JP2002182706A - Control system with learning function - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a control system with a learning function which can improve learning precision without increasing the use quantity and the operation time of a memory and without damaging versatility in learning of operation by FNN(Fuzzy Neural Network) and the like. SOLUTION: A part 17b for learning operation for intake air volume is provided with an output value conversion function reducing the dynamic range of an output value and an input value conversion function converting an input value with respect to an area where the output value with respect to the input value is large for reducing an error between a result after learning and teacher data without increasing the number of membership functions.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a control system with a learning function for setting a control amount in a control system for mechatronics or the like by learning using a fuzzy neural network (hereinafter, FNN) or the like. Without increasing the number of membership functions
The present invention relates to a control system with a learning function suitable for improving learning accuracy without impairing generalization at the time of learning.

[0002]

2. Description of the Related Art Conventionally, in controlling an engine of a vehicle such as a car or a motorcycle, a target A / F (air-fuel ratio) is set based on a throttle opening and an engine speed, and suction is similarly performed based on a throttle opening and an engine speed. Calculate the air amount and set the target A
In some cases, the fuel injection amount is calculated and set based on / F and the intake air amount. At this time, a learning function such as FNN is added to a calculation unit that calculates the intake air amount from the throttle opening and the engine speed in order to cope with individual differences in engine performance, aging, and the like. It is considered that the input / output relationship using the engine speed as an input value and the intake air amount as an output value is used as teacher data and learned in real time during traveling. The use of FNN is for imparting generalization to the learning function. Generalization is
That is, certain data affects the data around it. In this case, the control characteristics are changed smoothly with the learned data.

[0003]

However, in the above-described conventional learning method, when a region where the teacher data of the intake air amount changes rapidly is learned, the learning accuracy is increased for the reason. Since learning is performed by assigning many membership functions, if the number of membership functions increases, the required memory and the time required for computation increase, and the learning result increases and decreases as the number of membership functions increases. The range of impact on
There is a possibility that the learning result has a characteristic that an appropriate value is output only in a narrow range of the input value. That is,
Since the control characteristic after learning tends to show a steep change at the learning point, if the membership function is increased to increase the learning accuracy, generalization at the time of learning may be impaired.

Accordingly, the present invention has been made in view of such unresolved problems of the prior art. In learning operation by FNN or the like, the present invention has been made without increasing memory usage and operation time. Further, it is an object of the present invention to provide a control system with a learning function capable of increasing the learning accuracy without impairing generalization.

[0005]

In order to achieve the above object, a control system with a learning function according to the first aspect of the present invention learns a relationship between an input value and an output value as teacher data, and learns the learning. A control system with a learning function for setting a control amount according to contents, wherein the learning is performed after converting the output value so that the dynamic range of the output value is reduced.

Therefore, if the output value is converted by using a predetermined function or a data conversion table so that the dynamic range is reduced, the learning error of the output value with respect to the input value is reduced, and the learning accuracy is improved. That is, since learning is performed after reducing the dynamic range of the output value without changing the number of teacher data, for example, when FNN is used for learning, it is not necessary to allocate a large number of membership functions, and the generalized function is accordingly reduced. The learning accuracy can be improved without impairing the operability and without increasing the memory consumption or the time required for the calculation.

In the control system with a learning function according to the present invention, the relationship between the input value and the output value is learned as teacher data, and the control amount is set according to the learning content. In a region where the output value corresponding to the input value largely changes, learning is performed after converting the input value so that the output value in the region changes gradually.

Therefore, by converting the input value using a predetermined function or a data conversion table or the like, learning can be performed after the area where the output value changes rapidly becomes smooth without increasing the number of membership functions. . That is, for example, F
When the NN is used for learning, it is not necessary to assign a large number of membership functions to a rapidly changing area, and the generalization is not impaired by that much, and the memory consumption and the time required for the calculation are not increased. Learning accuracy can be improved.

[0009]

Embodiments of the present invention will be described below with reference to the drawings. 1 to 18 are diagrams showing an embodiment of a control system with a learning function according to the present invention. First, a configuration in a case where a control system with a learning function according to the present invention is applied to engine control of a vehicle will be described with reference to FIG.
FIG. 1 is a schematic configuration diagram in which a control system with a learning function according to the present invention is applied to control of an engine including an electronically controlled fuel injection device.

In FIG. 1, an engine control system 1 includes an engine power unit 2, a fuel tank 3, various sensors, and an electronic control unit (EC) for controlling the driving of the engine to an appropriate state.
A U (Electronic Control Unit) 4, an injector 5 for injecting fuel by electronic control, an ignition coil 6 for generating a high-voltage current for ignition, a spark plug 7 for generating an electric spark by the supplied high-voltage current, and The engine includes an intake pipe 12 for sucking air and supplying fuel injected by the injector 5 into the cylinder, and an exhaust pipe 13 for discharging combustion gas.

The engine power unit 2 includes a cylinder 2a, a piston 2b reciprocating by an expansion force generated by fuel combustion in the cylinder 2a, a crankshaft 2d for converting the reciprocating motion into a rotary motion,
A connecting rod 2c for interlocking the piston 2b and the crankshaft 2d, an intake valve 2e for opening and closing the passage of the intake pipe 12 by a cam (not shown), and an opening and closing mechanism for opening and closing the passage of the exhaust pipe 13 by a cam (not shown). And an exhaust valve 2f.

Various sensors include an air-fuel ratio sensor 8 for detecting an exhaust A / F, a throttle sensor 9 for detecting a throttle opening, and a detection of an engine speed and fuel injection and ignition timing. And a suction pipe negative pressure sensor 11 for detecting the amount of intake air inside the intake pipe. In addition, although other sensors such as an engine water temperature sensor are provided, they are not directly related to the present invention, and thus description thereof is omitted.

The fuel tank 3 includes a filter 3a for removing dust from the fuel, a fuel pump 3b for sending the fuel to a pressure regulator 3c described later, and a pressure regulator for maintaining a pressure at which the injector 5 injects the fuel. 3c. That is,
After the fuel in the fuel tank 3 is removed by the filter 3a, the fuel is sent to the pressure regulator 3c by the fuel pump 3b. Inject. The injected fuel is mixed with the air flowing into the cylinder 2 a through the intake pipe 12, and an electric spark is generated by the ignition plug 7 from the high-voltage current generated by the ignition coil 6 for the fuel in the mixed state. And then ignite. The combustion of the fuel by the ignition generates a large expansion force, and reciprocates the piston 2b in the cylinder 2a. Then, this reciprocating motion is transmitted to the crankshaft 2d by the connecting rod 2c, where it is converted into a rotational motion.

In such an engine control system 1, E
As one of its functions, the CU 4 has a role of electronically controlling the injection of fuel by the electronic fuel injection device having the injector 5, and a RAM (Ran) for storing input information.
dom Access Memory) and ROM (Read Only Memory) in which a program for controlling the operation of the engine is stored.
And a CPU for executing the program.

FIG. 7 is a block diagram showing the configuration of the electronically controlled fuel injection device. Hereinafter, the configuration will be described with reference to FIG. The electronically controlled fuel injection device 16 includes a fuel injection amount control program 17 stored in a ROM of the ECU 4.
And an injector 5 for injecting fuel in accordance with the fuel injection amount determined by the program 17.

The fuel injection amount control program 17 is used to output an appropriate intake air amount serving as a target A / F when an input value conversion unit 17a, which will be described later, calculates the intake air amount and the vehicle travels in a steady state. An intake air amount learning operation unit 17b for performing a learning operation, and a data inverse conversion unit 17c for performing an inverse conversion of the output value already converted at the time of the initial learning and returning the output value to a value that is actually applicable to the calculation of the fuel injection amount. A fuel injection amount calculation unit 17d that calculates a fuel injection amount from the calculated intake air amount,
When the vehicle travels in a steady state, the target A / F and the exhaust A / F
And a correction amount calculation unit 17e for calculating a correction amount for correcting the difference between the two.

The input value conversion unit 17a outputs the result after learning,
In order to reduce the error from the teacher data, the input value is converted so that the change is gentle in an area where the output value largely changes with respect to the input value. Further, in the present embodiment, the intake air amount learning calculation unit 17b uses FNN as a learning function, and this FNN is subjected to initial learning in a laboratory or the like in advance, and the engine used at that time is used. In the case of (1), a state in which learning is performed so as to calculate an intake air amount that becomes a target A / F is used. Here, for the initial learning of the FNN, the output value is also converted so that the dynamic range of the output value is reduced in order to reduce the learning error as in the input value conversion unit 17a. The output value from the FNN needs to be inversely converted by the data inverse conversion unit 17c into a value that can be actually applied to the calculation of the fuel injection amount.

More specifically, the operation of the engine will be described.
When step 7 is executed, the intake air amount learning calculation unit 17b calculates the intake air amount from the engine speed detected by the crank angle sensor 10 and the throttle opening detected by the throttle sensor 9. At this time, the input value conversion is performed on the throttle opening by the input value conversion unit 17a, and the converted value is input to the intake air amount learning calculation unit 17b. Then, the calculated intake air amount is inversely converted by the data inverse conversion unit 17c and then input to the fuel injection amount calculation unit 17c, where the fuel injection amount is determined according to a preset target A / F. Fuel is injected into the cylinder 2 a by the injector 5.

When the running of the vehicle is stabilized and the engine is operated steadily, the correction amount calculating section 17b detects the exhaust A / A in the exhaust pipe 13 detected by the air-fuel ratio sensor 8.
A correction amount is calculated from the difference between F and the target A / F, and the intake air amount is corrected so as to become the target A / F. That is, feedback control for correcting the intake air amount by feeding back the exhaust A / F is performed. Here, in the present embodiment,
In the steady state, the intake air amount learning calculation unit 17b uses FNN to set the intake air amount as an output value and the detected throttle opening and engine speed as input values.
The corrected intake air amount that becomes / F is recorded, and this value is used as teacher data to learn the relationship between the input value and the output value in real time during traveling.

Therefore, thereafter, even in the steady state,
When similar input values are detected by various sensors, feedforward control is performed so that the intake air amount determined by the input / output relationship after learning is supplied into the cylinder 2a. For an input value that is not controlled and has an unacceptable difference between the exhaust A / F and the target A / F, the intake air amount is corrected by feedback control, and the corrected value is used as the teacher data. And learning by FNN is performed.

FIG. 9 is a diagram showing the relationship between the throttle opening and the intake air amount serving as the target A / F in the engine control system by actual numerical values. This intake air amount serves as teacher data at the time of learning. As is clear from FIG. 9, the numerical value is particularly large as the engine speed becomes higher, the dynamic range becomes larger, and the throttle opening becomes 25%. In the low opening range of less than °, the change of the output value with respect to the input value is large.

Therefore, in the present embodiment, at the time of the initial learning, the learning is performed after performing the above-described conversion for both the input value and the output value.
The electronically controlled fuel injection device 16 using the NN includes an input value conversion unit 17a for converting an input value and a data inverse conversion unit 17c for inversely converting an output value after the FNN operation. Hereinafter, the effects of the conversion of the input / output values will be briefly described first based on the simple examples shown in FIGS.

FIG. 2 shows the range 1 of the input value x of the function y = x ^2.
It is a figure which shows the relationship with the output value y (1-100) with respect to. FIG. 3 shows the input / output relationship of FIG. 2 when the output value y (1 to 100) is normalized in the range of 0 to 1 and learned using FNN for the learning function. It is a figure which shows an output value, a learning result, and its error. Note that this error indicates an error when the learning result is shifted by 0.99 with respect to the normal output value, and is the dynamic range 99 (100−100) of the normal output value (1 to 100).
It is assumed that a 1% error (0.99) with respect to 1) occurs uniformly in the entire learning result. And FIG.
The same evaluation as in FIG. 3 is performed when the output value is reduced to 1 / x.

Here, for example, assuming that the allowable error is within 2%, as shown in FIG. 3, when the dynamic range of the output value is not converted, when the input value x is 1 to 7, the error is less than 2%. Is also large and deviates from the tolerance. In particular, x = 1
In this case, the error has a very large value of 99%. As a general tendency, the smaller the input value x, the greater the effect of the 1% error on the dynamic range.

On the other hand, in FIG. 4, as for the allowable error within 2%, the learning result shows that when the input value x is 1 to 4, the error is larger than 2% and is out of the allowable range. When the input value x is 5 to 10, the error is smaller than 2% and is within the allowable range. Also, if the input value x is 1 to
In the case of, the error is not less than 2%, but is significantly reduced as compared with that before the conversion. In particular, when x is 1, the error is 9%, so that the error with respect to the learning result before the conversion is obtained. It is about 1/10 for 99%.

FIGS. 5 and 6 show the effect of the input value conversion.
It will be described based on. FIG. 5 is a diagram illustrating an example of the input / output relationship when the change in the output value y is large in a region where the input value x <15. FIG. 6 is a diagram illustrating an input / output relationship when the input value x is converted to x ′ = 5 * x−60 in the region of x <15. As shown in FIG. 6, the input / output relationship after the conversion is smaller than that before the conversion by the input value x in FIG.
The large change in the output value existing in the region <15 has disappeared and the output value has become smooth.

That is, the learning accuracy of the FNN decreases as the change in the input / output relationship with a large dynamic range of the output value or the input / output relationship with a large change in the output value with respect to the input value decreases. This is compensated for by increasing the number of membership functions to increase learning accuracy.
However, in such a case, there is a fear that generalization property, which is a characteristic of a learning function such as FNN, is impaired. In order to improve the learning accuracy without impairing the generalizability, the present invention reduces the dynamic range of a large output value by the above-mentioned two transformations, and also reduces the dynamic range of the output value with respect to the input value. On the other hand, by converting the input values, the change is made smoother, and the aim is to increase the learning accuracy while keeping the number of membership functions unchanged.

Based on the effects of the two conversions, the operation of the control with the learning function for the electronically controlled fuel injection device 16 in the engine control system 1 will be described with reference to FIGS. Here, in order to improve the learning accuracy without increasing the number of membership functions, the above-described input value conversion unit 1 is used.
According to 7a, the throttle opening as an input value is converted into an input value having a throttle opening of less than 25 °, and the output value is converted during the initial learning so as to reduce the dynamic range. The data inversion unit 17 converts the output value of the air amount learning operation unit 17b into the intake air amount.
c is inversely converted into a value applicable to the calculation of the fuel injection amount.

FIG. 8 is a flowchart showing the processing of the fuel injection amount control program 17 during normal operation.
9 is a flowchart showing a process of the fuel injection amount control program 17 in a steady state and at the time of learning. First, the flow of processing during normal operation will be described with reference to FIG. As shown in FIG. 8, the process proceeds to step S800, where the input value conversion unit 17a sets the throttle sensor 9 and the crank angle sensor 1
From 0, the throttle opening and the engine speed are obtained, and the flow shifts to step S802.

In step S802, the input value converter 17
At a, it is determined whether or not the throttle opening is less than 25 °. If the throttle opening is less than 25 ° (Yes), the process proceeds to step S804. If it is not less than 25 ° (No), the intake air amount learning calculation unit 1
The value is transmitted as it is to 7b, and the process moves to step S806. When the process proceeds to step S804, since the throttle opening is less than 25 °, the input value conversion unit 17a converts the input value. The input value is converted according to the following equation (1), where the throttle opening is THL and the converted extended throttle opening is ETHL.

ETHL = ((THL-18) * 25) ÷ 7 (1) That is, 18 ° is subtracted from the input throttle opening THL (°), and the value is subtracted. The converted throttle opening ETH is obtained by multiplying 25 and further dividing the calculation result by 7.
L. When the conversion of the input value is completed, the converted value is transmitted to the intake air amount learning calculation unit 17b, and the flow shifts to step S806.

In step S806, in the intake air amount learning calculation unit 17b, the throttle opening degree is less than 25 °, the converted one is obtained, and the throttle opening degree is more than 25 °, the detected throttle amount is obtained. The opening degree is used as it is, and the forward calculation by the FNN is performed. And F
When the intake air amount is calculated by NN, step S808
Move to Here, the forward calculation means that an input value is normally input to a control system to which the FNN is applied and an output value is calculated without performing a learning calculation by the FNN.

In step S808, the intake air amount before conversion is M_at, and the intake air amount after conversion is EM_at.
The data inverse conversion unit 17c performs inverse conversion on the output converted intake air amount according to the following equation (2). M_at = (50 * EM_at) ÷ (THL + 30) (2) That is, the converted intake air amount EM_at is multiplied by 50, and the obtained value is obtained by adding 30 to the throttle opening THL. The value obtained by the division becomes the intake air amount M_at. Upon completion of the inverse transformation, the flow shifts to step S810. Therefore,
By this inverse conversion, the converted intake air amount returns to a value applicable to the calculation of the fuel injection amount in the fuel injection amount calculation unit 17d.

Here, from the equation (2), it is understood that the output value is converted by the following equation (3). EM_at = (M_at * (THL + 30)) ÷ 50 (3) That is, the result of integrating the calculated intake air amount M_at and a value obtained by subtracting 30 ° from the input throttle opening THL. Is further divided by 50 to obtain the converted intake air amount EM_at.

In step S810, the required fuel injection amount is calculated from the calculated intake air amount and the target A / F by the fuel injection amount calculation unit 17d, and a control signal is transmitted to the injector 5. And terminate the processing. That is, as long as the engine shifts to the steady operation and the input value to be learned is not input, the processes of steps S800 to S810 are repeated, and the feedforward control by the forward calculation of the FNN is performed.

Next, the flow of processing when the engine is in steady operation and an input value to be learned is input will be described with reference to FIG. First, as shown in FIG. 18, the process proceeds to step S1800, where it is determined whether the engine is operating in a steady state and is an input value to be learned. If (Yes)
Shifts to step S1802, and if not (No), shifts to step S800 shown in FIG. 8 to perform the processing during normal operation.

If the flow shifts to step S1802, feedback control is started, the exhaust air A / F in the exhaust pipe 13 is detected by the air-fuel ratio sensor 8, and the data is transmitted to the correction amount calculation unit 17e. Move to In step S1804, the correction amount calculation unit 1
7e, the difference between the exhaust air-fuel ratio and the target air-fuel ratio is calculated, and the difference is within an allowable range (for example, 0.05%).
), And if it is within the allowable range (Ye
s) moves on to step S1806, otherwise,
The amount of intake air is corrected by calculating the correction amount, and the process proceeds to step S18.
Move to 02. That is, the correction by the feedback control is performed until the difference between the exhaust A / F and the target A / F falls within an allowable range.

When the flow shifts to step S1806, the corrected intake air amount is acquired as teacher data by the intake air amount learning calculation unit 17b, and the flow shifts to step S1808. In step S1808, since the correction of the intake air amount with respect to the input value during the steady operation at this time has been completed, the feedback control is ended, and step S181 is performed.
Move to 0.

In step S1810, the intake air amount learning operation unit 17b repeats the learning operation by the FNN so that the output value corresponding to the input value in the forward operation during steady operation becomes the same value as the acquired teacher data. And the calculation result is within an allowable error range (for example, 0.0
When the value falls within 5%), the process ends. In other words, when the operation shifts to the steady operation and the acquisition of the teacher data for the input value and the learning at that time are not completed, the feedback control in steps S1802 to S1808 and the learning calculation by the FNN in step S1810 are performed. Further, for those for which the acquisition of the teacher data and the learning by the FNN for the teacher data at that time have been completed, the processing by the feedforward of steps S800 to S810 shown in FIG. 8 will be performed thereafter.

Further, an example using actual numerical values in the present embodiment will be described with reference to FIGS. FIG. 9 is a diagram showing a relationship between the throttle opening and the engine speed before the output value conversion and the corrected intake air amount which is the teacher data. Further, FIG. 10 shows the input / output relationship of FIG. FIG. 9 is a diagram illustrating a relationship between a throttle opening, an engine speed, and a learning result when learning is performed by normalizing to a numerical value of 1; Note that this learning result shows a result at the time of being shifted from the normal output value by 0.0000134, which is the dynamic range of the normal output value 0.0134 (0.0141−0.04).
007) 1% error 0.000134 (kg / sec)
Is assumed to occur uniformly in the entire learning result.
FIG. 11 is a diagram showing the relationship between the throttle opening and the compression ratio of the output value when the output value is converted by the above equation (3) with respect to the input / output relationship of FIG. Further, FIG.
2 is the output value with respect to the input / output relationship of FIG.
13 is obtained by performing the same evaluation as in FIG. 10 on the input / output relationship in FIG.

As shown in FIG. 10, the error is large near the throttle opening of 20 ° which is a low opening range. In particular, when the engine speed is 2000 rpm and the throttle opening is 18 °, The error is 20%.
In other words, the effect of a 1% error on the dynamic range on the learning result is very large in the low opening range compared to the high opening range.

Therefore, by performing output value conversion as shown in FIG. 12, the dynamic range of the output value is reduced. FIG. 12 shows that the dynamic range of the output value is compressed and reduced at the compression ratio shown in FIG. 11, and the conversion of the output value is performed while maintaining the low opening range as shown in FIG. There is a tendency that the higher the opening, the higher the compression. Therefore, as shown in FIG. 12, the minimum value of the output value in the relationship between the throttle opening and the compressed intake air amount is 0.007242 (kg / sec), and the maximum value is 0.00.
6079 (kg / sec), and the dynamic range is 0.05355, which is less than half that before conversion. Therefore, if learning is performed after normalizing the range from the minimum value to the maximum value to 0 to 1 and the error occurs uniformly in the entire output value, a 1% error is 0.00005355.
(kg / sec), as shown in FIG. 13, the maximum value of the error is 8% or less even at around 20 °, which is the throttle opening in the low opening range where the influence of the error is large, as compared with before conversion. The error is less than half, and the error in the learning result is greatly reduced as compared to before conversion.

Next, an example of the learning result by the conversion of the input value in addition to the conversion of the output value will be described. FIG. 14 is a graph showing the relationship between the throttle opening, which is an input value, and the throttle opening before conversion by the above equation (1), the engine speed, and the corrected intake air amount, which is teacher data after output value conversion. It was made. FIG. 15 normalizes the input / output relationship of FIG. 14 when the throttle opening is between 18 ° and 90 °,
FIG. 9 is a diagram illustrating a relationship between a throttle opening, an engine speed, and a learning result when learning is performed using an 8-bit CPU. Note that this learning result is obtained when the input value is 0.281.
The learning result in the case of a shift of 25 ° is shown, assuming that a shift of 1 bit (0.28125 °) occurs in the entire input value. FIG. 16 is a graph obtained by converting the input values in the input / output relationship of FIG. 10 by the above equation (1). FIG. 17 shows the same evaluation as FIG. 15 performed on FIG.

As shown in FIG. 15, the throttle opening 25
A large error has occurred in the low opening range of less than °,
Especially at a throttle opening of 20 °, the engine speed is 2
An error of about 0.14% occurs at 000 rpm. On the other hand, in FIG. 16, the throttle opening 25
The area of less than ° is smoother than in FIG. 14 before the conversion. When the same learning as in FIG. 15 is performed on the input value that has become smooth as a whole, as shown in FIG. 17, the error with respect to the learning result is within the range of the engine speed 2,000 rpm to 6000 rpm. 100
The error range falls within a very small range such as 0.05% or less for all the values in increments of 0.

As described above, the fuel injection amount control program 17 according to the present invention, as the intake air amount learning calculation unit 17b, converts an output value having a large dynamic range so that the dynamic range becomes small. Since the initial learning by the FNN is performed after that, the error generated in the output after learning can be reduced without increasing the membership function even in the actual real-time learning calculation.

In addition, the intake air amount learning calculation unit 17b determines the input value so that the region where the output value changes greatly with respect to the input value is gentle even in the initial learning and the real-time learning. Since the learning is performed after the conversion, it is possible to reduce an error generated in the output after the learning, particularly when the input value is shifted.

Further, since the learning is performed without increasing the number of membership functions and teacher data by the conversion of the input / output values, it is possible to prevent the memory usage and the operation time from being increased by the learning operation, and Thus, it is possible to perform learning without impairing generalization. Here, the input value conversion unit 17a and the intake air amount learning calculation unit 17b of FIG.
8 corresponds to the control system with a learning function described in FIG.
800 to step S810, step S180 in FIG.
0 to S1810 are performed.

In the above embodiment, the control system with the learning function by the FNN is applied to the control object as it is. However, the present invention is not limited to this. In the model control method for modeling and controlling the control object. A control system with a learning function may be applied. Thus, for example, by applying a control system with a learning function to engine control by a model control method as shown in FIG. 19, in a volume efficiency calculation unit using a learning calculation, etc., as in the above embodiment,
By compressing the output data and expanding the input axis depending on the shape of the data, the membership function can be reduced and the learning accuracy can be improved.

In the above embodiment, FNN is used as a learning function. However, the present invention is not limited to this.
MAC (Cerebeller Model Arithmetic Computer), etc.
Any learning calculation method may be used. Further, in the above embodiment, the control system with the learning function is applied to the engine control of the vehicle. However, the present invention is not limited to this, and may be applied to any control object without departing from the gist of the present invention. good.

Further, in the above embodiment, the throttle opening and the intake air amount are converted by the equations (1) and (3), respectively. However, the present invention is not limited to this. The expression to be converted may be changed to an appropriate expression, and the object to be converted is not limited to the throttle opening and the amount of intake air, but is related to the air-fuel ratio control and is a numerical value applicable to the present invention. Anything is acceptable.

In the above embodiment, the conversion of the input value and the output value is performed by a function. However, the present invention is not limited to this, and another means such as a data conversion table for tabulating the corresponding numerical values by an array or the like may be used. You may use it.

[0052]

As described above, according to the control system with the learning function according to the first aspect of the present invention, the dynamic value of the output value having a large dynamic range can be increased without increasing the resolution. Since learning is performed after converting the output value so that the range becomes smaller, it is possible to improve the learning accuracy without increasing the number of bits of the AD converter and the CPU used for the learning operation. When a network is used, the learning accuracy can be improved without increasing the membership function, so that learning without impairing generalization can be performed.

According to the control system with a learning function according to the second aspect of the present invention, the input value can be changed without increasing the resolution in an area where the output value greatly changes with respect to the input value. Since learning is performed after converting so that the output changes smoothly, the AD converter and C
The learning accuracy can be improved without increasing the number of PU bits, and when a neural network is used as the learning function, the learning accuracy can be improved without increasing the membership function. It is possible to make learning that does not impair the transformation.

[Brief description of the drawings]

FIG. 1 is a schematic configuration diagram in which a control system with a learning function according to the present invention is applied to control of an engine including an electronically controlled fuel injection device.

FIG. 2 is a diagram illustrating a relationship between an output value y and a range 1 to 10 of an input value x of a function y = x ² .

FIG. 3 shows 1% when the input / output relationship of y = x ² is learned.
11 is a table showing output results after learning with respect to input / output values before learning when an error (0.99) occurs uniformly in the entire learning result.

4 is a table showing a result after learning when an error of 1% occurs uniformly in the entire learning result as in FIG. 3 when learning is performed after converting the dynamic range of the output value to y = 1 / x. It is.

FIG. 5 is a diagram illustrating an example of an input / output relationship when a change in an output value is large in a region where an input value x <15.

FIG. 6 is a diagram illustrating an input / output relationship when an input value x is converted into x ′ = 5 * x−60.

FIG. 7 is a detailed block diagram illustrating a configuration of an electronically controlled fuel injection device.

FIG. 8 is a flowchart illustrating a process of a fuel injection amount control program during normal operation and during learning.

FIG. 9 is a diagram showing a relationship between a throttle opening and an engine speed before output value conversion and a corrected intake air amount which is teacher data.

FIG. 10 shows the relationship between the throttle opening and the learning result when an error of 1% with respect to the dynamic range of the intake air amount when the input / output relationship before the output value conversion is learned uniformly occurs in the entire learning result. FIG.

FIG. 11 is a diagram showing a relationship between a throttle opening and a compression ratio with respect to a converted output value.

FIG. 12 is a diagram illustrating a relationship between a throttle opening and an intake air amount after output value conversion.

FIG. 13 is a diagram illustrating a relationship between a throttle opening and a learning result when an input / output relationship after output value conversion is learned.

FIG. 14 is a diagram showing a relationship between a throttle opening and an engine speed before input value conversion and a corrected intake air amount which is teacher data after output value conversion.

FIG. 15 shows an input / output relationship before an input value conversion of an 8-bit CP.
FIG. 9 is a diagram illustrating a relationship between a throttle opening degree and an engine speed when learning is performed using U and a learning result.

FIG. 16 is a diagram showing a relationship between an extended throttle opening and an engine speed after input value conversion and a corrected intake air amount which is teacher data after output value conversion.

FIG. 17 shows the input / output relationship after input value conversion by using an 8-bit CP.
It is a figure which shows the relationship between the expansion throttle opening degree and engine speed at the time of learning using U, and a learning result.

FIG. 18 is a flowchart showing a process of a fuel injection amount control program when a steady operation is performed and an input value that requires a learning operation is input.

FIG. 19 is a diagram illustrating an example in which model-based control is applied to engine control.

[Explanation of symbols]

 Reference Signs List 1 engine control system 2 engine 3 fuel tank 4 ECU 5 injector 8 air-fuel ratio sensor 9 throttle sensor 10 crank angle sensor 11 intake pipe negative pressure sensor 12 intake pipe 16 electronically controlled fuel injection device 17 fuel injection amount control program 17a intake air amount Learning operation unit

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI Theme coat ゛ (Reference) G06N 3/00 550 G06N 3/00 550E 3/08 3/08 F term (Reference) 3G084 BA09 BA13 CA05 DA13 EA04 EB17 EC04 FA07 FA10 FA11 FA29 FA33 FA38 3G301 JA19 KA21 MA01 MA11 NA08 NA09 ND01 ND21 ND42 ND43 ND45 PA01Z PA07Z PA11Z PD03A PE01Z PE03Z 5H004 GA33 GB12 HA02 HA13 HB03 HB04 HB07 HB08 HB13 JB11 KA18

Claims (2)

[Claims] 1. A control system with a learning function for learning a relationship between an input value and an output value as teacher data and setting a control amount according to the learning content, wherein a dynamic range of the output value is reduced. A control system with a learning function characterized by learning after converting the output value as described above. 2. A control system with a learning function which learns a relationship between an input value and an output value as teacher data and sets a control amount in accordance with the learning content, wherein an output value corresponds to the input value. In areas where there are significant changes,
A control system with a learning function, wherein the learning is performed after converting the input value so that the change of the output value in the area becomes gentle.