JP3855739B2 - Neural network learning method and program - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention allows a learning control air conditioner including an air conditioner ECU that performs air conditioning control based on an air conditioning control amount calculated by a neural network to learn the neural network from teacher data used to calculate a coupling coefficient of the neural network. The present invention relates to a neural network learning method and a program for causing a computer to execute creation of teacher data.
[0002]
[Prior art]
Japanese Patent Application Laid-Open No. 11-301239 discloses an application of the learning control air conditioner to a vehicle air conditioner. This air conditioner is provided with a neural network that learns based on an occupant's set temperature operation and the like, and performs air conditioning control based on the air conditioning control amount calculated by this neural network.
[0003]
In such a learning control air conditioner, before the air conditioner is shipped to the market, the neural network is learned with teacher data consisting of tens of thousands to hundreds of thousands of points, and the coupling coefficient of the neural network is initialized in advance. Tuning work is required. Here, if there is a defect in the teacher data, dissatisfaction such as a passenger's feeling of warmth and sound with respect to the operation of the air conditioner increases. Therefore, conventionally, the tuning operation is performed while evaluating the quality of the teacher data according to the following procedure.
[0004]
First, the air conditioner is actually operated using a neural network learned based on predetermined teacher data. Then, evaluation of warm feeling, sound feeling, etc. with respect to this operation is performed for each setting of environmental conditions (for example, outside air temperature, solar radiation amount, etc.) and occupant operations (for example, set temperature, etc.) (1) air conditioning evaluation. In addition, the said evaluation is performed for every output item (for example, face blowing temperature, air volume, blower outlet mode) of a neural network.
[0005]
After that, based on the above evaluation, we extracted points that were dissatisfied with the evaluation of warmth, sound, etc., and examined how the dissatisfaction would be resolved if the teacher data was changed (▲ 2) Examination of air conditioning evaluation results ). Then, based on the above examination, the predetermined teacher data is changed and created ([3] teacher data creation).
[0006]
Then, the neural network is learned again based on the changed teacher data (4 learning). Then, it is verified whether or not the error between the output of the learned neural network and the output data of the teacher data is within a predetermined error range (5) Learning error verification. If it is not within the predetermined error range, (3) teacher data creation and (4) learning operations are performed again so as to fall within the predetermined error range.
[0007]
Then, the neural network within the predetermined error range is mounted on the ROM (6), and the air conditioner is actually operated by the neural network mounted on the ROM to check for defects (▲). 7 ▼ ROM check). Thereafter, the above (1) air conditioning evaluation is performed again ((8) re-air conditioning evaluation).
[0008]
[Problems to be solved by the invention]
However, among the above-mentioned numbers (1) to (8), the work of studying the air conditioning evaluation result of the number (2) is an operation for manually analyzing the evaluation result and determining the teacher data improvement points. This is a complicated task and is prone to human error.
[0009]
In addition, in the air conditioning evaluation of the turn (1), if the environmental conditions and occupant operation settings are changed, not only one output item (for example, the face blowing temperature) but also other output items (for example, the air flow rate and the outlet mode) Since it changes, it is difficult to find the optimum output, and the working time also increases.
[0010]
In view of the above points, an object of the present invention is to make it possible to quickly and easily create teacher data in an air conditioner that performs air conditioning control using a neural network, without human error.
[0011]
[Means for Solving the Problems]
In order to achieve the above object, according to the first aspect of the present invention, in the learning control air conditioner including the air conditioner ECU that performs air conditioning control based on the air conditioning control amount calculated by the neural network, the coupling coefficient of the neural network is calculated. A neural network learning method for learning a neural network from teacher data used to perform communication, wherein an air conditioner ECU is communicably connected to a computer having predetermined teacher data having a predetermined input value and an output value, and the teacher is connected to the computer. Interpolating the input value and output value of the data, inputting the input value during air conditioning operation from the air conditioner ECU to the computer, calculating the output value from the data interpolated result, and calculating the calculated output value to the computer To shift the output value shifted to The process of transmitting from the computer to the air conditioner ECU without learning the Ral network, and when the learning control air conditioner is operated based on the transmitted output value and the evaluation test is completed, the shifted output value and this output The method comprises a step of making the computer change and create final teacher data interpolated based on an input value corresponding to the value, and learning a neural network based on the final teacher data .
[0012]
As a result, there is no human error, and, soon, it is possible to easily create a teacher data.
[0013]
In the invention described in claim 2, in the learning control air conditioner with air conditioning ECU which performs air conditioning control based on the air-conditioning control amount calculated by the neural network, the teacher data used to calculate the coupling factor of the neural network A neural network learning method for learning a neural network, wherein an air conditioner ECU is communicably connected to a computer having predetermined teacher data having predetermined input values and output values, and the teacher data input values and output values are connected to the computer. DOO was data interpolation, by inputting an input value in the air-conditioning operation from the air conditioner ECU to the computer, a step of computing an output value from the result of data interpolation, the computed output values for each output item of the neural network computer out that in shifting, has been this shift The value is shifted when the evaluation control is completed by operating the learning control air conditioner based on the output value transmitted from the computer to the air conditioner ECU without learning the neural network. And a step of causing the computer to change and create final teacher data interpolated based on the input value corresponding to the output value, and learning the neural network based on the final teacher data. And
[0014]
As a result, there is no human error, and, soon, it is possible to easily create a teacher data.
[0016]
The invention according to claim 3 is characterized in that the contents of the teacher data created by reflecting the shifted output value and the input value corresponding to the output value can be confirmed by a computer. As a result, it is possible to confirm how much the change degree of the teacher data has affected as a result of the shift operation.
[0017]
By the way, it is difficult to learn a portion where the change in the air conditioning control amount is abrupt. On the other hand, in the invention according to claim 4, when the shifted output makes the change in the air conditioning control amount abruptly, the change in the abrupt portion is corrected so as to be gradual. Therefore, it is possible to eliminate the portion where the change in the air conditioning control amount, which is difficult to learn, is abrupt, at the time when the teacher data is created. Therefore, learning can be performed easily.
[0018]
Further, in the invention according to claim 5, when the shifted output causes abrupt change in the air conditioning control amount, the teacher data is generated by increasing the number of teacher data near the sharpened portion. Therefore, it is possible to increase the number of teacher data in a portion where the change in the air conditioning control amount is difficult to learn, and to make the learning result fit the teacher data as much as possible at the time of learning.
[0021]
In the invention according to claim 6, in the learning control air conditioner including the air conditioner ECU that performs air conditioning control based on the air conditioning control amount calculated by the neural network, the teacher data used for calculating the coupling coefficient of the neural network A program for causing a computer connected to be communicable with an air conditioner ECU to perform data interpolation between an input value and an output value of predetermined teacher data having a predetermined input value and an output value. Inputting the input value during air conditioning operation and calculating the output value from the data interpolated result, and shifting the calculated output value, without learning the neural network The process of transmitting to the air conditioner ECU and the evaluation by operating the learning control air conditioner based on the transmitted output value. When the test is completed, the computer causes the computer to execute the shifted output value and the step of changing and creating the final teacher data subjected to data interpolation based on the input value corresponding to the output value. .
According to a seventh aspect of the present invention, in a learning control air conditioner equipped with an air conditioner ECU that performs air conditioning control based on an air conditioning control amount calculated by a neural network, a teacher used to calculate a coupling coefficient of the neural network A program for causing a computer connected to an air conditioner ECU to communicate data, and interpolating an input value and an output value of predetermined teacher data having a predetermined input value and an output value, Input the input value during air-conditioning operation from the ECU, calculate the output value from the data-interpolated result, and shift the output value calculated for each output item of the neural network. , The step of transmitting to the air conditioner ECU without learning the neural network, and the transmitted output The learning control air conditioner is operated based on the evaluation test, and when the evaluation test is completed, the final teacher data interpolated based on the shifted output value and the input value corresponding to the output value is changed and created, and Is executed by a computer.
[0022]
DETAILED DESCRIPTION OF THE INVENTION
DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments shown in the drawings will be described below.
[0023]
(First embodiment)
In the present embodiment, the teacher data generation method of the present invention is applied to a vehicle air conditioner among air conditioners that are learning controlled by a neural network. FIG. 1 is a schematic diagram showing a hardware configuration of an air conditioning electronic control device (hereinafter referred to as an air conditioner ECU) 10 provided in a vehicle air conditioner and a general-purpose computer 20 for performing the teacher data creation method of the present embodiment. The vehicle air conditioner is the same as the vehicle air conditioner described in JP-A-11-301239.
[0024]
The air conditioner ECU 10 includes a neural network that learns based on a passenger's set temperature operation and the like, and performs air conditioning control of the vehicle air conditioner based on the air conditioning control amount calculated by the neural network. The air conditioner ECU 10 is similar to the ECU described in Japanese Patent Application Laid-Open No. 11-301239, and includes a combination of a known microcomputer (air conditioning computer) and its peripheral circuits.
[0025]
The vehicle air conditioner is provided with a temperature setter for setting a set temperature tset in the passenger compartment. This temperature setter is a temperature setting means that is provided on the air conditioning operation panel of the vehicle instrument panel and is manually operated. Further, the vehicle air conditioner is provided with an inside air sensor for detecting the inside air temperature tr in the vehicle interior and an outside air sensor for detecting the outside air temperature tam outside the vehicle compartment, and further a solar radiation sensor for detecting the amount of solar radiation ts into the vehicle interior. An evaporator temperature sensor for detecting the cooling temperature (blowing air temperature) te of the evaporator and a water temperature sensor for detecting the temperature Tw of the hot water flowing into the heater core are provided.
[0026]
The air conditioner ECU 10 receives the temperature setting signal tset, the inside air temperature sensor signal tr, the outside air temperature sensor signal tam, and the like, and functions as calculation means for calculating the target blowing temperature tao, the blower voltage VM, and the like using a neural network.
[0027]
Next, the outline of the neural network will be briefly described. When a neural network gives a certain input signal (for example, tset, tam, tr), its output is set to a desired value (teacher data) set in advance. Thus, the network has a hierarchical structure having an error back-propagation learning function (back-propagation function) for correcting the coupling coefficient (synaptic load) of the neural network.
[0028]
When the teacher data is changed, the coupling coefficient (synaptic load) is corrected by repeatedly “learning” so that the output for a certain input signal becomes the changed teacher data again. That is, the correlation function (coupling coefficient) is automatically generated from a large amount of data (teacher data).
[0029]
By the way, in the vehicle air conditioner having such a learning control function, before the air conditioner is shipped to the market, the neural network is learned by using teacher data consisting of tens of thousands to hundreds of thousands of points. Tuning work such as initial setting of the coupling coefficient is required. Here, if there is a defect in the teacher data, dissatisfaction such as a passenger's feeling of warmth and sound with respect to the operation of the air conditioner increases.
[0030]
Therefore, in this embodiment, the tuning operation is performed while evaluating the quality of the teacher data according to the procedure shown in the flowchart of FIG.
[0031]
First, the air conditioner is actually operated using a neural network learned based on predetermined teacher data. Then, evaluation of warm feeling, sound feeling (for example, sound feeling for blower noise), etc. for this operation is performed for each setting of environmental conditions (for example, tam, tr, ts, etc.) and occupant operations (for example, tset, etc.) (air conditioning evaluation). (Step S10)).
[0032]
After that, based on the above evaluation, points that are dissatisfied with the evaluation of warmth, sound, etc. are extracted, and examination is made as to how the dissatisfaction is resolved when the teacher data is changed (consideration of air conditioning evaluation results (step S20). )).
[0033]
Then, using the general-purpose computer 20, the predetermined teacher data is changed and created by a method described in detail later based on the above examination (teacher data creation (step S30)). Note that reference numeral 21 in FIG. 1 indicates the teacher data group modified and created in step S30 and is displayed on the display (display device) 22 of the general-purpose computer 20.
[0034]
The input values of the teacher data 21 are a deviation value tdd determined from the difference between the passenger compartment temperature tr and the set temperature tsetd, the outside air temperature tam, the solar radiation amount ts, and the output value (air conditioning control amount) is the blower voltage. VM. Note that the vehicle air conditioner of the present embodiment can air-condition the driver's seat side and the passenger's seat side in the passenger compartment independently on the left and right sides, and the deviation value tdd is the front seat set temperature tsetd on the driver's seat side. This is a deviation value td from the passenger compartment temperature tr.
[0035]
By the way, the air conditioner ECU 10 is connected to another ECU mounted on the vehicle by a communication line W, and can communicate with a predetermined communication protocol (for example, CAN). The general-purpose computer 10 is provided with a communication line connection card 22 for connecting the communication line W. By connecting the communication line W to the card 22, the air conditioner ECU 10 and the general-purpose computer 20 are connected to each other. Communication is possible using a communication protocol.
[0036]
Thereby, in step S30, data such as tr, tam, and ts can be input from the air conditioner ECU 10 to the general-purpose computer 20, and the teacher data 21 can be created using these input data. By using the communication line W and the communication protocol that are originally present in the vehicle, it is possible to eliminate hardware modifications of the air conditioning computer. Alternatively, such hardware modifications can be minimized. Thus, the design cost of the learning control air conditioner can be reduced.
[0037]
Next, the interpolation function is calculated by performing data interpolation (for example, linear interpolation, nearest neighbor interpolation, cubic interpolation, etc.) on the input value and the output value of the teacher data 21 that has been changed, and the input value ( In this case, the operation value of the optimization function for tdd, tam, ts) is set as an output value (in this case, VM) (emulator incorporation (step S40)). Since linear interpolation is employed as data interpolation in this embodiment, the time for calculating the optimization function is much shorter than the time for learning the neural network by the conventional step (4).
[0038]
The general-purpose computer 10 is installed with application software (for example, MATLAB) that calculates an optimization function by performing data interpolation as described above.
[0039]
Next, an output value (taod) that is a calculation value of the optimization function is transmitted to the air conditioner ECU 10, and the vehicle air conditioner is actually operated based on the transmitted output value (taod). Then, it is verified whether the communication input / output value and the calculated value are defective (operation check (step S50)).
[0040]
Here, in the verification as described above, it is configured to display on the display 22 as indicated by reference numeral 22a in FIG. 1 which part of the teacher data 21 the operation of the air conditioner is based on. As a result, when the air conditioning content is unsatisfactory, it is easy to determine whether the teacher data 21 is bad or the control other than the neural network is bad.
[0041]
Next, re-evaluation of air conditioning is performed, and it is confirmed whether the dissatisfaction in step S20 is reduced (re-air conditioning evaluation (step S60)).
[0042]
The input / output values are obtained by standardizing (normalizing) the sensor signals and the like from 0 to 1, and the actually output values need to be converted back from 0 to 1. For example, the actual detection range of the internal air temperature Tr detected by the internal air sensor is normally 0 ° C. to 50 ° C., and this detection value is assigned to 0 to 1 by the normalization unit and input to the input layer of the neural network. . Since the output result from the output layer is also a value of 0 to 1, it is inversely converted into an actual value corresponding to the sensor signal or the like by a conversion map set in advance in the output conversion unit.
[0043]
Next, a method for creating the teacher data 21 in step S30, which is a main part of the present invention, will be described.
[0044]
FIG. 3 is a flowchart showing an operation of creating the teacher data 21 by the general-purpose computer 20. First, the predetermined teacher data 21 used in step S10 is read into the general-purpose computer (step S310).
[0045]
Next, using a dedicated program, the read predetermined teacher data 21 is linearly interpolated to create a control pattern (step S320). The creation procedure will be described below. FIG. 4 is a characteristic diagram of the deviation value tdd and the blower voltage VM showing a control pattern of the blower voltage VM. This characteristic diagram is a general-purpose computer 20 as shown in FIG. Is displayed on the display 22.
[0046]
Further, the display 22 displays which part of the teacher data 21 the operation of the air conditioner is based on as indicated by reference numeral 22a in FIG. As a result, when the air conditioning content is unsatisfactory, it is easy to determine whether the teacher data 21 is bad or the control other than the neural network is bad.
[0047]
Further, the display 22 displays a toolbar A for shifting the control pattern of the blower voltage VM and a toolbar B for shifting the control pattern of the face blowing temperature TAV. Here, the dotted line in the characteristic diagram shown in FIG. 4A indicates the original control pattern based on the predetermined teacher data 21 used in step S10. By operating the toolbars A and B, the control pattern can be changed. Can be shifted.
[0048]
Specifically, when the first operation for changing the data D1 on the original control pattern to the data D2 is performed using the tool bar A, the inclined portion of the original control pattern is translated (shifted) so as to pass through the data D2. ) The blower voltage characteristic VM after learning the first change operation becomes a control pattern indicated by a solid line in FIG.
[0049]
Next, in step S320, it is determined whether or not the control pattern of the blower voltage VM has been changed by operating the toolbar A (step S330). If it is determined that there has been a change, the process proceeds to step S340, where the operation of the toolbar A is learned, and the control pattern is changed.
[0050]
A specific example of this change will be described below. A one-dot chain line in the characteristic diagram shown in FIG. 4B indicates the control pattern after the first learning, and the data D3 on the control pattern after the first learning is represented by When the second operation for changing to data D4 is performed, the inclination of the inclination of the control pattern after the first learning is changed so as to pass through data D2 and data D4. The blower voltage characteristic VM after learning the second changing operation is a control pattern indicated by a solid line in FIG.
[0051]
The two-dot chain line in the characteristic diagram shown in FIG. 4C shows the control pattern after the second learning, and the third operation for changing the data D5 on the control pattern after the second learning to the data D6. When this is done, the control pattern after the second learning is changed to a slope that approximates the least squares of the data D2, data d4, and data D6, and linear interpolation is performed. The blower voltage characteristic VM after learning the third change operation becomes a control pattern indicated by a solid line in FIG.
[0052]
For the third change operation or more, a control pattern is created by obtaining a slope that approximates each data after change by least squares. Then, when the above learning is completed and when it is determined in step S330 that there is no change, the process proceeds to step S350.
[0053]
In step S350, it is determined whether or not the evaluation test is completed. If not completed, the process returns to step S330. If completed, the process proceeds to step S360, and the control pattern learned in step S340 is determined at predetermined intervals. Data is sampled, and teacher data 21 is created using a dedicated program based on these sampling data.
[0054]
As described above, according to the present embodiment, in step S30, the final teacher data 21 is reflected on the general-purpose computer 20 by reflecting the shifted blower voltage VM and the deviation value tdd corresponding to the blower voltage VM. And the neural network can be learned by using the teacher data 21. Therefore, the teacher data 21 can be created quickly and easily without human error.
[0055]
Further, according to the present embodiment, since the contents of the created teacher data 21 can be confirmed on the display 22 of the general-purpose computer 20, it is possible to determine to what extent the change degree of the teacher data 21 has been affected as a result of the shift operation. It can be confirmed on the display 22.
[0056]
Moreover, according to this embodiment, since the teacher data 21 is created using the external general-purpose computer 20 that can communicate with the learning control air conditioner, for example, the face target blowing temperature TAV, the solar radiation correction amount ts, etc. An air conditioning control amount that cannot be operated by a normal learning control air conditioner can also be shifted using the general-purpose computer 20.
[0057]
Further, according to the present embodiment, since the control pattern of the face blowing temperature TAV can be set using the toolbar B in the general-purpose computer 20, the air conditioning control amount (TAV) that cannot normally be set from the air conditioning operation panel of the vehicle instrument panel. ) And the teacher data 21 can be easily created.
[0058]
Further, according to the present embodiment, step S40 functions as an emulating unit, and the general-purpose computer 20 can directly calculate the output value (VM) instead of the air conditioner ECU 10 using the teacher data 21. Therefore, it is possible to evaluate the quality of the teacher data 21 by omitting the learning work (4), which has taken an enormous amount of time, and to directly calculate the output value (VM) by a simple method that does not take time. The number of evaluation days of the data 21 can be shortened, and consequently the design cost of the learning control air conditioner can be reduced.
[0059]
(Second Embodiment)
FIG. 5 is a schematic diagram illustrating the hardware configuration of the air conditioner ECU 10 and the general-purpose computer 20 according to the present embodiment, and the hardware configuration is the same as that of the first embodiment. In the first embodiment, the display 22 displays the tool bar A for shifting the control pattern of the blower voltage VM and the tool bar B for shifting the control pattern of the face blowing temperature TAV. In the present embodiment, these are displayed. In addition to the toolbars A and B, toolbars C, D, E, and F are also displayed.
[0060]
These toolbars A to F shift the output by the computer 20 for each output item of the neural network. Specifically, the toolbar C indicates the control pattern of the target outlet temperature taod on the driver's seat side, the toolbar D is for shifting the control pattern for the target blowout temperature taop on the passenger seat side, the toolbar E is for shifting the control pattern for the solar radiation amount ts, and the toolbar F is for shifting the control pattern for the outlet mode calculated value s.
[0061]
FIG. 6 is a flowchart showing an operation of creating the teacher data 21 by the general-purpose computer 20 according to the present embodiment, and step S330 of the flowchart shown in FIG. 3 of the first embodiment is changed to step S331 shown in FIG. The other steps are the same as in the first embodiment.
[0062]
In step S320, each control pattern can be changed by operating the toolbars A to F. In step S331, it is determined whether or not the control pattern has been changed by operating these toolbars A to F. It has become.
[0063]
(Third embodiment)
FIG. 7 is a flowchart showing an operation of creating the teacher data 21 by the general-purpose computer 20 according to the present embodiment. Step S340 of the flowchart shown in FIG. 3 of the first embodiment is replaced with steps S341, S342, shown in FIG. This is a change to S343.
[0064]
In step S341, the control pattern is changed in the same manner as the control pattern is changed in the manner shown in FIG. 4 of the first embodiment. Then, it is determined in step S342 whether or not the control pattern shifted by this change has caused the change in the blower voltage VM to be abrupt. In the present embodiment, the above determination is made based on whether or not the change in the blower voltage VM per deviation value tdd1 ° C. is greater than 1.6V.
[0065]
If it is determined in step S342 that the change in the blower voltage VM is a control pattern that changes abruptly, as shown in step S343 in FIG. 7, the change in the portion that has become abrupt in step S343 is changed. Correct it so that it is gentle. Thereafter, the process proceeds to step S350 shown in FIG. 3 of the first embodiment, and thereafter, the operation is performed in the same flowchart as in the first embodiment.
[0066]
As described above, according to the present embodiment, when the shifted output makes the change in the air conditioning control amount abruptly, correction is made so that the change in the abrupt portion becomes gentle, and the teacher data is corrected. 21 is created, it is possible to eliminate a portion where learning is difficult and the air-conditioning control amount is abruptly changed when the teacher data 21 is created. Therefore, learning can be performed easily.
[0067]
(Fourth embodiment)
FIG. 8 is a flowchart showing an operation for creating the teacher data 21 by the general-purpose computer 20 according to the present embodiment. Steps S342 and S343 of the flowchart shown in FIG. 7 of the third embodiment are abolished, and the first embodiment. Steps S351 and S352 shown in FIG. 8 are added between step S350 and step S360 in the flowchart shown in FIG.
[0068]
In step S351, similarly to step S342 shown in FIG. 7, it is determined whether or not the control pattern changed in step S341 has caused the change in the blower voltage VM to be abrupt.
[0069]
If it is determined in step S352 that the change in the blower voltage VM is a control pattern that changes suddenly, as shown in step S352 of FIG. 8, the teacher near the portion that has become sharp in step S343. Increase the number of data. Note that the teacher data indicated by the underline in the figure is increased data.
[0070]
As a result, when the shifted output suddenly changes the air conditioning control amount, the teacher data 21 is created by increasing the number of teacher data in the vicinity of the abrupt portion, which makes learning difficult. By increasing the number of teacher data in the portion where the change in the air conditioning control amount is abrupt, the learning result can be fitted to the teacher data 21 as much as possible at the time of learning.
[Brief description of the drawings]
FIG. 1 is a schematic diagram showing hardware configurations of an air conditioner ECU and a general-purpose computer according to a first embodiment of the present invention.
FIG. 2 is a flowchart showing a procedure for performing a tuning operation while evaluating the quality of teacher data according to the first embodiment.
FIG. 3 is a flowchart showing an operation of creating teacher data by the general-purpose computer according to the first embodiment.
FIG. 4 is a characteristic diagram showing a control pattern of a blower voltage VM according to the first embodiment, (a) is a characteristic diagram showing a control pattern after the first learning, and (b) is a characteristic diagram after the second learning. It is a characteristic view which shows a control pattern, (c) is a characteristic view which shows the control pattern after the 3rd learning.
FIG. 5 is a schematic diagram showing a hardware configuration of an air conditioner ECU and a general-purpose computer according to a second embodiment of the present invention.
FIG. 6 is a flowchart showing an operation of creating teacher data by a general-purpose computer according to the second embodiment.
FIG. 7 is a flowchart showing an operation of creating teacher data by a general-purpose computer according to the third embodiment of the present invention.
FIG. 8 is a flowchart showing an operation for creating teacher data by a general-purpose computer according to a fourth embodiment of the present invention.
[Explanation of symbols]
20 ... general-purpose computer, 21 ... teacher data.

Claims (7)

In a learning control air conditioner having an air conditioner ECU that performs air conditioning control based on an air conditioning control amount calculated by a neural network, the neural network that learns the neural network from teacher data used to calculate a coupling coefficient of the neural network Learning method,
The air conditioner ECU is communicably connected to a computer including predetermined teacher data having a predetermined input value and an output value,
Causing the computer to interpolate an input value and an output value of the teacher data, causing the computer to input an input value during air conditioning operation from the air conditioner ECU, and calculating an output value from the result of the data interpolation; ,
Shifting the calculated output value by the computer, and transmitting the shifted output value from the computer to the air conditioner ECU without learning a neural network;
When the learning control air conditioner is operated based on the transmitted output value and the evaluation test is completed, the final data-interpolated data is based on the shifted output value and the input value corresponding to the output value. A method for learning a neural network, comprising: a step of causing the computer to change and create teacher data and learning a neural network based on the final teacher data . In a learning control air conditioner having an air conditioner ECU that performs air conditioning control based on an air conditioning control amount calculated by a neural network, the neural network that learns the neural network from teacher data used to calculate a coupling coefficient of the neural network Learning method,
The air conditioner ECU is communicably connected to a computer including predetermined teacher data having a predetermined input value and an output value,
Causing the computer to interpolate an input value and an output value of the teacher data, causing the computer to input an input value during air conditioning operation from the air conditioner ECU, and calculating an output value from the result of the data interpolation; ,
Shifting the calculated output value for each output item of the neural network by the computer, and transmitting the shifted output value from the computer to the air conditioner ECU without learning the neural network;
When the learning control air conditioner is operated based on the transmitted output value and the evaluation test is completed, the final data-interpolated data is based on the shifted output value and the input value corresponding to the output value. A method for learning a neural network, comprising: a step of causing the computer to change and create teacher data and learning a neural network based on the final teacher data . Neural network according to claim 1 or 2, characterized in that the content of the teacher data created by reflecting an input value corresponding to the output value and the shifted output value was identifiable by the computer Learning method . If the shifted output value is assumed to be rapidly a change in the air-conditioning control amount, that a change of the became suddenly portion creates the teacher data and corrected to be gentle learning method of the neural network according to any one of claims 1 to 3, wherein. If the shifted output value is assumed to be rapidly a change in the air-conditioning control amount, characterized in that to create the training data by increasing the number of training data in the vicinity of the portion which became the rapidly The method for learning a neural network according to any one of claims 1 to 3 . In a learning control air conditioner equipped with an air conditioner ECU that performs air conditioning control based on an air conditioning control amount calculated by a neural network, the creation of teacher data used to calculate the coupling coefficient of the neural network can be communicated with the air conditioner ECU A program for causing a computer connected to the computer to execute
Data interpolation is performed between an input value and an output value of predetermined teacher data having a predetermined input value and an output value, an input value during air conditioning operation is input from the air conditioner ECU, and an output value is calculated from the result of the data interpolation. A step of calculating,
Shifting the calculated output value, and transmitting the shifted output value to the air conditioner ECU without learning a neural network;
When the learning control air conditioner is operated based on the transmitted output value and the evaluation test is completed, the final data-interpolated data is based on the shifted output value and the input value corresponding to the output value. A program for causing the computer to execute a step of changing and creating teacher data . In a learning control air conditioner equipped with an air conditioner ECU that performs air conditioning control based on an air conditioning control amount calculated by a neural network, the creation of teacher data used to calculate the coupling coefficient of the neural network can be communicated with the air conditioner ECU A program for causing a computer connected to the computer to execute
Data interpolation is performed between an input value and an output value of predetermined teacher data having a predetermined input value and an output value, an input value during air conditioning operation is input from the air conditioner ECU, and an output value is calculated from the result of the data interpolation. A process of calculating;
Shifting the calculated output value for each output item of the neural network, and transmitting the shifted output value to the air conditioner ECU without learning the neural network;
When the learning control air conditioner is operated based on the transmitted output value and the evaluation test is completed, the final data-interpolated data is based on the shifted output value and the input value corresponding to the output value. A program for causing the computer to execute a step of changing and creating teacher data .

Images