conditioner.

[0041]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings. In the first embodiment,
The present invention is applied to a vehicle air conditioner mounted on a vehicle that provides an environment with optimal air conditioning according to the environment inside and outside the vehicle. The vehicle air conditioner according to the present embodiment uses a neural network, which is a non-linear prediction technology in which a neural network of a higher animal is modeled by engineering.

FIG. 4 schematically shows a vehicle air conditioner 40 according to the present embodiment. This vehicle air conditioner 40
Are a computer 10 for inputting data and the like, a computer main body 42 for predicting and calculating an environment provision by optimal air conditioning using a neural network based on a nonlinear prediction method according to a program stored in advance, and a calculation result of the computer main body 42 Is displayed on the monitor 56.

The computer main body 42 includes a microcomputer, and includes a CPU 44, a RAM 4
6, ROM 48, memory 5 for storing data for constructing a neural network (details will be described later)
0, an input / output device (hereinafter referred to as I / O) 52 for exchanging data and the like between the main body and another device, and a bus 54 connected to these so that data and commands can be input and output Have been. The ROM 48 stores a processing program described later.

An input device 58 for an occupant in the vehicle to set and input an environmental state is connected to the I / O 52. The environmental conditions that can be set include a set temperature of each of a driver's seat and a passenger's seat, and a blowing mode such as a step and a face. Further, a sensor 60 for detecting an environment inside and outside the vehicle is connected to the I / O 52. The environment that can be detected by the sensor 60 includes the temperature inside and outside the vehicle, for example, the indoor temperature and the outdoor temperature, and the amount of solar radiation to each of the driver's seat and the passenger's seat.

A driving device 62 for providing an environment by air conditioning is connected to the I / O 52. Examples of the driving device 62 include a cooling / heating device for adjusting the temperature of the air blown to the driver's seat and the passenger seat, and an air volume adjusting device for adjusting the air volume to the driver's seat and the passenger seat.

Further, a read / write device (FD device) 64 for reading and writing information such as data and programs on a magnetic storage medium (FD) 66 such as a floppy disk can be connected to the I / O 26. Therefore, a processing program to be described later is not limited to being stored in the ROM 48, but may be stored in the FD 66, or a processing program stored in the FD 66 may be read out by the FD device 64 and stored in the memory 50. . Further, a storage medium such as a hard disk device may be provided inside the computer main body 42 or as an external device, and may be stored in the hard disk device. These stored processing programs are:
The program may be executed on the FD 66 stored in advance, a processing program described later may be stored, and the program may be executed on the FD 66, or may be executed by being stored in the memory 50 or the like. The recording medium is a CD-ROM
There are optical discs such as MD and magneto-optical discs such as MD and MO. When these are used, a CD-ROM device, MD device, MO device, etc. may be used instead of the FD device 64 or further.

FIG. 5 is a block diagram showing a schematic functional configuration of the vehicle air conditioner 40 of the present embodiment. The vehicle air conditioner 40 according to the present embodiment outputs output data for setting the inside of the vehicle living space to an optimal environmental condition based on the temperatures of the driver's seat and the passenger's seat.

The vehicle air conditioner 40 has a non-linear operation unit 70 (detailed later) constituted by a neural network as a main element, and the non-linear operation unit 70 determines the temperature of the driver's seat and the passenger's seat and the like in the vehicle living space. It functions as a calculation unit that calculates the correspondence with environmental conditions. Although the details will be described later, the non-linear operation unit 70 is for obtaining a model in which the temperature and the like of the driver's seat and the passenger's seat are associated with the environmental conditions in the vehicle living space based on the input data. In addition,
The model referred to here is a conversion system that can perform conversion and inverse conversion such that the temperature of the driver's seat and the front passenger's seat correspond to the environmental conditions in the vehicle living space on a one-to-one basis. When a network is represented by a mathematical expression, the expression includes a mathematical expression and its coefficient.

The data input by the input device 58 and the data detected by the sensor 60 are input to the nonlinear operation unit 70. In the present embodiment, as input data, driver's seat set temperature x ₁ , passenger seat set temperature x ₂ , indoor temperature x ₃ , outdoor temperature x ₄ , driver's seat solar radiation x ₅ , passenger seat solar radiation x ₆ , A driver's seat mode x ₇ and a passenger seat mode x ₈ are employed. These data are used to digitize and input the current environmental state (setting).

As a result of the calculation, data representing the optimum environmental condition for the current environmental state (setting) based on the input data is output from the non-linear calculation unit 70. In addition,
When the data resulting from the operation of the non-linear operation unit 70 is a value representing the performance of the product as it is, the value can be displayed or stored on the monitor 56 as it is,
The monitor 56, the storage device, and the like may correspond to the data output device. In this embodiment, as the data output, the driver's seat air temperature y _1, a passenger seat air temperature y _2, the driver's seat air volume y _3, employs a passenger seat air volume y _4. These data are quantified into device values of the driving device 62 in order to optimize environmental conditions.

In this embodiment, the nonlinear operation unit 7
Numeral 0 is configured using the hardware resources shown in FIG. 4 and software resources to be described later, has a conversion function formed by a conceptual neural network as described later, and has a learning function to learn the same. I have. That is, the non-linear operation unit 70 includes a computer main body 42 (CPU 44, RA
M46, ROM 48, memory 50, I / O 52, and bus 54) and hardware resources described below. In addition, the correspondence between the temperature of the driver's seat and the passenger seat and the environmental conditions in the vehicle living space is learned in advance by another neural network, and the learned conversion coefficient of the other neural network is input. A neural network may be constructed using the coefficients.

The above-mentioned non-linear operation unit 70 is provided as an input layer for controlling the temperature and setting of the driver's seat and the passenger's seat, etc., in order to enable the input of environmental values such as the temperature and setting of the driver's seat and the passenger's seat. A neural network having neurons corresponding to the number of environmental conditions and having neurons corresponding to environmental conditions in the vehicle living space as an output layer via an intermediate layer, and forming a neural network in which each neuron is connected by a synapse. ing. When the values of the current environmental conditions, such as the temperature and settings of the driver's seat and the passenger's seat, are input after learning, which will be described later, the nonlinear arithmetic unit 70 outputs the corresponding environmental conditions in the vehicle living space. Is done. At the time of learning, the environmental conditions in the known optimal vehicle living space corresponding to the current environmental conditions such as the temperature and settings of the driver's seat and the passenger seat are input as a teacher,
Depending on the magnitude of the error difference between the environmental conditions in the vehicle's living space and the known environmental conditions with the optimal output, each value of the current environmental conditions, such as the temperature and settings of the driver's seat and the passenger's seat, and the corresponding optimal It is set so as to correspond to the environmental conditions in the vehicle living space.

As shown in FIG.
An example of a neural network that has been
A unit corresponding to the
It is configured by connecting the input to the output. Real truth
In the embodiment, a predetermined number of units I₁, I_Two, ...,
I_jAn input layer consisting of (j> 1) connected to this input layer
Unit M1₁, M1_Two, ..., M1_k(K> 1)
1st intermediate layer consisting of, connected to the unit M1 of the 1st intermediate layer
Unit M2₁, M2_Two, ..., M2 _p(P>
A second intermediate layer comprising 1), and a unit of the second intermediate layer
Unit O connected to M2₁, O_Two, ..., O
_q(Q> 1). Ma
In addition, each unit of the first and second intermediate layers and each unit of the output layer
To allow the unit to apply a predetermined value to the output value
Connected to the bias units B1, B2, B3
You.

In this embodiment, each unit in the intermediate layer is constituted by a neural circuit element having a sigmoid characteristic whose input / output relationship is represented by a sigmoid function, and the unit in the input layer and the unit in the output layer are input / output. The relationship is composed of linear neural circuit elements. By configuring so as to have this sigmoid characteristic, the output value becomes a real value (positive number).

The unit I _j of the input layer of the nonlinear operation unit 70
Is composed of linear neurons, and the output of each unit _Ij can be expressed by the following equation (1). That is, when an input signal is u _j (corresponding to input data x _{1 to} x ₆ ) for a certain unit I _j , the output vj _j can be represented by a linear function gj.

Vj _j = gj (u _j ) (1)

The unit M1 of the first intermediate layer M1₁~
M1_kMeans that the input-output relationship is represented by a sigmoid function
Composed of neural network elements having sigmoidal characteristics
Then, each output of the unit of the first intermediate layer M1 is expressed by the following (2)
It can be represented by an equation. That is, the k-th unit
Weights (units) corresponding to the strength of each synaptic connection.
, The unit I of the j-th input layer
_jThe coupling coefficient for _jkAnd the input signal is vj_jIn
Virtual equivalent of the average value of the neuronal membrane potential.
Part state variable rk_kAnd the output vk_kIs new
Can be expressed by a nonlinear function f
You.

[0058]

(Equation 1)

Where b1: threshold value supplied from the bias unit B1 f (): sigmoid function

The unit of the second intermediate layer M2 is the first
As with the unit in the middle layer M1, the input / output relationship is
Nerve with sigmoidal properties represented by
Each of the units of the second intermediate layer M2.
The output can be expressed by the following equation (3). That is,
For the p-th unit, the strength of each synaptic connection
The corresponding weight (coupling coefficient of the unit),
The coupling coefficient for the unit M1 of the first intermediate layer is w_kpWhen
And the input signal is vk_kAnd the flatness of the membrane potential of the neuron
Virtual internal state variable rp corresponding to the mean_pTo represent
And output vp _pIs a nonlinear function that describes the characteristics of neurons.
It can be represented by the number f.

[0061]

(Equation 2)

Where b2 is a threshold value supplied from the bias unit B2. F () is a sigmoid function.

[0063] Unit O _q of the output layer, as well as the units of the input layer is composed of a linear type neuron, the output of each unit O _q can be expressed by the following equation (4). That is, for a certain unit O _q, when an input signal vp _p, the output vq _q can represent a linear function gq. That is, for the q-th unit, a weight corresponding to the strength of each synaptic connection (the unit coupling coefficient,
That is, the coupling coefficient) for the unit M2 of the p-th second intermediate layer and w _pq, the input signal is a vp _p, the output vq _q can represent a linear function gq.

[0064]

(Equation 3)

Where b3 is a threshold value supplied from the bias unit B3.

[0066] Therefore, by inputting each value x _i of temperature of the driver seat and the front passenger seat to the unit of the input layer, each value from the unit in the output layer representing the environmental conditions of the vehicle living space is output.

It should be noted that the characteristics of each unit in the input layer described above may be such that the input is output as it is. Further, the weight (coupling coefficient) of each unit of the non-linear operation unit 70 (neural network) is learned and corrected by a learning process to be described later so that an error in known experimental data is minimized. That is, it is thought that the nervous system of an organism learns by changing the strength of synaptic connections between neurons. For this reason, in the neural network, learning / correction refers to changing the weight (coupling coefficient) and threshold value of each unit corresponding to the synapse weight. Further, the number of units in the input layer and the number of units in the intermediate layer may be optimized by a genetic algorithm, that is, the value of data may be varied by a genetic algorithm and set according to environmental conditions. The processing of this genetic algorithm is described in Japanese Patent Application No. Hei.
The technology described in Japanese Patent No. 128977 can be employed.

Next, the details of the general neural network learning process in the non-linear operation section 70 will be described with reference to FIG.
This will be described with reference to FIG. In the present embodiment, regarding the correspondence between the temperature of the driver's seat and the front passenger seat and the environmental conditions in the vehicle living space, the temperature and the like are measured by each value of the use environmental conditions, and the environmental conditions in the vehicle living space in that case are measured. Is determined in advance. By obtaining a plurality of them, the correspondence between the temperature of the driver's seat and the passenger's seat, etc., and the environmental conditions in the vehicle living space is obtained in advance as experimental data, and a plurality of teacher data used in learning. Note that a predetermined number (for example, 70 to 80% of the total) of experimental data among a plurality of teacher data is used as learning data, and the other (for example, the remaining 20
(.About.30%) is regarded as unlearned data. This is because the experimental data is used for data used in learning of the neural network and data for confirming whether the learned neural network has been optimally learned. In addition, each value such as the temperature of the driver's seat and the passenger's seat is used as input teacher data, and each value representing an environmental condition in the vehicle living space is used as output teacher data.

In setting the unlearned data from all the data, the set unlearned data can be set so as not to be extrapolated data. That is, data having a short Mahalanobis general distance can be used as unlearned data. This is because an analysis method such as a multivariate analysis targets an interpolation problem, and a solution to an extrapolation problem may include a solution having a large error. This is because even a neural network corresponds to a nonlinear multiple regression model, and it is considered that setting unlearned data that does not become extrapolated data contributes to improvement in reliability of obtained data.

First, in step 200,
Read learning data and unlearned data. In the next step 202, initialization is performed by setting a coupling coefficient (weight) and a threshold value of each unit in the neural network to predetermined values. In the next step 204, an error of each unit of the intermediate layer and the output layer is obtained in order to train the neural network using a plurality of learning data whose values such as the temperature of the driver's seat and the passenger's seat are known.

The output layer error, ie, the unit error, can be minimized by slightly changing at least one of the coupling coefficient and the threshold value. Further, the error of the intermediate layer can be obtained by an inverse calculation such as an error back propagation method (a so-called back propagation method) using the error of the output layer.

In the next step 206, the obtained coupling coefficients and threshold values are updated (rewritten), and in the next step 208, the unlearned data is updated by the neural network based on the updated coupling coefficients and threshold values. Are tested to obtain data representing an environmental condition in the vehicle living space as a value of the test result. Next step 210
Then, it is determined whether or not the value of the test result obtained in step 208 is within a predetermined range which is a criterion of convergence determination to determine whether or not convergence is achieved, or the above process is repeated a predetermined number of times. It is determined whether or not the determination has been made. If the determination is affirmative, the present routine is terminated. On the other hand, if the determination is negative, the process returns to step 204, and the above processing is repeated. In this way, when learning data is input, each coupling coefficient and threshold value are determined so that the error of each unit of the intermediate layer and the output layer is minimized.

As described above, the neural network is learned using a plurality of experimental data in which the correspondence between the temperature of the driver's seat and the passenger seat and the environmental conditions in the vehicle living space is known. That is, learning is performed so that the error of the output value from the output layer of the neural network with respect to the teacher signal is minimized. As described above, when values such as the temperatures of the driver's seat and the passenger's seat are input to the non-linear calculation unit 70 by learning, the non-linear calculation unit 70 outputs a value representing an environmental condition in the vehicle living space.

After the above processing is completed and learning of the neural network is sufficiently performed, the network structure, that is, the coupling coefficient and the threshold value may be stored in a memory to construct a conversion system. Good.

By the way, as described above, the correspondence between the temperature of the driver's seat and the passenger's seat, etc. and the environmental conditions in the vehicle living space is unitarily constructed by one multilayer neural network (see FIG. 6). Since all units must be connected to each other and teacher data must be prepared for all combinations of input data and output data, the number of teacher data becomes enormous, and learning is completed quickly and easily. I can't let it.

Therefore, in this embodiment, inputs are classified for significant data, and a neural network is constructed for the classified data. Then, the output value from the neural network constructed in each case is set as an intermediate value, and this can be reused.

That is, in the case of the vehicle ECON apparatus, the causal relationship between the input and the output (for example, which sensor output value is used to determine the driving amount) is often clear. For this reason,
The input can be classified.

The input data includes a direct environmental temperature consisting of the driver's seat set temperature x ₁ , the passenger's seat set temperature x ₂ , the indoor temperature x ₃ , and the outdoor temperature x ₄ , the driver's seat insolation x ₅ , and the passenger's seat insolation and indirect environmental state consisting x _6, driver's seat mode x _7,
It is classified into a set conditions that from the passenger seat mode x _8. From the direct environmental temperature and the indirect environmental state, an arbitrary state temperature for each of the driver's seat and the passenger's seat can be estimated. In addition, the driver's seat outlet temperature y ₁ , the passenger seat outlet temperature y ₂ , the driver's seat air volume y ₃ , and the passenger seat air volume y, which are output data.
₄ are interrelated. Therefore, the driver's seat air volume y ₃ and the passenger seat air volume y ₄ are the driver's seat air temperature y ₁ and the passenger seat air temperature y.
_The optimum value can be set by determining each of the _two .

Therefore, as shown in FIG. 8, the non-linear operation unit 70 of the present embodiment includes a plurality of neural networks (hereinafter abbreviated as nn) 70A, 70B, 70C, 7
OD, 70E, 70F, and 70G, and as a whole, a single neural network (hereinafter abbreviated as NN).

The nn70A has an input having the driver's seat set temperature x ₁ , the passenger's seat set temperature x ₂ , the indoor temperature x ₃ , and the outdoor temperature x ₄ as input data, and intermediate data dt ₁ and pt ₁ as outputs. Configuration.

[0081] nn70B is driver's set temperature x _1, room temperature x _3, the outdoor temperature x _4, the driver's seat insolation x ₅ as constituting an input for receiving data, the intermediate data is output d
The configuration is assumed to be t2. The nn70C has a configuration in which the passenger seat set temperature x ₂ , the indoor temperature x ₃ , the outdoor temperature x ₄ , and the passenger seat insolation x ₆ are input data, and the output intermediate data is pt2.

[0082] nn70F as constituting an input for receiving data to intermediate data dt1, dt2 related to the driver's seat, a configuration in which the output driver seat air temperature y _1. Also, nn70G is the intermediate data pt related to the passenger seat.
1, pt2 as constituting an input for receiving data, a configuration in which the output passenger seat air temperature y _2.

[0083] nn70D is input with the driver's seat air temperature y ₁ and driver's mode x _7, the output driver seat air volume y _3. Further, Nn70E is input with the passenger seat air temperature y ₂ and the front passenger's seat mode x _8, the output passenger seat air volume y _4.

[0084] Here, the driver's seat air temperature y ₁ is the driver's seat temperature setting x _1, room temperature x _3, the outdoor temperature x _4, can be determined from the driver's seat insolation x _5. However, this driver's seat outlet temperature y
₁ is determined by the driver's seat set temperature x ₁ and the indoor temperature x ₃
And nonlinearity weak component of outdoor temperature x _4, the driver's seat temperature setting x _1, since non-linearity of alternating action of the indoor temperature x ₃ and the outdoor temperature x ₄ and driver's seat insolation x ₅ is stronger component, By separating these components, the construction of the NN is easy and the calculation is easy.

Similarly, the passenger seat outlet temperature y ₂ can be determined from the passenger seat set temperature x ₂ , the indoor temperature x ₃ , the outdoor temperature x ₄ , and the passenger seat solar radiation x ₆ . However, this passenger seat outlet temperature y
₂ is determined by the passenger seat set temperature x ₂ and the indoor temperature x ₃
And nonlinearity weak component of outdoor temperature x _4, the passenger seat set temperature x _2, since non-linearity of alternating action of the indoor temperature x ₃ and the outdoor temperature x ₄ and the front passenger's seat insolation x ₆ is stronger component, By separating these components, the construction of the NN is easy and the calculation is easy.

[0086] Therefore, in the present embodiment, the non-linear with respect to the driver's seat air temperature y ₁ is responsible for a weak component nn7
0A and nn70B which is responsible for a component having strong nonlinearity. The output of nn70A Is the driver's seat air temperature y ₁ when the driver's seat insolation x ₅ = 0 and the intermediate data dt1. The output of nn70B Is the correction value of the nonlinear by the driver's seat insolation x ₅ and the intermediate data dt2.

Note that regarding the driver's seat, the input is the driver's seat set temperature.
Degree x₁, Indoor temperature x_Three, Outdoor temperature x _Four, Driver's seat insolation x
_FiveNeural network that outputs intermediate data dt1 as
Regarding the work and the passenger seat, the input is the set passenger seat temperature x_Two,
Indoor temperature x_Three, Outdoor temperature x_Four, Passenger seat solar radiation x₆As
Neural network for outputting intermediate data pt1
And are very similar. Therefore, in the present embodiment,
Together with the driver's seat set temperature x₁, Passenger seat set temperature
x_Two, Indoor temperature x_Three, Outdoor temperature x_Four, Driver's seat solar radiation
x _Five, Passenger seat solar radiation x₆With intermediate data dt
The neural network that outputs 1, pt1 is nn7
OA. This is the case if the nonlinearity is weak.
The generated neural network is more connected between units
This is because the number decreases.

Driver air volume y_ThreeIs determined by the driver's seat
Constant temperature x₁, Indoor temperature x_Three, Outdoor temperature x_Four, Driver's seat insolation
Quantity x_Five, Driver mode x₇It is. Above driver's seat outlet temperature
Degree y ₁Is the driver's seat set temperature x₁, Indoor temperature x_Three， Outdoor temperature
x_Four, Driver's seat insolation x_FiveIs determined by
Driver air volume y_ThreeIs the driver's seat outlet temperature y₁And driver's seat
Code x₇Can be obtained from

Similarly, the determination of the passenger seat air volume y ₄ is as follows.
Passenger seat set temperature x _2, room temperature x _3, the outdoor temperature x _4, the passenger seat insolation x _6, a front passenger seat mode x _8. The above-described passenger seat outlet temperature y ₂ is a passenger seat set temperature x ₂ , a room temperature x ₃ ,
Outdoor temperature x _4, since they are determined by the passenger seat insolation x _6, the front passenger's seat air volume y ₄ can be determined from the passenger's seat air temperature y ₂ and the front passenger's seat mode x _8.

Each of these nns 70A to 70G can be constructed with the general configuration shown in FIG. In this case, the above-described configuration can be realized by changing hardware and software (subroutines) having the same configuration only with coefficients and the like.

With the above configuration, the number of connections between units in the neural network can be reduced.

Next, the details of the learning process of the NN by the nonlinear operation unit 70 (FIG. 8) having the above configuration will be described with reference to FIG. First, in step 200, nn for learning is set by setting the variable i to 1. Note that i
= 1, 2,..., 7, nn70A, 70B,
70C, 70F, 70G, 70D, and 70E are assumed to be supported. Next, the learning data and the unlearned data are read in the same manner as in the process of FIG. 7 (step 200),
A coupling coefficient (weight) and a threshold value of each unit are set and initialized (step 202), and an error of each unit of the intermediate layer and the output layer is obtained using a plurality of known learning data (step 204). Next, each coupling coefficient and the like are updated (step 206), tested (step 208), and the above processing is repeated until convergence (Yes at step 210). When nn converges, in the next step 214, it is determined whether or not the above processing has been completed for all the classified nns. If the processing has been completed for all the nns, this routine ends. On the other hand, if a negative determination is made in step 214, the variable i is incremented by 1 in the next step 216, and the process returns to step 200 to repeat the above processing for the next nn. Thereby, when learning data is input, the coupling coefficient and threshold value of each nn are determined so that the error of each unit of the intermediate layer and the output layer is minimized.

In this way, the neural network is trained using a plurality of experimental data in which the correspondence between the temperature of the driver's seat and the passenger's seat, etc. and the environmental conditions in the vehicle living space is known. That is, learning is performed so that the error of the output value from the output layer of the neural network with respect to the teacher signal is minimized. As described above, when values such as the temperatures of the driver's seat and the passenger's seat are input to the non-linear calculation unit 70 by learning, the non-linear calculation unit 70 outputs a value representing an environmental condition in the vehicle living space.

Next, the operation of the vehicle air conditioner 40 of the present embodiment will be described with reference to the flowchart of FIG. When the power of the vehicle air conditioner device 40 is turned on or an instruction to start execution is given from the keyboard, the process proceeds to step 220 in FIG. 10 and the values from the input device 58 and the sensor 60 are read as input data. The input data, the driver's seat temperature setting x _1, a passenger seat set temperature x _2, room temperature x _3, the outdoor temperature x _4, the driver's seat insolation x _5, passenger seat insolation x _6, the driver's seat mode x _7, passenger seat it is a mode x _8. The driver's mode x ₇ and the front passenger's seat mode x _8, the value representing the blowing mode to set the direction of the wind blown to the feet or the face or the like is input.

In the next step 222, a driving temperature is first calculated from the read input data x _{1 to} x ₈ . That is, the driver seat outlet temperature y ₁ and the passenger seat outlet temperature y ₂ are calculated, and in the next step 224, the driver seat air volume y ₃ and the passenger seat air volume y ₄ are calculated using the drive temperature and the mode. In the next step 226, the computed driver's seat air temperature y _1, a passenger seat air temperature y _2, the driver's seat air volume y _3, and outputs the passenger seat air volume y ₄ to the drive unit 62 as output data.
As a result, the driving device 62 sets the driver's seat outlet temperature y ₁ ,
The front passenger seat air temperature y _2, driver's seat air volume y _3, the front passenger seat air volume y ₄
And the vehicle interior space is set to an optimal environment.

As described above, in the present embodiment, inputs are classified for significant data, and a neural network is constructed for the classified data. Then, the output value from the neural network constructed in each case is set as an intermediate value, and the intermediate value is reused to increase the number of units connected in the NN to make the configuration simple, without causing a complicated configuration. Thus, the NN can be constructed, and the calculation load can be suppressed.

Next, a second embodiment will be described. In the present embodiment, the present invention is applied to an air conditioner for a vehicle that enables a season or a control intended by a driver without increasing the scale. In this embodiment, since the configuration is substantially the same as that of the above-described embodiment, the same portions are denoted by the same reference numerals and detailed description thereof will be omitted.

Since the vehicle air conditioner is used in a relatively narrow cabin, it has high accuracy (for example, an error of 0.5%).
) Is required, and the degree of nonlinear control is large because the influence of solar radiation is great. Therefore, N
Control by N is suitable, but fine adjustment by the operators may be required in some cases. In this case, there is a problem that it is necessary to facilitate a known data correspondence from the beginning, and that learning and reconstruction are required.

Therefore, in the present embodiment, the correction amount required for such fine adjustment is obtained by fuzzy inference.

The nonlinear arithmetic unit 70 constructed by a neural network and constructed in the vehicle air conditioner 40 of the present embodiment has the same basic configuration as that of the above embodiment. The different parts are a combiner 74 on the output side of nn70F and a combiner 76 on the output side of nn70G, as shown in FIG.
Has been established. These combiners 74 and 76 are connected to the output side of the fuzzy inference unit 72. Input side of the fuzzy inference device 72, the driver's seat temperature setting x _1, a passenger seat set temperature x _2, room temperature x _3, the outdoor temperature x _4, the driver's seat insolation x _5, each data of the passenger's seat insolation x ₆ The input is connected to the sensor 60.

As is well known, the fuzzy inference device 72
The output value is estimated from the input value by fuzzy inference. In the present embodiment, the driver output temperature x ₁ , the passenger seat temperature x ₂ , the indoor temperature x ₃ , the outdoor temperature x ₄ , the driver's seat insolation x ₅ , the passenger's seat insolation x are input sensor outputs. Estimate the current season from ₆ . The description of the fuzzy inference is omitted, but in brief, the linguistic value is quantified by a membership function using a plurality of conditional statements “if,” and finally the desired output value (estimated value ).

In this embodiment, the fuzzy inference device 72
Has a map 72A. This map 72A is
The correspondence between the estimated season and the correction amount is mapped and stored, and an example is shown in Table 1 below.

[0103]

[Table 1]

The above-mentioned correction amount indicates a step value (operation scale) of the driving device 62. For example, once the temperature is set, the adjustment range is divided into 10 steps in the case of the air flow. It is set in advance such as one step.

Next, the operation of the present embodiment will be described. What
Note that the learning of NN is the same as in the above embodiment,
Omitted. The power of the vehicle air conditioner 40 is turned on or
When an instruction to start execution is given from the keyboard, FIG.
The processing routine is executed, and the input device 58 and the sensor 60
Is read as input data (step 22).
0), read input data x₁~ X₈First drive
Dynamic temperature (driver's seat outlet temperature y ₁, Passenger seat outlet temperature y_Two)
Calculation is performed (step 222).

In the next step 230, the current season is estimated by fuzzy inference based on the read input data x _{1 to} x ₈ , a correction amount is obtained by referring to the map 72A, and in the next step 232, the above calculation is performed. The corrected driving temperature (driver seat outlet temperature y ₁ , passenger seat outlet temperature y ₂ ) is corrected. Next, using the corrected drive temperature and mode driver seat air volume y _3, calculates the passenger seat air volume y ₄ (Step 22
4). Then, the calculated driver's seat air outlet temperature y ₁ , passenger seat air outlet temperature y ₂ , driver air flow y ₃ , passenger air flow y ₄ are output to the drive device 62 as output data (step 22).
6). Thus, the driving device 62, the driver's seat air temperature y _1, a passenger seat air temperature y _2, the driver's seat air volume y _3, driven by the passenger seat air volume y _4, the vehicle living space is set to the optimum environment.

As described above, in the present embodiment, the current season is estimated by fuzzy inference, and the output data is corrected. Therefore, there is no need to learn the NN for each season or to reconstruct the NN. A wide variety of drive controls are possible only by building a basic NN. For this reason, the NN can be constructed with a simple configuration without increasing the number of units connected in the NN to make the configuration complicated or reconstructing, and the calculation load can be suppressed.

In the above embodiment, the case where the driving amount is corrected by obtaining the correction amount by fuzzy inference of the season has been described. However, the present invention is not limited to this.
The present invention is also applicable to a case where a fine adjustment based on a difference of a vehicle type or a correction according to a passenger's preference is performed. In this case, operability is improved because re-learning is unnecessary.

Next, a third embodiment will be described. In addition,
This embodiment has substantially the same configuration as the above-described embodiment, and therefore, the same portions are denoted by the same reference numerals and detailed description thereof will be omitted.

The non-linear operation unit 71 constructed by a neural network and constructed in the vehicle air conditioner 40 of the present embodiment has the same basic configuration as that of the above embodiment. The difference is that, as shown in FIG. 13, a novel neural network 78 is provided in place of the fuzzy inference device 72 of FIG. The output side of this nn78 is connected to one input side of the combiners 74 and 76. nn
Input of 78, the driver's seat temperature setting x _1, a passenger seat set temperature x _2, room temperature x _3, the outdoor temperature x _4, the driver's seat insolation x _5, each data of the passenger's seat insolation x ₆ input Connected to the sensor 60 as described above. Note that nn78 is
As in the above embodiment, learning is performed using known data as teacher data.

Next, the operation of the present embodiment will be described. What
Note that the learning of NN is the same as in the above embodiment,
Omitted. The power of the vehicle air conditioner 40 is turned on or
When an instruction to start execution is given from the keyboard,
The processing routine is executed, and the input device 58 and the sensor 60
Is read as input data (step 22).
0), read input data x₁~ X₈First drive
Dynamic temperature (driver's seat outlet temperature y ₁, Passenger seat outlet temperature y_Two)
Calculation is performed (step 222).

In the next step 240, a correction amount corresponding to the current season by nn78 is obtained from the read input data x _{1 to} x ₈ , and in the next step 232,
The calculated drive temperatures (driver's seat outlet temperature y ₁ , passenger seat outlet temperature y ₂ ) are corrected. Next, using the corrected drive temperature and mode driver seat air volume y _3, calculates the passenger seat air volume y ₄ (step 224). The driving device 62 uses the calculated driver's seat air temperature y ₁ , passenger seat air temperature y ₂ , driver air volume y ₃ , passenger air volume y ₄ as output data.
(Step 226). As a result, the driving device 62 sets the driver seat outlet temperature y ₁ , the passenger seat outlet temperature y ₂ ,
Driver's seat air volume y _3, driven by the passenger seat air volume y _4, the vehicle living space is set to the optimum environment.

As described above, in the present embodiment, the correction value is obtained by nn having a simple configuration of only correction and the output data is corrected, so that the entire NN can be learned every season,
There is no need to construct again, and various drive controls are possible only by constructing a basic NN and adding nn for correction. For this reason, the NN can be constructed with a simple configuration without increasing the number of units connected in the NN to make the configuration complicated or reconstructing, and the calculation load can be suppressed.

As described above, in each of the above embodiments, there is no need to create a special sensor, ECU, or special processing for each control device, and it is possible to easily apply the above processing only by realizing the above processing. Thus, it is possible to provide a control device that is simple and has a reduced calculation load.

In the above embodiment, the case where the present invention is applied to the driver's seat and the front passenger's seat has been described. However, the present invention is not limited to this, and can be applied to a rear seat or an arbitrary seat position of a large vehicle. Is also possible.

[0116]

As described above, according to the first aspect of the present invention, the connection relation of the neural circuit elements is determined so as not to be affected by the input data of the type having low dependency. Connection relationship can be reduced,
There is an effect that operation efficiency can be improved.

Further, even when the input data interacts with a plurality of output data, each of them can be input in duplicate, so that the connection relation of the neural circuit elements can be easily determined, Has the effect of reducing the connection relationship and improving the calculation efficiency.

According to the second aspect of the present invention, even when input data and output data have a linear or non-linear relationship, they can be separately learned. When the linearity is weak, that is, when the nonlinearity is strong, insufficient learning and when the linearity is weak, overlearning can be suppressed, and there is an effect that the calculation efficiency can be improved.

According to the third aspect of the invention, the output data can be corrected to a predetermined tendency by the correction means, so that there is an effect that fine adjustment can be performed without deteriorating the calculation efficiency. is there.

According to the invention described in claim 6 , the input data is environmental information around the vehicle and predetermined setting information, and the output data is device information such as the set temperature and air volume of the vehicle air conditioner. There is an advantage that the multi-input multi-output control device can be easily used for an air conditioner.

[Brief description of the drawings]

FIG. 1 is a conceptual configuration diagram for explaining a multi-input multi-output control device of the present invention.

FIG. 2 is a conceptual configuration diagram for explaining another example of the multi-input multi-output control device of FIG. 1;

FIG. 3 is a conceptual configuration diagram for explaining another example of the multi-input multi-output control device of the present invention.

FIG. 4 is a schematic configuration diagram of the vehicle air conditioner according to the first embodiment.

FIG. 5 is a block diagram showing a schematic configuration of each function of the vehicle air conditioner according to the first embodiment.

FIG. 6 is a general conceptual configuration diagram of a neural network.

FIG. 7 is a flowchart showing a flow of a learning process of the neural network.

FIG. 8 is a conceptual configuration diagram of a neural network of the vehicle air conditioner according to the first embodiment.

FIG. 9 is a flowchart illustrating a flow of a learning process of the neural network according to the first embodiment.

FIG. 10 is a flowchart showing a flow of operation of the vehicle air conditioner according to the first embodiment.

FIG. 11 is a conceptual configuration diagram of a neural network of a vehicle air conditioner according to a second embodiment.

FIG. 12 is a flowchart showing an operation flow of the vehicle air conditioner according to the second embodiment.

FIG. 13 is a conceptual configuration diagram of a neural network of a vehicle air conditioner according to a third embodiment.

FIG. 14 is a flowchart showing a flow of operation of the vehicle air conditioner according to the third embodiment.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 10 Corresponding means 12 Input data 14 Input means 16 Intermediate means 18 Output means 20 Output data 30 Construction means 40 Vehicle air conditioner 42 Computer main body 56 Monitor 58 Input device 60 Sensor 62 Drive device 70 Nonlinear operation part

──────────────────────────────────────────────────続 き Continuation of the front page (51) Int.Cl. ⁷ Identification code FI G06F 15/18 550 G06F 15/18 550C (56) References JP-A-5-265509 (JP, A) JP-A-6-99830 (JP, A) JP-A-6-191255 (JP, A) JP-A-8-16788 (JP, A) JP-A-4-51385 (JP, A) JP-A-5-324013 (JP, A) Kaihei 5-101028 (JP, A)

Claims (18)

(57) [Claims] 1. A multi-input multi-output control device for outputting a plurality of types of output data having a correspondence including a non-linear correspondence to a plurality of types of input data that has been input, wherein the input data is input. Input means, output means for outputting the output data, and intermediate means for transmitting data from the input means to the output means, and a number of neural circuits are provided between the input means and the output means. A neural network model having a multilayer structure in which the connection relation of the elements is modeled, and a linearity characteristic for each of a plurality of types of intermediate data as an output of a predetermined neural circuit element included in the intermediate means. Classifies the input means with an input group consisting of a type of input data having a high dependency determined by a high data relationship and a highly nonlinear data relationship, and corresponds to the classified input group. The connection relation of the neural circuit elements is determined so as not to be affected by the input data of the low dependence type until the neural circuit element of the intermediate means, and between the intermediate means and the output means. The corresponding means that determines the connection relationship of the neural circuit element of, The correspondence between the multiple types of input data and the multiple types of output data is known in advance using a plurality of relationship data, And a construction means for constructing a correspondence relationship between the types of input data and the many types of output data. 2. The classification means according to claim 1, wherein the correspondence between the input data belonging to the classified input group and the plurality of types of output data is predetermined using specific data for each input group. A correspondence relationship between input data belonging to a group and a large number of types of output data is constructed, and a connection relationship of the neural circuit elements between the intermediate means and the output means is based on the constructed correspondence relationship for each input group. The multi-input / multi-output control device according to claim 1, wherein adjustment is performed. 3. The multi-input multi-output control device according to claim 1, further comprising a correction unit that corrects output data of the output unit to a predetermined tendency. 4. Estimating seasonal or vehicle type differences or occupant preferences from the input data by fuzzy inference,
A correction amount is obtained by referring to a map in which the correspondence between the estimated season or vehicle type or the occupant preference and the correction amount is mapped, and the output data of the output means is corrected by the obtained correction amount. 3. The multi-input multi-output control device according to claim 1, further comprising a fuzzy inference device. 5. A neural network for calculating a correction amount corresponding to a season or a vehicle type difference or a passenger's preference from the input data, and correcting the output data of the output means with the obtained correction amount. The multi-input multi-output control device according to claim 1 or 2, wherein 6. The vehicle according to claim 1, wherein the input data is environmental information around the vehicle and predetermined setting information, and the output data is device information of a vehicle air conditioner. 5. The multi-input multi-output control device according to any one of 5. 7. A multi-input multi-output control method for outputting a plurality of types of output data having a correspondence including a non-linear correspondence with respect to a plurality of types of input data which has been input, the method comprising: Between the input and the output, a multi-layered neural network model that models the connection relationship of a large number of neural circuit elements, and a predetermined predetermined neural circuit element included between the input and the output. Classifying the input data into an input group consisting of a highly dependent input data type determined by a highly linear data relationship and a highly nonlinear data relationship for each of a plurality of types of intermediate data as outputs, The connection relationship of the neural circuit elements is determined so that the neural network elements corresponding to the input group are not affected by the input data of the type having low dependency. A connection relationship between the plurality of types of input data and the plurality of types of output data, wherein a connection relationship between the plurality of types of input data and the number of types of output data is determined in advance. data using, before
A multi-input multi-output control method, wherein a correspondence relationship between the plurality of types of input data from the input to the output of the input data and the plurality of types of output data is established. 8. The construction of the correspondence relation is such that the correspondence relation between the input data belonging to the classified input group and the plurality of types of output data is determined using specific data for each input group in which the correspondence relation is known in advance. A correspondence between input data belonging to the input group and a large number of types of output data is constructed, and the neural network between the output of the intermediate data and the output of the output data is constructed based on the constructed correspondence for each input group. The multi-input / multi-output control method according to claim 7, wherein a connection relation of the circuit elements is adjusted. 9. The multi-input multi-output control method according to claim 7, wherein the output data is further corrected to a predetermined tendency. 10. A seasonal or vehicle type difference or occupant preference is estimated from the input data by fuzzy inference, and a correspondence relationship between the estimated season or vehicle type difference or occupant preference and a correction amount is mapped. obtain a correction amount by referring to a map with claim 7 or 8, characterized in that correcting the output <br/> force data by the correction amount determined the
2. The multi-input multi-output control method according to 1. 11. The method of claim obtain a correction amount corresponding to the difference or the passenger's preference seasons or vehicle type from the input data by a neural network, and corrects the output data by the correction amount determined the 7 Or a multi-input / multi-output control method according to item 8. 12. The vehicle according to claim 7, wherein the input data is environmental information around the vehicle and predetermined setting information, and the output data is device information of a vehicle air conditioner. 12. The multi-input multi-output control method according to any one of the eleventh to eleventh aspects. 13. A multi-input multi-output control program for outputting, by a computer, many types of output data having a correspondence including a non-linear correspondence with respect to a plurality of types of input data inputted. A recording medium, wherein a period from the input to the output of the input data is constituted by a multilayer-structured neural network model that models a connection relationship between a large number of neural circuit elements, and is included between the input and the output. An input comprising a highly dependent input data type determined by a highly linear data relationship and a highly nonlinear data relationship for each of a plurality of types of intermediate data as outputs of predetermined predetermined neural circuit elements The input data is classified by groups, and the influence from the input data of the type having low dependency on the neural network elements corresponding to the classified input group is obtained. The connection relations of the neural circuit elements are each determined so as not to be received, and the connection relations of the neural circuit elements from the output of the intermediate data to the output of the output data are determined. the correspondence relationship between the number of types of output data in advance using a known relationship data and the previous
A recording medium storing a multi-input multi-output control program for establishing a correspondence relationship between said many types of input data from input to output of said input data and many types of output data. 14. The construction of the correspondence relation is such that the correspondence relation between the input data belonging to the classified input group and the plurality of types of output data is determined by using specific data for each input group in which the correspondence relation is known in advance. A correspondence between input data belonging to the input group and a large number of types of output data, and constructing the neural network between the output of the intermediate data and the output of the output data based on the correspondence established for each of the input groups. 14. The recording medium according to claim 13, wherein the connection relation of the circuit elements is adjusted. 15. The recording medium according to claim 13, wherein the output data is further corrected to have a predetermined tendency. 16. A seasonal or vehicle type difference or occupant preference is estimated from the input data by fuzzy inference, and a correspondence between the estimated season or vehicle type difference or occupant preference and a correction amount is mapped. obtain a correction amount by referring to the map was, recorded multiple-input multiple-output control program according to claim 13 or 14, characterized in that correcting the <br/> force data out the by correction amount obtained the Recording medium. 17. The method according to claim 13, wherein a correction amount corresponding to a season or a difference in vehicle type or a passenger's preference is obtained from the input data by a neural network, and the output data is corrected by the obtained correction amount. 14. A recording medium recording the multi-input multi-output control program according to 14.
Body . 18. The vehicle according to claim 13, wherein said input data is environmental information around the vehicle and predetermined setting information, and said output data is device information of a vehicle air conditioner. A recording medium recording the multi-input / multi-output control program according to any one of claims 17 to 17.