Neural network learning device

• US 5,727,131 A
• Filed: 10/22/1993
• Issued: 03/10/1998
• Est. Priority Date: 10/23/1992
• Status: Expired due to Term

• Current Assignee: Nippondenso Co., Ltd. (DENSO Corporation)
• Original Assignee: Nippondenso Co., Ltd. (DENSO Corporation)
• Inventors: Nakamura, Joji, Ito, Yuji, Tanaka, Hiroaki, Kawai, Takayoshi, Takeo, Yuji, Yoshimi, Tomohisa
• Primary Examiner(s): Hafiz, Tariq R.
• Application Number:
  US08/139,710
• Time in Patent Office: 1,600 Days
• Field of Search: 382/12-15, 382/155-159, 395/20-27, 165/42, 62/228.4, 62/179
• US Class Current: 706/23
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

A learning device that affects only the input-output relationships that should be additionally learned. A learning NN unit 8 capable of executing additional learning is provided separately from a learned NN unit 4 which is a basic control unit. The learned NN unit 4 produces a basic output in response to an input signal from a signal input unit 14, the learning NN unit 8 produces a correction amount desired by an individual person, and a desired control is performed based on the total value. When the output is changed, a difference is calculated between the changed output value and the basic output value from a first output unit 15, and the learning NN unit 8 executes the additional learning based upon the difference and the input value at this moment in compliance with a back-propagation method.

Claims

1. A neural network learning device comprising:

signal input means for providing input signals;

a basic control unit, connected to said signal input means, which has previously learned a basic control pattern and which generates basic output values based on the basic control pattern when said input signals are provided by said signal input means;

a learning neural network unit that receives said input signals from said signal input means, said learning neural network generating additional output values in response to said input signals based on an additional control pattern;

a total output unit for generating total output values based on a sum of said basic output values and said additional output values;

output changing means for changing at least one of said total output values;

difference calculation means for calculating a difference between said changed total output values and said basic output values;

a storage unit for storing additional learning data, said additional learning data including said differences calculated by said difference calculation means and said input signals, said learning neural network being trained on said additional learning data, and said additional control pattern being based on said additional learning data;

comparator means for comparing a capacity of a storage region of said storage unit with an amount of additional learning data when said total output values are changed by said output changing means;

selection means for selecting stored data from within said storage means when said comparator means determines that said amount of additional learning data exceeds said storage region capacity, said stored data being selected based on criteria including at least one of an amount of time stored and a deviation from said changed output values;

judging means for comparing said selected stored data with said additional learning data, and for determining whether said additional learning data will be stored based on said criteria; and

stored data control means for writing said additional learning data in said storage means when said judging means determines that said additional learning data will be stored.

2. A device according to claim 1, wherein said signal input means includes at least one environmental sensor for generating input signals based on data related to environmental conditions and said output changing means includes a touch panel that is manipulated by a passenger of said vehicle, said device further comprising:

additional storage means for storing said data related to environmental conditions, said basic output values, and said changed total output values, wherein said additional learning data also includes said data related to environmental conditions so that said learning neural network is also trained based on said data related to environmental conditions, and wherein said additional storage means stores said data when said touch panel is manipulated while said air-conditioning device is controlled by said basic output values in a predetermined manner under predetermined environmental conditions, and when said air-conditioning device is supplied with input signals that place said air-conditioning device in a condition different from a condition in which it is controlled by said basic control unit.

3. A device as recited in claim 1, wherein said device is employed as a control means for controlling an air-conditioning device for a vehicle.

4. A device as recited in claim 3, wherein said air-conditioning device has an air duct within which is provided at least an internal air/external air switching damper, a blower, an evaporator, an air-mixing damper, a heater core and a blow port switching damper, said control means controlling at least one of said internal air/external air switching damper, blower, evaporator, air-mixing damper, heater core or blow port switching damper in response to an input signal from said input signal means.

5. A device as recited in claim 4, wherein said input signal means includes at least one environmental sensor for generating input signals based on data related to environmental conditions, and wherein said basic control unit constitutes a neural network unit that is previously trained on said data related to environmental conditions and basic control characteristics of said air-conditioning device under said environmental conditions, said basic control unit generating said basic control values for controlling said air-conditioning device in response to at least one of said input signals generated by said environmental sensors.

6. A device as recited in claim 3, wherein said input signal means has a plurality of input signal terminals, and said basic control unit produces one or more of said basic output values.

7. A device as recited in claim 3, wherein said signal input means includes at least one environmental sensor for generating input signals based on data related to environmental conditions and said output changing means includes a touch panel that is manipulated by a passenger of said vehicle, said device further including:

suitable storage means for storing said data related to environmental conditions, basic output values, and said changed total output values, wherein said additional learning data also includes said data related to environmental conditions so that said learning neural network is also trained based on said data related to environmental conditions, and wherein said suitable storage means stores said data when said touch panel is manipulated while said air-conditioning device is controlled by said basic output values in a predetermined manner under predetermined environmental conditions, and when said input signals supplied to said air-conditioning device place said air-conditioning device in a condition different from a condition in which it is controlled by said basic control unit.

8. A device as recited in claim 7, wherein said learning neural network unit is disconnected from the basic control unit while being trained on said additional learning data.

9. A device as recited in claim 1, wherein said signal input means includes at least one of an internal air temperature sensor, an external air temperature sensor, a solar radiation sensor, an outlet port temperature sensor and a water temperature sensor that constitutes an environment sensor.

10. A device as recited in claim 1, wherein said output changing means is a touch panel that can be manipulated by a passenger of a vehicle.

11. A neural network learning device comprising:

signal input means for providing input signals;

a basic control unit, connected to said signal input means, which has previously learned a basic control pattern and which generates basic output values based on the basic control pattern when said input signals are provided by said signal input means;

a learning neural network unit that receives said input signals from said signal input means, said learning neural network generating additional output values in response to said input signals based on an additional control pattern;

a total output unit for generating total output values based on a sum of said basic output values and said additional output values;

output changing means for changing at least one of said total output values;

difference calculation means for calculating a difference between said changed total output values and said basic output values;

a storage unit for storing additional learning data, said additional learning data including said differences calculated by said difference calculation means and said input signals, said learning neural network being trained on said additional learning data, and said additional control pattern being based on said additional learning data;

judging means for determining whether said differences calculated by said difference calculation means have positive values or negative values; and

absolute value conversion means for calculating absolute values of said differences when said differences are determined to have negative values;

wherein said learning neural network unit comprises;

a first neural network unit for generating positive additional output values based on a positive-side additional control pattern when input signals are provided by said signal input means;

a second neural network unit for generating negative additional output values based on a negative-side additional control pattern when input signals are input from said signal input means;

output conversion means for generating converted negative additional output values by multiplying said negative additional output values generated by said second neural network unit by -1; and

an output unit which generates an output based on a sum of said positive additional output values and said converted negative additional output values,

wherein said first neural network unit learns said positive-side additional control pattern based on said differences calculated by said difference calculation means when said differences have positive values, and

wherein said second neural network unit learns said negative-side additional control pattern based on said differences calculated by said difference calculation means when said differences have negative values.

12. A neural network learning device for controlling an air-conditioning device for a vehicle, comprising:

signal input means for providing input signals, said signal input means including at least one environmental sensor for generating input signals based on data related to environmental conditions;

a basic control unit connected to said signal input means which has previously learned a basic control pattern and which generates basic output values based on the basic control pattern when said input signals are provided by said signal input means;

a learning neural network unit that receives said input signals from said signal input means, said learning neural network generating additional output values in response to said input signals based on an additional control pattern;

a total output unit for generating total output values based on a sum of said basic output values and said additional output values;

output changing means for changing at least one of said total output values, said output changing means including a touch panel that is manipulated by a passenger of said vehicle;

a difference calculation means for calculating a difference between said changed total output values and said basic output values;

a storage unit for storing additional learning data, said additional learning data including said differences calculated by said difference calculation means and said input signals, said learning neural network being trained on said additional learning data, and said additional control pattern being based on said additional learning data;

a suitable storage means for storing said data related to environmental conditions, basic output values, and said changed total output values, wherein said additional learning data also includes said data related to environmental conditions so that said learning neural network is also trained based on said data related to environmental conditions, and wherein said suitable storage means stores said data when said touch panel is manipulated while said air-conditioning device is controlled by said basic output values in a predetermined manner under predetermined environmental conditions, and when said air-conditioning device is supplied with input signals that place said air-conditioning device in a condition different from a condition in which it is controlled by said basic control unit;

selecting means for selecting additional learning data stored within said storage means that have a smallest associated control level;

judging means for comparing said selected additional learning data with said additional learning data to be stored based on respectively associated control levels; and

stored data control means for writing said additional learning data to be stored in said storage means when said additional learning data to be stored have smaller associated control levels than said selected additional learning data within said storage means.