JP2003076404A - Computer with learning function - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a computer with a learning function for performing a partial learning to correct an output value by a certain quantity of learning on the basis of the relation between an output value and a target value in relation to the output value with a change of the output value and for performing the whole learning to correct the whole of coefficients by the specified quantity of learning when a learning result after the partial learning shows the predetermined tendency. SOLUTION: An intake air quantity learning computer unit 20 is formed of an engine speed computing unit 20a for computing the rotating speed of a crank shaft as an output shaft of an engine as the engine speed on the basis of a time change ratio of the crank angle, an intake air quantity computing unit 20b for computing the intake air quantity on the basis of the engine speed and the intake negative pressure to perform the learning computing on the basis of the decided learning quantity, a Pmi computing unit 20c for computing a mean effective pressure Pmi inside of a cylinder 2a on the basis of the crank angle, an A/F estimation unit 20d for estimating A/F on the basis of the computed Pmi, and a learning quantity decision unit 20e for deciding the learning quantity on the basis of the estimated A/F and a target A/F.

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD The present invention relates to a fuzzy neural network and a CMAC (Cerebller Model Arithmetic Comp).
and a learning function suitable for learning a predetermined relationship between an input value and an output value related to control in a control system in which noise is mixed in a control signal, etc. .

[0002]

2. Description of the Related Art Conventionally, a CMAC (Cerebellar Model Ari) has been used as a learning function in engine control of vehicles such as cars and motorcycles as a control system using an arithmetic unit with a learning function.
thmetic computer) or fuzzy neural network, etc., using the target A / F as teacher data to learn and control in real time during running so that the intake air amount for obtaining that A / F becomes an appropriate value. There is.

[0003]

However, in the above conventional learning method, when noise is mixed in the A / F to be actually detected, the data including the noise is fed back for learning. In addition, if the obtained data is learned as it is, the accuracy of learning becomes poor. In such a case, the influence of noise can be reduced by collecting a large number of data and processing them statistically.
For that, it is necessary to operate repeatedly under the same conditions,
In the case of engine control, it is not realistic to keep various factors such as the temperature and humidity of the atmosphere, the water temperature and hot water temperature of the engine, the condition of the road surface, the number of passengers, the weight of luggage, etc., under constant conditions. Furthermore, since learning cannot be started until a certain amount of data has been collected, the real-time learning is lacking,
Moreover, a huge memory capacity is required to buffer such data.

Further, in order to reduce the influence of noise at the time of learning, the learning gain may be made small. Even in that case, however, if the detected data contains data with a very large error, the gain of There is a risk that even a small learning will cause a learning error that cannot be ignored. Therefore, the present invention has been made by paying attention to the unsolved problem of such a conventional technique, and based on the magnitude relationship between the output value and the target value with respect to the output value, only a certain learning amount is obtained. If partial learning for correcting the output value is performed and the learning result shows a predetermined tendency after partial learning,
It is an object of the present invention to provide an arithmetic unit with a learning function that performs overall learning by correcting all coefficients by a specific learning amount.

[0005]

In order to achieve the above object, an arithmetic unit with a learning function according to a first aspect of the present invention is an arithmetic unit for calculating an output value based on a predetermined relationship according to an input value. An arithmetic unit with a learning function, comprising: a means and a learning means for performing learning so that the output value becomes a predetermined target value, wherein the learning means comprises: A magnitude relationship determining unit that determines the magnitude relationship between the target value and the output value by comparing the target value is provided, and the coefficient representing the predetermined relationship is corrected by a fixed learning amount based on the determination result. It is characterized in that a partial learning process for repeatedly performing the above is executed.

With such a configuration, the magnitude relationship determining means compares the output value with the target value, and determines whether the output value is above or below the target value. Is determined, and the learning means repeatedly performs partial learning in which the coefficient is corrected by a constant learning amount based on the determination result. Therefore, for example, when the present invention is applied to engine control of a vehicle, even if noise is mixed in the calculated output value, only a certain learning amount is calculated based on the magnitude relationship between the output value and the target value. Since the correction is performed only, the learning amount does not depend on noise, and the learning error can be reduced.

Here, the coefficient represents a predetermined relationship between an input value and an output value in the arithmetic unit with a learning function of the present invention, which is a learning arithmetic method, for example, C with two inputs and one output.
It is a numerical value that is actually corrected by learning, such as an output value in the MAC. According to a second aspect of the present invention, in the arithmetic unit with a learning function according to the first aspect, the constant learning amount is set such that a change in output value before and after learning falls within a predetermined numerical range. It is characterized by being.

That is, when the coefficient is corrected by a constant learning amount, the learning amount is set so that the change in the output value does not exceed the target value. Therefore, it is possible to reduce inappropriate correction of the coefficient, which is useful for improving the convergence speed in learning.
The invention according to claim 3 is the arithmetic unit with a learning function according to claim 1 or 2, wherein the learning means is 1
Alternatively, an n-dimensional (n is an integer of 2 or more) space formed by the relationship between a plurality of types of input values and the corresponding coefficients is divided into a plurality of subspaces, and the partial learning processing for the coefficients has the same tendency. Is indicated, the whole learning process for correcting all the coefficients by the learning amount for whole learning determined based on the tendency is executed.

That is, the learning means converts an n-dimensional (n is an integer of 2 or more) space (for example, three-dimensional) formed by the relationship between one or a plurality of types of input values and the corresponding coefficients into a plurality of subspaces. If the partial learning processing for the coefficients in several subspaces shows the same tendency, all the coefficients are determined based on the coefficients of the subspaces showing the tendency. Only the learning amount for use is corrected.

Therefore, the tendency of the learning content in several subspaces is checked to correct all the coefficients collectively, which is useful for improving the convergence speed in learning. According to a fourth aspect of the present invention, in the arithmetic unit with the learning function according to the third aspect, the learning amount for overall learning is a coefficient at the time of starting learning and learning in each of the plurality of subspaces showing the tendency. It is characterized by the average value of the difference from the coefficient at the end.

That is, when partial learning in some subspaces shows the same tendency, the average value of the difference between the coefficient at the start of learning and the coefficient at the end of learning in each of the subspaces is used as the learning amount for overall learning. As a result, all the coefficients are corrected by the learning amount. Also,
According to a fifth aspect of the present invention, in the arithmetic unit with a learning function according to the third aspect, the learning amount for overall learning has a coefficient at a learning start time and a learning end time in each of the plurality of subspaces showing the tendency. It is characterized by being the minimum value of the difference from the coefficient of.

That is, when partial learning in several subspaces shows the same tendency, the smallest value of the differences between the coefficient at the start of learning and the coefficient at the end of learning in each of the subspaces is taken as the whole. As the learning amount for learning, all the coefficients are corrected by the learning amount.

[0013]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings. 1 to 5 are diagrams showing an embodiment of an arithmetic unit with a learning function according to the present invention. First,
The configuration of a vehicle engine system to which the arithmetic unit with a learning function according to the present invention is applied will be described with reference to FIG. Figure 1
FIG. 3 is a block diagram showing a configuration of an engine system.

In the figure, an engine system 1 is an engine power unit 2, a fuel tank 3, various sensors, and an electronic control unit E for driving and controlling the engine in an appropriate state.
A CU (Electronic Controll Unit) 4, an injector 5 for injecting fuel by electronic control, an ignition coil 6 for producing a high-voltage current for ignition, an ignition plug 7 for producing an electric spark by the supplied high-voltage current, and a cylinder. It is composed of 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 which reciprocates by an expansion force generated by combustion of fuel in the cylinder 2a, and a crankshaft 2d which converts 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 intake valve 2e for opening and closing the passage of the exhaust pipe 13 by a cam (not shown). And an exhaust valve 2f.

As various sensors, an air-fuel ratio sensor 8 for detecting an exhaust A / F, a throttle sensor 9 for detecting a throttle opening, an engine speed, fuel injection and ignition timing are detected. A crank angle sensor 10 for detecting the intake air pressure and an intake pipe negative pressure sensor 11 for detecting the intake air amount inside the intake pipe are provided, and the detection results are transmitted to the ECU 4. Further, although a sensor such as an engine water temperature sensor is also provided, it is not directly related to the present invention, and therefore its explanation is omitted.

The fuel tank 3 includes a filter 3a for removing dust from the fuel, a fuel pump 3b for delivering the fuel to a pressure regulator 3c, which will be described later, and a pressure regulator 3c for holding the fuel injection pressure in the injector 5. And are equipped with. That is, the fuel in the fuel tank 3 is sent to the pressure regulator 3c by the fuel pump 3b after the dust is removed by the filter 3a, and the fuel in the fuel tank 3 is introduced into the cylinder 2a through the intake pipe 12 by the injector 5 while being kept at a high pressure. Inject fuel. The injected fuel is
Similarly, air mixed with the air flowing into the cylinder 2a through the intake pipe 12 is generated, 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 this mixed state, whereby ignition is performed. To do. Then, the combustion of the fuel by the ignition produces a large expansion force to reciprocate the piston 2b in the cylinder 2a. Furthermore, this reciprocating motion is transmitted to the crankshaft 2d by the connecting rod 2c, and is converted into a rotating motion there.

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 has a RAM (Random RAM) that stores input information.
m Access Memory), a ROM (Read Only Memory) in which a program for controlling the operation of the engine is stored,
It is provided with a CPU for executing a program, and an intake air amount learning calculation unit for calculating the intake air amount by the calculator with a learning function of the present invention.

Further, the process of calculating the fuel injection time in the ECU 4 will be described with reference to FIG. FIG. 2 is a block diagram showing a fuel injection time calculation process. The fuel injection time calculation process is performed by the intake air amount calculation CMAC described later.
An intake air amount learning calculation unit 20 that calculates an appropriate intake air amount by learning with 20b _1, a fuel amount conversion unit 21 that converts the calculated intake air amount into a fuel amount, and a conversion based on the injector QT characteristics. Processing is performed by the injection time conversion unit 22 that converts the fuel amount into the injection time.

Here, the CMAC is a method for expressing a high-order vector by a low-order vector by mapping, thereby performing learning by replacing a complicated operation with a plurality of simple operations, and multidimensional input. It is possible to suppress the complication of the processing due to the increase of the number of intermediate layers. In addition, learning is CMA
The learning calculation method is not limited to C, and other learning calculation methods may be used. The intake air amount learning calculation unit 20 acquires the crank angle detected by the crank angle sensor 10, and calculates the rotation speed of the crankshaft, which is the output shaft of the engine, as the engine speed from the time change rate of the crank angle. Engine speed calculator 20a, engine speed calculated by engine speed calculator 20a, and intake pipe negative pressure sensor 1
The intake air amount calculation unit that calculates the intake air amount from the intake negative pressure detected by 1 and performs the learning calculation of the intake air amount calculation CMAC 20b ₁ based on the learning amount determined by the learning amount determination unit 20e described later. 20b and a Pmi calculator 2 for calculating an average effective pressure Pmi in the cylinder 2a from the crank angle detected by the crank angle sensor 10.
0c, the A / F predicting unit 20d that predicts the exhaust A / F based on the calculated Pmi, and the learning amount for partial learning is determined based on the predicted exhaust A / F and the target A / F. The learning amount determination unit 20e.

Further, the structure of the intake air amount calculation unit 20b will be described with reference to FIG. FIG. 3 shows the intake air amount calculation unit 20.
It is a figure which shows the structure of b. The intake air amount calculation unit 20b
The intake air amount calculation CMAC 20b ₁ and the entire offset unit 20b ₂ are included. CMA for intake air amount calculation
C20b ₁ has an engine speed and an intake negative pressure as input values and an intake air amount as an output value. Based on the learning amount for partial learning determined by the learning amount determination unit 20e, C20b ₁ is an actual value. CMAC2 for calculating the intake air amount so that an appropriate intake air amount can be calculated according to the running conditions of the vehicle
Learning for correcting the coefficient of 0b ₁ is performed. Here, the correction of the coefficient is performed by the A / F predicted by the A / F prediction unit and the target A.
/ F is compared to determine whether the predicted A / F is on the plus side or the minus side with respect to the target value, for example, to determine the magnitude relationship between the two, and based on the determination result, The coefficient is corrected so that the predicted A / F approaches the target A / F. Here, C for intake air amount calculation
As the MAC 20b ₁ , a MAC 20b ₁ that has been subjected to initial learning so as to output a target control value set in a laboratory or a laboratory is used.

Further, the intake air amount, which is the output value after learning, is transmitted to the overall offset unit 20b ₂ before being transmitted to the fuel amount conversion unit 21, where the output value after learning and the output value before learning are transmitted. And the tendency of the learning result of the partial learning at that time is under a predetermined condition (for example, the results of the partial learning in some learning sections of the intake air amount calculating CMAC 20b ₁ described later all show an upward tendency). When the above condition is satisfied, the coefficients of all sections are corrected by the learning amount for overall learning, and the corrected intake air amount is transmitted to the fuel amount conversion unit 21. In the present embodiment, as the learning amount for the whole learning, the whole offset value is set based on the partial learning result showing the tendency, and the intake air amount calculating C is set.
A value obtained by multiplying the output value of the MAC 20b ₁ by this overall offset value is calculated as the final intake air amount.

Further, the average effective pressure Pmi is obtained by dividing the total work amount for one cycle by the stroke volume, and it is considered that the integration result for one cycle is proportional to the integration result between arbitrary crank angles. Well, especially the explosion (expansion) stroke has a strong correlation with the mood result. Therefore, Pmi has a correlation with the crank angular velocity obtained by differentiating the crank angle, and has a stronger correlation with the deviation of the crank angular velocity. Therefore, in the present embodiment, after the compression top dead center, a value integrated between the crank angle of 30 deg. And 80 deg. Is set as the average effective pressure Pmi, and a map and the like from this, the engine speed, and the intake pipe negative pressure. To predict A / F. Since this prediction has no particular relation to the present invention, detailed description will be omitted.

Further, in the learning amount determining section 20e, the target A
A learning amount for partial learning that falls within a certain range is determined so that the actual A / F predicted after correction does not exceed or fall below the target value with respect to / F. It is like this. On the other hand, when the intake air amount after learning is calculated, the fuel amount conversion unit 21 converts the intake air amount calculated by the intake air amount calculation unit 20b into the fuel amount.

Further, when the converted fuel amount is acquired, the injection time conversion unit 22 converts the fuel amount into the injection time based on the injector QT characteristic and transmits the pulse signal of the injection time to the injector 5. And the injector 5
Receives a pulse signal of the injection time, and injects fuel into the cylinder 2a based on the pulse signal.

Further, the flow of learning calculation processing in the intake air amount calculation section 20b will be described with reference to FIG. Figure 4
6 is a flowchart showing a learning calculation process in the intake air amount calculation unit 20b. Here, C for intake air amount calculation
The learning calculation by the MAC 20b ₁ is performed while the vehicle is driving, and after the vehicle is stable and the engine is in the steady state, the learning is not performed during acceleration or deceleration. I will.

As shown in FIG. 4, first, step S40.
0, the learning section in the intake air amount calculation CMAC 20b ₁ is divided, and the process proceeds to step S402. Here, the learning section is divided into an n-dimensional numerical space (n is 2) serving as a reference in accordance with the type of input value and the area of input / output values used in the steady operation of the vehicle.
The above integer) is formed, and the partial space is evenly divided within a certain range according to the number of divisions of the numerical space. Therefore, in the present embodiment, since the input value is of two types, the engine speed and the intake negative pressure, a three-dimensional numerical space is formed and further set together with the intake air amount which is the output value for these. In the region, each input axis is equally divided into three to form nine output value planes. In the present embodiment, this output value plane is called a learning section.

In step S402, the intake air amount calculating CMAC 20b ₁ set in the learning section is set.
Is duplicated and the process moves to step S404. Here, the copy is used when comparing the intake air amount calculating CMAC 20b ₁ before and after learning to determine the tendency of partial learning and when calculating a correction value of the overall offset value described later. This is because.

In step S404, the learning amount determination unit 20
Inhalation using the learning amount for partial learning determined in e
CMAC20b for air amount calculation₁Part of each learning section
Minute learning is performed and the process proceeds to step S406. Where suck
Air intake amount calculation CMAC20b₁Each learning sex in
Partial learning is predicted by the A / F prediction unit 20d.
Between the actual A / F and the target A / F set in advance
Based on the relationship, determined by the learning amount determination unit 20e
Inhale the learning amount that is a constant value so as to approach the target A / F
CMAC20b for air amount calculation ₁Person in charge of each learning section of
It is done by adding or subtracting to a number. This partial learning
Is repeated a set number of times (eg 500 times)
Becomes

When the process proceeds to step S406, it is determined whether there are three or more sections for which partial learning has finished,
If it is determined that there are three or more (Yes), step S40.
If not (No), the process proceeds to step S404. In the flowchart shown in FIG. 4, the number of sections for which partial learning is completed is determined to be three or more at the time of determination, but it may be set to an arbitrary number depending on the number of learning sections to be formed.

When the process proceeds to step S408, it is determined whether or not the learning results all show the same tendency in the learning section in which the partial learning is completed. If the tendency is the same (Yes), the process proceeds to step S410. If not (No), the process proceeds to step S418. Here, the tendency of the learning result means that a condition is satisfied in some (here, three) learning sections among the divided learning sections, for example, when the learning results all tend to increase. Is.

When the process proceeds to step S410, correction teacher data is created from the learning section related to the tendency determination, and the process proceeds to step S412. Here, the correction teacher data is for performing finer correction on the CMAC after the overall learning described later. In the present embodiment, the input / output relationship of all learning sections is used as the correction teacher data. Create a saved one.

In step S412, an offset correction value is calculated based on the learning section used for the tendency determination, and the offset correction value is added to the total offset value (this is called total learning), and the flow proceeds to step S414. That is, when the learning results of the partial learning in the three learning sections for which the partial learning is completed in the intake air amount calculating CMAC 20b ₁ tend to increase, for example, the coefficient and step of the learning section related to the tendency determination Overall learning in which a difference from the coefficient of the corresponding learning section among the learning sections in the initial state copied in S402 is calculated, and the smallest value among those differences is set as an offset correction value, and the value is added to the overall offset value. It is carried out.

For example, among the divided learning sections, the ones for which the partial learning has been completed are defined as the first to third sections.
When learning is performed such that the difference between the coefficient after learning and the coefficient before learning is 10 in the first section, 11 in the second section, and 14 in the third section, the minimum value of increase in the overall offset value Therefore, the correction of adding only 10 will be performed.

[0035] After the transition to step S414, proceeds to step S416 to copy the intake air quantity calculation CMAC20b ₁ for control using the intake air amount calculation CMAC20b ₁ replication initial state to the actual calculation of the intake air amount To do. That is, the intake air amount calculation CMAC 20 for control
b ₁ is returned to the state before partial learning. When the process proceeds to step S416, the control intake air amount calculating CMAC 20b ₁ in the initial state before the partial learning is corrected and learned by the correction teacher data, and the process proceeds to step S402.

That is, the intake air amount calculating CMAC 20b ₁ which has returned to the initial state before the partial learning is returned to the state at the time when the partial learning of three or more learning sections before the calculation of the offset correction value is completed. Therefore, the correction teacher data is used to perform correction so that the content of the partial learning of each learning section is reflected in the corresponding section. Then, the intake air quantity calculation CMAC20b ₁ is used as the intake air amount calculation CMAC20b ₁ reference for calculating the intake air amount, the output value of the intake air quantity calculation CMAC20b _1, the corrected overall offset value The multiplied value becomes the final intake air amount. Further, in the learning from the next time, a copy of the control intake air amount calculating CMAC 20b ₁ used as this reference is created and the same learning calculation as above is performed.

On the other hand, when it is determined that the learning results of the learning sections for which partial learning has ended are different in tendency, and the process proceeds to step S418, a flag indicating the end of partial learning of the learning section used for the tendency determination is set. After clearing, the process proceeds to step S404. Furthermore, the relationship between the A / F output value in the learning calculation and the A / F output value after learning is shown in FIG.
It will be described based on. Fig. 5 shows A / F with noise
It is a figure which shows the relationship between an output value and the A / F output value after learning. In the figure, the A / F output value 51 is data in which noise given by a normal distribution random number with a standard deviation of 0.5 is mixed, and is created on a trial basis. Where the target A / F
When learning is started from 13.5 with 14.5 as the learning value, the learning value A /
The F output value 50 converges to a value near 14.5 which is the target value. Here, the partial learning is to repeatedly perform a correction such that the injection amount is slightly increased when the A / F output value 51 is below the target value, and the injection amount is slightly decreased when the A / F output value 51 is above the target value. , A / F output value 50 after learning is 1
It is a constant small value that can be accommodated between 3.5 and 14.5.

In this way, the learning amount determining section 20e determines the learning amount for correcting the coefficient of the intake air amount calculating CMAC 20b ₁ based on the relationship between the detected output value and its target value, and the coefficient is calculated. Since the partial learning to be corrected by this learning amount is repeated, even if noise is mixed in the output value, the coefficient is corrected by a fixed learning amount in one partial learning. Can be reduced.

Further, the learning amount determined by the learning amount determining unit 20e is determined within a range in which the output value calculated after learning does not exceed the target value, which is useful for improving the convergence speed of learning. In addition, for learning calculation, intake air amount calculation C
Applying MAC20b ₁ to calculate intake air amount CMAC2
When the learning results of the three learning sections that have completed the partial learning have the same tendency in the plurality of learning sections at 0b ₁ , the overall offset value is corrected based on the tendency. CM for calculation
The coefficients of all learning sections of the AC 20b ₁ become equivalent to the state of being offset by the overall offset value, which helps improve the convergence speed in learning.

Further, in order to reflect the correction contents of the coefficient by the partial learning in the CMAC after the whole learning, the correction teacher data is created, and after the correction of the whole offset value, the learning result is obtained for the coefficients of all the learning sections. Since it reflects, it helps improve the speed of learning convergence. Here, FIG.
The function for determining the magnitude relationship between the predicted A / F and the target A / F in the intake air amount computation unit 20b shown in FIG. 2 corresponds to the magnitude relationship determination means according to claim 1, and the intake air amount computation unit shown in FIG. The learning calculation function in 20b is defined in claims 1, 3, and 4.
Corresponds to the described learning means.

In the above embodiment, when correcting the overall offset value, the minimum value of the difference between the coefficient before learning and the coefficient after learning in the learning section showing the same tendency in the learning result is set as the offset correction amount. Although the overall offset value is corrected as described above, the present invention is not limited to this, and the coefficient may be corrected so as to be the maximum value of the differences or the average value of the differences.

Further, in the above-mentioned embodiment, only the tendency to increase and correct the coefficient at the time of determining the tendency of the learning result is described, but the present invention is not limited to this, and the coefficient to decrease and correct may be used. Alternatively, the learning may be performed by determining another tendency, for example, when there are a certain number or more of coefficients having a certain numerical value or more. Further, in the above-described embodiment, although learning is performed for one numerical space in CMAC,
Not limited to this, learning is performed by dividing the numerical space itself into multiple by forming a new numerical space by shifting the input axis of the numerical space by the quantization unit that is the minimum unit of the input value. The accuracy may be increased.

[0043]

As described above, according to the arithmetic unit with a learning function according to the first aspect of the present invention, the output value and the target value are compared with each other by the magnitude relation determining means to determine the output value as the target. Partial learning is performed by determining the magnitude relationship between the output value and the target value such as whether it is above or below the value, and correcting the coefficient by a certain learning amount based on the determination result by the learning means. Therefore, even if noise is mixed in the calculated output value, the learning amount is not influenced by the noise and the learning error is reduced by correcting only the fixed learning amount. Is possible.

According to the arithmetic unit with a learning function of the second aspect, in addition to the effect of the first aspect, when the coefficient is corrected by a fixed learning amount, the change in the output value is, for example, the target value. Since the learning amount is set so as not to exceed the numerical value, it is possible to reduce that the coefficient is inappropriately corrected, which is useful for improving the convergence speed in learning.

According to the arithmetic unit with a learning function of claim 3, in addition to the effect of claim 1 or 2, the learning means has a relationship between one or a plurality of types of input values and corresponding coefficients. The n-dimensional (n is an integer of 2 or more) space (for example, three-dimensional) configured by
When the same tendency is shown, all the coefficients are corrected by the learning amount for the whole learning determined based on the coefficient of the subspace showing the above-mentioned tendency. , The whole coefficient is corrected all at once, which is useful for improving the convergence speed in learning.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an engine system.

FIG. 2 is a block diagram showing a fuel injection time calculation process.

FIG. 3 is a diagram showing a configuration of an intake air amount calculation unit 20b.

FIG. 4 is a flowchart showing a learning calculation process in an intake air amount calculation unit 20b.

FIG. 5: A / F output value containing noise and A after learning
It is a figure which shows the relationship with / F output value.

[Explanation of symbols]

1 engine system 20 intake air amount learning calculation unit 20a engine speed calculation unit 20b intake air amount calculation unit 20c Pmi calculation unit 20d A / F prediction unit 20e learning amount determination unit 20b ₁ intake air amount calculation CMAC 20b ₂ overall offset unit

   ─────────────────────────────────────────────────── ─── Continued front page    F term (reference) 3G084 BA13 DA04 DA20 EA04 EB02                       EB06 EB08 EB12 EB18 EB20                       EB21 FA10 FA11 FA20 FA29                       FA33 FA38                 5H004 GA18 GB12 KD63

Claims (5)

[Claims] 1. A learning system comprising: a computing unit that computes an output value based on a predetermined relationship in accordance with an input value; and a learning unit that performs learning so that the output value reaches a predetermined target value. A computing unit with a function, wherein the learning unit includes a magnitude relationship determining unit that compares the output value with respect to a predetermined input value and the target value to determine a magnitude relationship between the target value and the output value. An arithmetic unit with a learning function, which is configured to execute a partial learning process in which the coefficient representing the predetermined relationship is corrected by a constant learning amount based on the determination result. 2. The arithmetic unit with a learning function according to claim 1, wherein the constant learning amount is set so that a change in output value before and after learning falls within a predetermined numerical range. 3. The learning means divides an n-dimensional (n is an integer of 2 or more) space formed by the relationship between one or a plurality of types of input values and the corresponding coefficients into a plurality of subspaces. , When the partial learning processes for the coefficients show the same tendency, the whole learning process for correcting all the coefficients by the learning amount for the whole learning determined based on the tendency is executed. The arithmetic unit with a learning function according to claim 1 or 2, characterized in that. 4. The learning amount for overall learning is an average value of differences between a coefficient at a learning start and a coefficient at a learning end in each of the plurality of subspaces showing the tendency. An arithmetic unit with a learning function according to claim 3. 5. The learning amount for overall learning is a minimum value of differences between a coefficient at the start of learning and a coefficient at the end of learning in each of the plurality of subspaces showing the tendency. The arithmetic unit with a learning function according to claim 3.