JPH1182137A - Parameter estimating device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enhance estimating accuracy in only a bad region in the case that the bad region of estimated accuracy is learned again by judging whether a neural network belongs to which one of a plurality of regions from a plurality of input parameters, and changing a way of bonding of the neural network based on the judged result. SOLUTION: An air fuel ratio control device is constituted with a state detecting part 3 for detecting a physical quantity such as engine revolution speed, intake air pressure, and slot the opening, indicating the state of an engine E, and a control unit C, and the control unit C is constituted with a neural operation part 1 serving as a parameter estimating device and a fuel injection amount enumeration part 2. In the neural operation part 1, a physical quantity detected with a neural estimating part 11 is inputted as a parameter, and an air fuel ratio is estimated with a neural network. In a region judging part 12, an engine operation state is judged whether it belongs to which one of a plurality of regions, and the route of the neural network is varied with a route varying part 13 corresponding to the region judged.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a parameter estimating apparatus using a neural network, and more particularly to a parameter estimating apparatus capable of learning corresponding to a range of input parameters.

[0002]

2. Description of the Related Art Even when it is difficult to theoretically derive a causal relationship between an input and an output in a physical or chemical system using a neural network, an output value is converted from an input value based on a learning function. It is possible to make an estimate. Utilizing this fact, neural networks are being applied to control devices for controlling complicated control systems, in particular, control devices for controlling highly nonlinear controlled objects. For example, as a control device using such a neural network, there is an air-fuel ratio control device for an internal combustion engine of an automobile.

[0003] A method of reducing NOx, CO, and HC, which are toxic gases contained in exhaust gas emitted from automobiles, using a catalyst is employed. For example, a three-way catalyst is used as a typical catalyst. In order for such catalysts to purify these harmful gases more effectively, it is necessary to maintain the air-fuel ratio at a constant value at which the catalyst can work effectively. It is necessary to control the air-fuel ratio to keep constant regardless of the air-fuel ratio.

In such air-fuel ratio control, feedforward control for increasing or decreasing the fuel injection amount is normally performed in accordance with a change in the throttle opening and the like, and feedback control is also used. These controls are
Good results can be obtained in a steady operation range such as when idling or running at a constant speed. However, in a transient state such as acceleration or deceleration, the air-fuel ratio sensor delays or the amount of fuel actually flowing into the cylinder changes due to the operating state or the external environment. It is actually very difficult to keep the fuel ratio at a constant value only by simple feedforward control or feedback control.

In order to improve the accuracy of the air-fuel ratio control, for example, Japanese Unexamined Patent Publication No. Hei 8-74636 teaches a non-linear element such as the above-mentioned fuel adhesion by a neural network. There has been proposed a control device that controls the amount of correction so as to improve the response performance during a transition. FIG. 22 shows an example of an air-fuel ratio control device using a conventional neural network. In this air-fuel ratio control device, a plurality of physical quantities representing the state of the engine E, in this case, the engine speed (Ne), the intake air pressure (Pb), the throttle opening (THL), the fuel injection amount (Gf) ), Intake air temperature (T
a), the cooling water temperature (Tw) and the air-fuel ratio (A / F) are detected,
A plurality of parameters detected by the state detection unit 210 are input by the neuro operation unit 220, and the true air-fuel ratio (A / F) which cannot be followed by the air-fuel ratio sensor in a transient state or the like.
r) The behavior is estimated by a neural network. Then, the target air-fuel ratio is realized by the fuel injection amount calculation unit 230 by feedback control so as to reduce the deviation between the estimated value of the air-fuel ratio (A / F) and the target air-fuel ratio (A / Fref). The fuel injection amount (Gb) is calculated. As described above, by obtaining the value of the air-fuel ratio (A / F) in a transient state or the like that cannot be obtained by a normal sensor, it is possible to control the air-fuel ratio appropriately.

FIG. 23 shows a configuration of a neural network used in the above-described neuro-calculating section 220 to enable such control. This neural network is 3
It is composed of layers. For the first layer, input parameters such as engine speed (Ne), intake air pressure (Pb), throttle opening (THL), fuel injection amount (Gf), intake air temperature (Ta), cooling water temperature (Tw), and detection air are input to the first layer. Fuel ratio (A / Fk)
Is entered. This detected air-fuel ratio (A / Fk) is the latest air-fuel ratio detected by the air-fuel ratio sensor in the control cycle.
However, this air-fuel ratio (A / Fk) does not indicate the value of the true air-fuel ratio (A / Fr) due to the sensor delay. Each input parameter is multiplied by its own weight value, summed up by each neuron in the second layer, and then a threshold is added.
It is converted by the transfer function. The values converted in the second layer are further multiplied by respective weight values to obtain a third value.
Summed by neurons in layers. Then, a value obtained by adding another threshold value to the sum of the values in the neurons of the third layer is converted by another transfer function, and the estimated air-fuel ratio (A /
FNN) is calculated.

A learning process of such a neural network will be described with reference to FIG. FIG. 24 shows the estimated air-fuel ratio (A) using the engine speed (Ne) and the like as input parameters.
FIG. 6 is a schematic diagram showing a learning process of a neural network for obtaining (/ FNN). First, the same state detection unit 210 as that shown in FIG. 22 is provided in the engine E of the vehicle for learning data collection, and the vehicle is actually driven to collect learning data. The collected data is converted by the learning data generator a into learning data in which the response delay of the sensor is adjusted. The learning data includes an air-fuel ratio (A / F) for teacher data.
t) and the corresponding engine speed (Ne) for the input parameter. Air-fuel ratio for teacher data (A / F
t) can be obtained from the air-fuel ratio (A / Fk) detected by the air-fuel sensor in consideration of the detection delay. It is desirable that each input parameter also include past data in a time series, but here, past parameters are not input in order to simplify the description. The generated learning data is stored in the learning data storage unit b. Then, the learning performing unit c performs learning of the neural network using the accumulated learning data. That is, the engine speed (Ne) and the like for the input parameters are input to the neural network, and the air-fuel ratio (A /
FNN) and a teacher's air-fuel ratio (A / Ft) deviation e is detected, and the deviation e is set within a certain allowable value, for example, an average of 0.1 or less in air-fuel ratio conversion. , That is, the weight value and the transfer function are changed.

[0008]

In a control system using a neural network as described above, it is indispensable to perform highly accurate learning of the neural network in order to improve control performance. Therefore, when the above-described learning is performed, learning is performed without setting the learning data in a specific range and aligning learning data within an operation range of a control target in advance. However, it is only possible to know from the results whether or not this training data is sufficient.
It is necessary to repeat a development routine such as collecting data in an area with a poor evaluation result, creating learning data again, and performing learning again. The “area” is at least one
5 is a concept representing an operating state of an engine determined from a combination of ranges of input parameters of a unit.

By the way, as a result of evaluation based on the evaluation data, when an error of some data is large, in order to reduce the error, usually, learning data in a region where the error is large is increased and re-learned. At this time, it is necessary to select additional data and create learning data again so as to balance the entire learning data distribution, that is, considering the balance of the number of learning data in each area. However, as a result of re-learning using the selected additional data, the region where good results were obtained last time may not be able to obtain good results by this learning, or the whole region may be deteriorated There is. In such a case, the learning data must be selected again, and the data must be collected and learned again many times, so that it takes a long time until a good result is obtained. As described above, it is very difficult to improve the estimation accuracy of the neural network over a wide estimation region, and this can only be solved by trial and error at present.

Further, in learning of a neural network, it is not possible to learn and evaluate all possible operating conditions in advance. For this reason, learning in an operation range that is usually considered to be possible is performed by trial and error as described above, and high-precision control performance is guaranteed in the learning range. However, the control performance cannot be guaranteed because the estimation accuracy cannot be guaranteed outside the learning range, and a fail-safe function such as stopping the control when there is an input outside the learning range is usually required.

For example, in an air-fuel ratio control device using a neural network as described above, a neural network that has been learned and evaluated in advance using learning data in a normally conceivable operating mode range is used. For this reason, high-precision control cannot be guaranteed in a driving state outside the learning range. For example, when the driver mainly drives on mountain roads, or when the air-fuel ratio sensor or the like has deteriorated, a good air-fuel ratio Control cannot be performed.

Therefore, in the present invention, even if a region having a poor estimation accuracy is re-learned in a parameter estimating apparatus using a neural network, the estimation accuracy of only a bad region can be reduced without deteriorating other regions having a good estimation accuracy. It is an object of the present invention to improve the parameter estimating device and improve the efficiency of the development thereof.

[0013]

In order to solve the above problem, the present invention receives an input of a plurality of input parameter values related to a parameter to be estimated, and calculates the parameter value to be estimated by a neural network from the received input parameter values. In the parameter estimating device for estimating, it is assumed that the neural network has learned by changing a coupling method as previously determined for each region represented by some of the plurality of input parameters, and Area determining means for determining which of the areas belongs to the partial parameter values of the plurality of input parameter values, and a method of connecting the neural networks according to the area determined by the area determining means. A route changing means for changing the route in the same manner as when learning is provided. In the present application, when referring to a part of the parameter, the whole parameter is included.

[0014] Further, assuming that the neural network has been learned by changing the connection path for each area represented by some of the plurality of input parameters, the area determination means has determined the path change means. The connection route of the neural network can be selected according to the area. Further, assuming that the neural network has learned by changing the weight of the connection path for each of the areas represented by some of the plurality of input parameters, the path change means is set to the area determined by the area determination means. The weight of the connection path of the neural network may be changed accordingly.

[0015]

Embodiments of the present invention will be described below with reference to the drawings. (Embodiment 1) FIG. 1 is a schematic diagram showing an outline of an air-fuel ratio control device using a parameter estimating device according to the present invention. The operation of the air-fuel ratio control device will be briefly described below. First, a throttle opening sensor l, an intake air pressure sensor m, an intake air temperature sensor n, a water temperature sensor o, a crank angle sensor p, and an air-fuel ratio sensor q Is input to the control unit C. Then, based on this input, the control unit C determines the actual air-fuel ratio (A / Fr).
Is calculated to maintain the target air-fuel ratio (A / Fref), and the calculated fuel injection amount (Gf) is output to the injector I. Then, the injector I injects fuel into the engine by the calculated fuel injection amount (Gf). Such an operation makes it possible to maintain the air-fuel ratio at the target air-fuel ratio (A / Fref). Note that the output of the air-fuel ratio sensor q is not used during a cold period.

FIG. 2 shows a hardware configuration diagram of the control unit C. As shown in FIG. 2, the control unit C includes a CPU 101 for performing arithmetic processing, a ROM 10, which stores a control program, various maps, weight values of neural networks, threshold values, transfer functions, and the like.
2. A RAM 103 used for storing numerical values at the time of arithmetic processing, an A / D converter 104 for converting a sensor input into a digital value,
D / A converter 1 for converting digital output to analog output
05.

FIG. 3 shows a functional block diagram of an air-fuel ratio control device including a control unit C realized by the above hardware. The air-fuel ratio control device includes a state detection unit 3 for detecting a physical quantity representing a state of an engine E to be controlled and a control unit C. The control unit C includes a neuro-calculation unit 1 and a fuel injection amount calculation unit 2. You. Note that the parameter estimating device according to the present invention is configured by the neuro operation unit 1.

The state detecting section 3 is a section for detecting a plurality of detectable physical quantities representing the state of the engine E. Here, the engine speed (Ne), intake air pressure (Pb), throttle opening (THL), fuel injection amount (Gf), intake air temperature (T
a), the cooling water temperature (Tw), and the air-fuel ratio (A / Fk) are detected. The state detection unit 3 performs a predetermined calculation on the parameters using the various sensors for detecting these physical quantities and the values of the sensors. And a RAM that stores the fuel injection amount and the like in the past control cycle. The fuel injection amount (Gf) is the fuel injection amount one control cycle before stored in the RAM. Further, the air-fuel ratio detected by the air-fuel ratio sensor does not represent the true air-fuel ratio (A / Fr) due to a delay in the response of the sensor. Here, the air-fuel ratio detected by this sensor is represented as a detected air-fuel ratio (A / Fk). Further, the plurality of physical quantities to be detected are not limited to these, and it goes without saying that more physical quantities may be detected.

The neuro operation unit 1 in the control unit C comprises a neuro estimation unit 11, an area determination unit 12, and a route change unit 13. The neuro estimation unit 11 inputs the physical quantity detected by the state detection unit 3 as a parameter,
The air-fuel ratio (A / FNN) is estimated using a neural network. FIG. 4 shows a conceptual diagram of a neural network used in the neuro estimation unit 11. This neural network is composed of four layers. To the first layer, as input parameters, the engine speed (Ne) and the intake air pressure (P
b), throttle opening (THL), fuel injection amount (G
f), the intake air temperature (Ta), the cooling water temperature (Tw), and the detected air-fuel ratio (A / Fk) are input. The air-fuel ratio (A / F
k) is 0 because the air-fuel ratio sensor does not work in the cold state.
Becomes It is also desirable to add data in the past control cycle to each input parameter. Each input parameter is multiplied by a weight value indicated by each coupling coefficient and summed up by each neuron in the second layer. This sum result is added with a threshold value and converted by a transfer function. The converted values are further multiplied by the respective weight values, summed up in the third layer, and output to the fourth layer after the same operation as in the second layer.
Then, the values are summed up by the neurons in the fourth layer and another threshold value is added. This value is converted by another transfer function to calculate the value of the estimated air-fuel ratio (A / FNN). The feature of this neural network is that the air-fuel ratio is estimated by a different path configuration depending on the input parameter. This is related to the neural network learning method described later in detail, and will be described together therewith.

The region determining unit 12 determines from a part of the parameters input to the neuro estimating unit 11 to which of the regions determined by the ranges of the predetermined parameters the engine operating state belongs. . Here, the engine speed (N
e), the intake air pressure (Pb), the intake air temperature (Ta), and the cooling water temperature (Tw) are used, and it is determined which of the areas divided in advance in the four-dimensional space constituted by these four parameters. I do. For selection of these parameters, input parameters having a large non-linear tendency in relation to the air-fuel ratio are selected.

Here, since the "region" in the four-dimensional space is used, this cannot be illustrated, but it is assumed that the region is divided by two parameters of the engine speed (Ne) and the intake air pressure (Pb). For example, the regions can be divided as shown in FIG. In FIG. 5, the range of the engine speed (Ne) and the range of the intake air pressure (Pb) are each divided into three, thereby dividing the region into nine. In other words, the region in the four-dimensional space means each part surrounded by the intersection of the range divided for each parameter in the four-dimensional space. The dimension of the space is determined by the number of parameters, and when there is one parameter, the range is a straight line. Here, this range is also referred to as a region.

The route changing unit 13 changes the route of the neural network according to a route predetermined according to the area determined by the area determining unit 12. As a route of the neural network corresponding to the region, for example, for each region divided in the four-dimensional space as described above, only one route from each neuron of the first layer to the second layer is directed to a specific neuron. Or just a few. Further, the weight of each route may be changed for each area.

FIG. 6 is a diagram schematically showing a part of the route changing unit 13. FIG. 6A shows the engine speed (Ne).
Is shown, a route is determined by switching the switch X1 according to the region to which the region determination result is output. Similarly, the neuron to which the other parameters (Pb, THL,...) Are input is configured to select a path according to the input area. The actual switching of the path is performed by processing of a program stored in the ROM.

As a route selection, for example, for a certain area, all switches select the top route in the figure,
In another area, it is conceivable that all the switches select the route to the same neuron in the second layer, for example, all the switches select the second switch from the top. FIG.
Shows the path obtained when all switches select the top path. As can be understood from the figure, with such a configuration, the output is concentrated on a specific neuron of the second layer for each region, so that the coupling coefficient of the path from the second layer to the third layer can be changed. In addition, it is possible to avoid affecting other neurons when re-learning each area.

The selection of the route is not limited to this, and a route toward a group of neurons in the second layer can be selected for a certain region, and the group of neurons corresponding to each region share a common part with each other. It may be included, or may be completely different. FIG. 6B shows a configuration in which the weight of each path is changed for a neuron that receives the engine speed (Ne) as an input. The weighting coefficients (S1 to S8) are changed in accordance with the area. Is shown to multiply the respective paths by the weighting coefficient of.

For example, in the route selection as shown in FIG. 7, only one neuron in the second layer is used for a specific area, but the specific neuron is weighted as the main and some other neurons are weighted. By performing weighting, other neurons can be used. Also, by changing the weighting, the number of types of routes can be increased, so that a flexible route corresponding to the type of region can be configured.

FIG. 8 shows an example of weighting for each path from the neurons of the first layer to the neurons of the second layer which receives the engine speed (Ne) as an input. In the figure, it is assumed that the same weight is sequentially applied to neurons to which other parameters are input from the upper path. In this way, the path from all the neurons in the first layer to the top neuron in the second layer is multiplied by “0.6” as a weight value, so that the top neuron is a main neuron. By multiplying the path to the other neuron in the second layer by a light weight value, the other neuron can be made an auxiliary neuron. Then, various paths can be constructed by changing the main neuron for each area determined by the parameters and changing the weight of the auxiliary neuron.

The fuel injection amount calculation unit 2 in the control unit C calculates the air-fuel ratio (A /
The fuel injection amount (Gf) is calculated from the deviation D between the FNN) and the target air-fuel ratio (A / Fref (= 14.7)). Specifically, the deviation D between the estimated air-fuel ratio (A / FNN) and the target air-fuel ratio
From the P, P, and PID control systems. For example, it is calculated by the following equation. Here, Kp, Kd, and Ki are PID gains, which are experimentally obtained.

Gf = Kp ・ D + Kd ・ ΔD + KiΣΣD Here, the fuel injection amount (G) is directly obtained from the deviation D between the estimated air-fuel ratio (A / FNN) and the target air-fuel ratio (A / Fref).
Although f) was calculated, this may be performed as follows. That is, the respective physical quantities detected by the state detection unit 3 are input to the fuel injection amount calculation unit 2, and a basic fuel injection amount (Gfb) is obtained from these values by feedforward control using a map obtained by experiment. On the other hand, the air-fuel ratio (A / FNN) estimated by the neuro estimation unit 11 and the target air-fuel ratio (A / FNN)
Fref), the fuel injection amount correction amount (ΔG
f) is calculated. Then, by adding the basic fuel injection amount (Gfb) and the correction amount (ΔGf), the fuel injection amount (G
f) is calculated. This can prevent abnormal fuel injection from being performed in the event that a large noise is added to the neural network and the estimated value of the neural network deviates significantly.

Next, a description will be given of a learning process of the neural network 11 having the above configuration. FIG. 9 shows the engine speed (Ne), the intake air pressure (Pb), the throttle opening (THL), and the fuel injection amount (G
f), the intake air temperature (Ta), the cooling water temperature (Tw), and the detected air-fuel ratio (A / Fk) as input items, and the air-fuel ratio (A / FNN)
FIG. 7 is a schematic diagram showing a learning process of a neural network for estimating.

First, the same state detector 3 as that shown in FIG. 4 is provided in the engine E of the automobile for collecting learning data. Then, the user actually drives the car and collects learning data.
Since the air-fuel ratio sensor of the state detection unit does not operate at a low temperature, the air-fuel ratio sensor is warmed first so that the air-fuel ratio can be detected even when the engine itself is in a low temperature state.

The collected data is subjected to phase delay compensation for the detection delay of the sensor by the learning data generator a,
Engine speed for learning (Ne), intake air pressure (Pb)
And the like are stored as input data for learning and the air-fuel ratio (A / Ft) corresponding thereto as teacher data in the learning data storage unit b in association with these. However, since the detected air-fuel ratio (A / Fk) at the time of cold is not inputted in the actual operation, it is deleted from the input data.

Using the learning data, the learning execution unit c uses the input data as input parameters of the neural network, and estimates the air-fuel ratio (A / FNN) as the output and the teacher data (A / FNN). The deviation e of Ft) is detected, and the coupling coefficient of the neural network, that is, the weight value integrated in each path from the neuron to the neuron, is changed so as to reduce the deviation e.

At this time, the engine speed (Ne), the intake air pressure (Pb), the intake air temperature (Ta),
Each data of the cooling water temperature (Tw) is input to the same region determination unit 12 as that of the air-fuel ratio control device shown in FIG. 4, and the region determination is performed on the input data. The route changing unit 13 that is the same as the air-fuel ratio control device shown changes the route of the neural network according to the input data. That is, learning is performed while the configuration of the neural network is changed according to the area corresponding to each input data.

In order to change the coupling coefficient of the neural network by the deviation e, each input parameter is converted by a transfer function using a back propagation method.
This is performed by converting the value of the weight value multiplied by this and the value of the threshold value in each processing unit by a predetermined method according to the deviation e. Here, the tangent sigmoid function (f (x) = tanh
(X)) is used. Various types of neural networks can be adopted as long as an air-fuel ratio value obtained from each parameter can be obtained.

When the above learning is completed, learning data is generated from data not used for learning or new data obtained by moving the automobile, and input parameters obtained from the learning data are generated. The air-fuel ratio (A / FNN) input to the trained neural network and estimated for each region is the air-fuel ratio (A / Ft) of the teacher data.
Then, it is determined whether the degree of estimation accuracy and generalization, that is, the degree of good estimation with respect to input parameters other than learning data is good. For example, the estimated air-fuel ratio (A /
It is determined that the estimation accuracy or generalization is poor in a region where 20% or more of the data has a deviation between the FNN) and the air-fuel ratio (A / Ft) of the teacher data exceeding 0.5.

When there is an area with poor estimation accuracy or the like, a path of the neural network is set by the path changing unit 13 in correspondence with the area.
Re-learning is performed using the learning data of the area. By performing such learning, the path for estimating the air-fuel ratio using the input parameters belonging to other regions has almost no effect, and the input parameters belonging to the regions having poor estimation accuracy and generalization properties are used. Only the route for estimating the air-fuel ratio can be improved.

As a result of such learning, the above-described neural network receives the air-fuel ratio (A / A
F) can be predicted and estimated. Note that a similar method is employed in manufacturing the neural network described in the following embodiments during manufacturing. Next, the operation of the air-fuel ratio control device having the above configuration will be described. FIG. 10 is a flowchart illustrating an operation in one control cycle of the air-fuel ratio control device.

First, the state detector 3 determines the physical state of the engine E as physical quantities such as engine speed (Ne), intake air pressure (Pb), throttle opening (THL), and fuel injection amount (G
f), intake air temperature (Ta), cooling water temperature (Tw), and air-fuel ratio (A / F) are detected (S101). Next, among the physical quantities detected by the area determination unit 12, the engine speed (N
e), the region of the engine operating state is determined from the values of the intake air pressure (Pb), the intake air temperature (Ta), and the cooling water temperature (Tw) (S102). Based on this region determination result, the route changing unit 13 changes the configuration of the route of the neural network of the neuro estimation unit 11 (S103).

Then, the neural estimating unit 11 estimates the air-fuel ratio (A / FNN) using the plurality of detected parameters as input parameters by the neural network having the changed route configuration (S104). Finally, the fuel injection amount calculation unit 2 calculates the fuel injection amount (Gf) using the estimated air-fuel ratio (A / FNN) (S105). The calculated fuel injection amount (Gf) is output to the injector I of the engine, and this amount of fuel is injected into the engine E. With such an operation, appropriate air-fuel ratio control can be performed even during a transition or the like.

In the above description, the regions are clearly divided according to the range of the parameters, and the route configuration of the neural network is changed for each of the clearly divided regions. Is not clear, but the boundary is ambiguous, and the area determining unit 12 determines to which area the input parameter belongs by the fuzzy inference rule, and changes the route of the neural network accordingly. Is also conceivable. Hereinafter, a method of determining an area and changing a route using such a fuzzy inference rule will be described.

For the sake of simplicity, here, the parameters for determining the region are two, the engine speed (Ne) and the intake air pressure (Pb), and as shown in FIG. 11, the engine speed (Ne) and the intake air pressure ( Pb) is divided into three small (S), medium (M), and large (B), respectively, and roughly divided into nine regions in which the boundaries between D1 to D9 are ambiguous.

The second layer of the neural network of the neuro estimating unit 11 has a configuration in which four neurons are assigned to each of the above regions, and the route changing unit 13 only assigns routes to the four neurons assigned to each of the regions. Shall be selected. For example, the route changing unit 13 sets the area D
With respect to 5, only the routes to the four neurons assigned to D5 are selected as shown in FIG.

The respective areas D1 to D9 can be expressed as follows according to the fuzzy control rules. if Ne = S and Pb = S then area F = D1 if Ne = M and Pb = S then area F = D2 if Ne = B and Pb = S then area F = D3 if Ne = S and Pb = M then area F = D4 ... (α) if Ne = M and Pb = M then area F = D5 ... (β) if Ne = B and Pb = M then area F = D6 if Ne = S and Pb = B then area F = D7 if Ne = M and Pb = B then Region F = D8 if Ne = B and Pb = B then Region F = D9 Membership function of each variable of engine speed (Ne) and intake air pressure (Pb) Are triangular as shown in FIGS. 13A and 13B, respectively. Various membership functions such as a bell shape and a trapezoidal shape can be adopted in addition to the triangular shape.

Now, for example, the value of the engine speed (Ne) is 1000 rpm, and the value of the intake air pressure (Pb) is -300 mmH.
g. From these values, the area determination unit 12 makes the following fuzzy inference. Based on the membership functions shown in FIGS. 13A and 13B, the fitness levels of the engine speed (Ne) and the intake air pressure (Pb) are as follows.

Ne: S = 0.8, M = 0.2, B = φ Pb: S = φ, M = 1, B = φ From these, a combination considered to be compatible with the above fuzzy control rule is obtained. (Α) and (β), and the degree of conformity of (α), that is, the area F = D4 by the MIN operation is 0.8 from S = 0.8 (Ne) and M = 1 (Pb). Yes, (β), that is, the degree of conformity for the region F = D5 is 0.2 from M = 0.2 (Ne) and M = 1 (Pb).

This is illustrated as a hatched portion in FIG. FIG. 14 shows the membership function of the fuzzy variable in the area F. And, by the MAX operation,
The center of gravity G of the hatched portion is determined, and this point is determined by setting the engine speed (Ne) to 1000 rpm and the intake air pressure (P
The value of b) is an area represented by -300 mmHg. From the center of gravity G obtained in this way, the route changing unit 13 selects a route as follows. That is, the position of the center of gravity G is D
Since it is located at a position slightly shifted from the area of 4 to the area of D5,
The composition ratio of the neurons corresponding to D4 and D5 in the second layer is determined according to the position of the center of gravity. More specifically, as shown in FIG. 15, when the corresponding neurons in the second layer of the neural network are arranged in correspondence with each region of the membership function of the region F, they correspond to one region centered on the position of the center of gravity. Select neurons that fall within the width so that only routes to these neurons are selected.

Here, the route changing unit 13 changes the ratio of the selected route according to the result of the fuzzy inference. This is done by changing the weight for each route according to the center of gravity. Is also good. For example, in a configuration in which only a path to one neuron in the second layer is selected for each area, as a result of fuzzy inference, about 80% of the area to be directed to neuron A is directed to neuron B. Is determined to be 20%. In this case, a weighting coefficient of 0.8 is added to each path toward the neuron A, and a weighting factor of 0.2 is added to each path toward the neuron B. In this way, the result of the fuzzy inference can be reflected in the change of the route configuration by weighting.

(Embodiment 2) Next, a second embodiment will be described. FIG. 16 shows a functional block diagram of the air-fuel ratio control device according to the present embodiment. This air-fuel ratio control device is different from the air-fuel ratio control device according to Embodiment 1 in that it has a configuration that allows online learning. For this reason, the configuration of the neuro calculation unit 1 is different from that of the air-fuel ratio control device according to the first embodiment.

That is, the neuro operation unit 1 includes a reference value acquisition unit 14, an estimation data recording unit 15, an estimation accuracy judgment unit 16, a generalization judgment unit 17a, and a relearning unit 17 as new components.
c, learning data generation unit 18a, learning data storage unit 1
8b, a learning execution unit 18c, a neural network updating unit 20, an initial coupling coefficient storage unit 19a, and an initial state return unit 1
9b is further added. Other configurations are the same as those of the air-fuel ratio control device according to the first embodiment.

The reference value acquisition unit 14 calculates the air-fuel ratio (A / FNN) estimated by the neuro estimation unit 11 in one control cycle.
Is a part for obtaining the actual air-fuel ratio (A / Fr) as a reference value with respect to the value of the air-fuel ratio (A / Fk) detected by the air-fuel ratio sensor of the state detection unit 3. Of the air-fuel ratio (A / Fr), that is, the reference value.

The estimated data recording unit 15 records the input parameters detected by the state detecting unit 3, the area determined by the area determining unit 12 based on the input parameters, and the reference value corresponding to the input parameters. I will do it. The estimation accuracy determination unit 16 calculates the estimated air-fuel ratio (A
/ FNN) and the corresponding reference value are compared to determine whether or not the estimation accuracy in each region is good. Specifically, if the average of the absolute value of the difference between the estimated air-fuel ratio (A / FNN) and the reference value in each region within a predetermined accuracy determination period is equal to or smaller than a predetermined set value, the region Is determined to be good, and if it is equal to or larger than the set value, it is determined that the estimation accuracy of the area is poor. Here, the accuracy judgment period is one hour,
The above set value is set to 0.8.

However, if the number of estimated air-fuel ratios (A / FNN) does not exceed a predetermined amount, for example, 10, the judgment of the estimation accuracy is not performed. The reason for this is to eliminate such a case because it is not appropriate to perform new learning when the estimation accuracy deterioration is caused by a temporary cause such as sensor noise. Note that the accuracy determination period may be defined by a time interval or by the number of times of estimation by the neuro estimation unit 11, and may be a period in which the number of data for each region reaches a predetermined number, for example, 50. Good. Also,
Here, it is determined whether the average of the difference between the estimated air-fuel ratio (A / FNN) and the reference value is equal to or smaller than a predetermined set value, but it is determined whether the value of the variance of the difference is equal to or smaller than the predetermined set value. Is also good. The same applies to each of the following periods.

Further, without judging whether the average or variance of the difference between the estimated air-fuel ratio (A / FNN) and the reference value is equal to or smaller than a predetermined set value, the estimated air-fuel ratio ( A
/ FNN) and the absolute value of the difference between the reference value and a predetermined value (for example,
0.8) or more are counted, and this is counted for a predetermined number (for example, one hour) during a predetermined time (for example, one hour).
20) is determined as a region with poor estimation accuracy,
A region that does not exceed the region may be determined as a region having a good estimation accuracy.

The learning data generation unit 18a reads, from the estimation data recording unit 15, the input parameters and the reference values corresponding to the region for which the estimation accuracy determination unit 16 has determined that the estimation accuracy is low, Using these, the input data for learning of the neural network of the neuro estimation unit 11 and the data for teachers are generated. The learning data storage unit 18b records and stores the generated learning data for each area.

The learning execution unit 18c uses the learning data stored in the learning data storage unit 18b during a predetermined storage period to perform neural learning on the neural network having a connection method corresponding to each region. To obtain a new coupling coefficient. Specifically, the learning performing unit 18c calls a program for realizing the neural network of the neuro estimating unit 11, and uses this to change the route of the neural network by the route changing unit 13 for each region. By changing the configuration of the neural network on the path changed using the learning data corresponding to the above by the back propagation method, a new coupling coefficient is obtained. The learning unit 18 is formed by the learning data generation unit 18a, the learning data storage unit 18b, and the learning execution unit 18c.

When a new coupling coefficient is obtained by the learning performing unit 18c, the neural network updating unit 20 updates the neural network of the neuro estimation unit 11 with the new coupling coefficient. The generalization determining unit 17a compares the estimated air-fuel ratio value (A / FNN) estimated using the updated neural network with the reference value, and generalizes the updated neural network for each region. The quality of the sex is judged. The generalization is mainly based on the learning execution unit 1
If the learning performed by 8c is too much learning or the learning result does not converge to a minimum value (local minimum)
The generalization determining unit 17a determines whether the generalization is good or not in order to deal with such a case.

More specifically, a predetermined number of estimated air-fuel ratio values (A / FNN) in the region where learning has been performed and the reference value acquisition unit 1
If the average value of the difference from the corresponding reference value acquired by No. 4 is equal to or less than a predetermined set value, it is determined that generalization is good, and if it is equal to or more than the set value, it is determined that generalization is bad. Here, the predetermined number is set to 100 after the neural network is updated by the neural network updating unit 20.
The set value is 0.3.

The learning invalidating unit 17b records the coupling coefficient before updating, and when the generalization determining unit 17a determines that the generalization is poor in a specific area, the learning invalidating unit 17b updates the area. Returns the coupling coefficient to its previous value. The initial coupling coefficient storage unit 19a stores the initial coupling coefficient of the neural network set at the manufacturing stage, and stores the initial state return unit 1
9b returns the neural network of the neuro estimating unit 11 to the initial state using the coupling coefficients stored in the initial coupling coefficient storage unit 19a. This initial state return unit 19b
The system is activated by a user operating a predetermined switch. It should be noted that the activation of the initial state return unit 19b is designed to emit a predetermined signal when the air-fuel ratio sensor whose performance greatly changes due to deterioration is replaced, or the like.
It is also conceivable to activate by this signal.

The operation of the air-fuel ratio control device having the above configuration will be described below. The control operation of the air-fuel ratio is the same as that of the air-fuel ratio control device according to the first embodiment, and thus the description thereof is omitted here, and the online learning operation will be described. In the following embodiments, the control operation of controlling the air-fuel ratio is the same as that according to the first embodiment.

FIG. 17 is a flowchart showing the operation in one learning cycle of the learning operation. First, the reference value acquisition unit 14 sets the actual air-fuel ratio (A / Fr) in one control cycle.
Is obtained as a reference value, and is output to the estimation accuracy determination unit 16. Then, the estimation accuracy determination unit 16 determines the air-fuel ratio (A / A) estimated by the neuro estimation unit 11 in this control cycle.
FNN) and a deviation between the output reference value (A / Fr) are calculated, and the absolute value of the deviation is recorded and added in association with the area determined by the area determination unit 12. At the same time, the estimated data recording unit 15 records the input parameters of the neuro estimation unit 11 in the control cycle, the reference value, and the area in association with each other.

After the above operation is repeated for one hour, the estimation accuracy judgment section 16 calculates an average from the sum of absolute values of the deviations recorded for each area, compares the average with a set value, and compares this with a set value. It is determined whether or not there is an area that exceeds (S201). In the estimation accuracy determination unit 16,
Here, returning to S201, the following processing is performed in parallel.

Next, the learning data generation unit 18a reads the input parameters and the reference values recorded in the estimation data recording unit 15 for the area in which the estimation accuracy determination unit 16 determines that the estimation accuracy is low. The learning input data and the teacher data are generated and stored in the learning data storage unit 18b. Then, the learning performing unit 18c calls a program for realizing the neural network of the neuro estimating unit 11, and uses this to change the route changing unit 13 according to the region.
, The path of the neural network is changed, and a new coupling coefficient is obtained for the neural network for each region by the back propagation method (S202).

When learning for each region extracted as having poor estimation accuracy is completed and a final coupling coefficient is obtained, the neural network is updated in a steady state when the air-fuel ratio control is not a problem, such as when the engine is idling. The unit 20 updates the neural network of the neuro estimation unit 11 with the new coupling coefficient (S203). When the neural network is updated, the generalization determining unit 17
a is the air-fuel ratio (A / FNN) estimated by the neuro estimation unit 11 for the learned region and the air-fuel ratio (A / F
Add the absolute value of the difference from r). Then, when 100 values are added for one region, the generalization determining unit 1
7a calculates the average value of these differences, and determines whether or not this value is equal to or greater than a predetermined value (0.3) (S204).

Here, if the average value is less than the predetermined value, the generalization property is good and nothing is performed as it is. On the other hand, if the average value is equal to or larger than the predetermined value, the generalization property is poor, so the neural network updating unit 20 returns the neural network of the neuro estimation unit 11 to the state before the update (S205). When the above operation is completed, here, the learning data storage unit 1
8b is cleared (S206). Of course, the data in the learning data storage unit 18b may be used together with new learning data in the next learning cycle without clearing, and the latest one-hour data may always be left, and the remaining old data may be used. It is also conceivable to clear the data.

By the above operation, even if the accuracy of the estimation of the air-fuel ratio is deteriorated due to the driving in the operating state in the region where the learning was not performed at the time of manufacturing or the deterioration of the air-fuel ratio sensor, etc., the estimation accuracy is improved by the re-learning. Good value,
Further, since the learning is performed on the path of the neural network corresponding to the region where the estimation accuracy has deteriorated, the estimation of the region where the estimation accuracy has been good is not significantly affected.

Next, the operation of the initial coupling coefficient storage section 19a and the initial state return section 19b will be described. First,
The user operates a predetermined switch to send a signal to the initial state return section 19b. Then, the initial state return unit 19b
Restores the neural network to the original state by the initial neural network coupling coefficient set at the manufacturing stage recorded in the initial coupling coefficient storage unit 19a.

For example, if the estimated value of the air-fuel ratio is deteriorated due to the deterioration of the air-fuel ratio sensor and online learning is performed and the coupling coefficient of the neural network of the neuro estimation unit 11 is updated, the air-fuel ratio sensor is replaced with a new one. A good estimate cannot be obtained with the updated neural network if it is replaced. In such a case, it is appropriate to return to the original coupling coefficient. In addition, if the neural network configuration becomes uncontrollable as a result of online learning, or if the owner of the car changes and the driving environment changes, it is better to return to the original coupling coefficient when shipping There is. Assuming such a case, an initial coupling coefficient storage unit 19a and an initial state return unit 19b are provided.

(Embodiment 3) In the second embodiment, the learning result is invalidated and returned to the pre-update coupling coefficient for the region determined to have no generalization by the generalization determining section 17a. It is also conceivable to perform relearning using the result as it is. In the present embodiment, an air-fuel ratio control device configured to re-learn an updated neural network will be described. FIG. 18 shows a functional block diagram of the air-fuel ratio control device according to FIG.

The air-fuel ratio control device shown in FIG. 18 is different from the air-fuel ratio control device shown in FIG. 16 in that the learning invalidation unit 17b is replaced by a relearning unit 17c, and the other configuration is the same. The re-learning unit 17c causes the learning unit 18 to perform re-learning on the region determined to be poor in generalization by the generalization determining unit 17a. That is, the learning data generation unit 18a reads input parameters and reference values corresponding to the area determined to have poor generalization from the estimation data recording unit 15, generates learning data, and generates the learning data storage unit 18b. The learning execution unit 18c uses the stored learning data to re-learn the neural network changed to a connection corresponding to the region.

Here, in this re-learning, considering that there is a high possibility that over-learning or a minimum value is obtained,
The condition is changed from that at the time of the first learning, and for example, re-learning is performed by thinning out the number of learning data. However, it goes without saying that various data can be adopted in re-learning, such as adding new data. The operation of the air-fuel ratio control device having such a configuration in one cycle during online learning will be described below. FIG. 19 is a flowchart showing this operation. This flowchart is almost the same as the flowchart shown in FIG. 17, but differs in that re-learning is performed without invalidating the learning in S205 in the flowchart of FIG.

The operations from the determination of the estimation accuracy (S201) to the determination of generalization (S204) are the same as those of the air-fuel ratio control device according to the second embodiment, and will not be described. As a result of the determination by the generalization determining unit 17a in S204, if the generalization of the update area is good, the data in the learning data storage unit 18b is cleared as in the case of the second embodiment (S20).
6). On the other hand, as a result of the determination by the generalization determining unit 17a, if the generalization of the update region is poor, the relearning unit 17c causes the learning unit 18 to perform relearning on the update region with poor generalization, and A new coupling coefficient corresponding to the update area is acquired (S305).

When a new coupling coefficient is obtained, S
203, the neural network updating unit 20
Further updates the neural network with this coupling coefficient. Hereinafter, the operations of S204, S305, and S203 are repeated until the generalization of the update area is improved. By such an operation, the generalization property can be further improved by re-learning by using the neural network whose estimation accuracy has been improved by learning as it is, and the configuration of the neural network can be efficiently changed to a desired one. .

(Embodiment 4) In the air-fuel ratio control devices according to Embodiments 2 and 3, when a new coupling coefficient of the neural network is obtained by the learning execution unit 18c, the neural network updating unit 20 immediately starts. Updated the neural network of the neuro estimation unit 11,
In the air-fuel ratio control device according to the present embodiment, it is first determined whether or not a new coupling coefficient obtained before updating gives a neural network with good generalization.

FIG. 20 shows a functional block diagram of the air-fuel ratio control device according to the present embodiment. This air-fuel ratio control device differs from the air-fuel ratio control device according to Embodiment 3 in that a provisional generalization determining unit 17a * is provided instead of the generalization determining unit 17a, and the function of the neural network updating unit 20a Has been changed. The temporary generalization determining unit 17a * compares the estimated air-fuel ratio (A / FNN) estimated using the temporary neural network in which the new coupling coefficient obtained by the learning performing unit 18c is temporarily replaced with a reference value. Then, based on a predetermined criterion, the quality of the generalization of the temporary neural network is determined for each region. Specifically, the temporary generalization determining unit 17a * obtains, as a temporary neural network, a program obtained by replacing the program for realizing the neural network of the neuro estimation unit 11 with the coupling coefficient obtained by the learning performing unit 18c. Average of the absolute value of the difference between the estimated air-fuel ratio value (A / FNN) estimated by the provisional neural network and the corresponding reference value acquired by the reference value acquiring unit 14 for each region during the predetermined generalization determination period If the value is equal to or less than a predetermined set value, it is determined that generalization is good. If the value is equal to or more than the set value, it is determined that generalization is poor. Here, Embodiment 3
Similarly to the generalization determining unit 17a according to the above, the predetermined number is set to 100 after the neural network is updated by the neural network updating unit 20, and
The set value is 0.3.

The relearning unit 17c is substantially the same as that according to the third embodiment, except that the tentative generalization determining unit 17a * re-learns an area determined to have poor generalization. Different. That is, the re-learning unit 17c operates the learning data generation unit 18a to input the input parameters and the reference values corresponding to the region determined to be poor in generalization by the provisional generalization determination unit 17a * to the estimation data recording unit. 15, the learning data is generated and stored in the learning data storage unit 18b, and then the learning data is used to re-learn the neural network changed to a connection corresponding to the region. Note that changing learning data to some extent in re-learning is the same as in the third embodiment.

The neural network updating unit 20a uses the coupling coefficient obtained by the learning execution unit 18c to execute the neural estimation unit 11 only when the temporary generalization determining unit 17a * determines that generalization is good. Update the neural network of. The operation of the air-fuel ratio control device having such a configuration in one cycle of online learning will be described below. FIG. 21 is a flowchart showing this operation. The operations shown in this flowchart are the same as those in the flowchart shown in FIG. 17 up to S201 and S202, and therefore description thereof is omitted.

When a new coupling coefficient is obtained by learning in S202, the temporary generalization determining unit 17a * obtains a temporary neural network from the learning coefficient and obtains a new neural network for a predetermined period, that is, 100 new data. It is determined whether or not the average of the absolute values of the estimated air-fuel ratio (A / FNN) and the reference value (A / Fr) during the given period is equal to or less than a set value (S403). Here, for the region determined to have good generalization, the neural network updating unit 20a updates the neural network of the neuro estimation unit 11 using the coupling coefficient obtained by the learning execution unit 18c (S404).

On the other hand, as a result of the determination by the provisional generalization determining section 17a *, if the generalization of the temporary neural network is poor, the re-learning section 17c operates the learning section 18 for the area with poor generalization and re-learns. To obtain a new coupling coefficient corresponding to the area of the temporary neural network (S405). When a new coupling coefficient is further obtained, the process returns to S403, and the temporary generalization determining unit 17a * obtains a new temporary neural network from the learning coefficient to determine whether generalization is good or not again. If the performance is poor, the operations of S405 and S403 are repeated until the generalization is improved. Then, when it is determined that the generalization property is good, the neural network updating unit 20a updates the neural network of the neuro estimation unit 11 using the coupling coefficient obtained by the learning execution unit 18c (S404). When the above operation is completed, the data in the learning data storage unit 18b is cleared as in the case of the second embodiment, and the processing is terminated (S
206).

By such an operation, since the neural network is not updated using the coupling coefficient that gives a poor generalization result, the accuracy of the estimation of the air-fuel ratio suddenly decreases in an operation state other than the learning data due to erroneous learning. It is possible to surely prevent the deterioration of the control accuracy and the control accuracy, thereby ensuring the safety in online learning.

In each of the above embodiments, the control target is the air-fuel ratio of the engine E. However, the control target has a strong nonlinearity and can detect a parameter related to a state quantity to be estimated as a neuro output. Then, the above-described neuro-calculating unit 1 can be used, and further, besides the control device, it can be used for any device that estimates an output value that has a causal relationship but is difficult to analyze from some input values. It is possible.

[0082]

As described above, the present invention has the following effects. First, the parameter estimating device according to the present invention is a parameter estimating device that receives input of a plurality of input parameter values related to an estimation target parameter and estimates an estimation target parameter value by a neural network from the received input parameter values. The neural network has learned by changing the coupling method as predetermined for each region represented by a part of the plurality of input parameters, and the plurality of regions received by the region determination means. When it is determined which of the areas belongs to the partial parameter values of the input parameter values, and the route changing means learns how to connect the neural networks according to the area determined by the area determining means. Make the same changes as

By such an operation, the parameter value to be estimated is estimated in a combining manner according to the area represented by the input parameter. Then, the estimation target parameter value is estimated by the neural network having a different connection configuration for each region determined by the input parameters as described above, so that learning can be performed for each region when learning the neural network.

That is, by dividing the learning area (possible operating range of the controlled object) by the input parameters and changing the connection of the coupling coefficient from the input layer of the neuro to the intermediate layer according to the area, One neuro configuration can have a small neuro configuration corresponding to each region. With this, when the estimation accuracy or the like is low for a specific area, learning needs to be performed only for a neural network having a connected configuration corresponding to the area, so that the influence on the connected configuration of the neural networks corresponding to other areas is not affected. With less, learning can be performed for each area.

Therefore, since the coupling coefficients and the like of the regions that do not need to be learned with high accuracy are almost preserved, the problem that the estimation accuracy of the regions that were good in the conventional additional learning cannot be guaranteed can be solved. It becomes possible to increase learning efficiency by reducing trial and error at the time. Furthermore, since it is sufficient to re-learn using only the data in the inaccurate area, it is not necessary to re-learn using all the data as in the conventional case, and the number of data is reduced to greatly reduce the learning time. Therefore, the learning can be performed at high speed.

Further, it is assumed that the neural network has learned by changing the connection route for each of the regions represented by some of the plurality of input parameters, and the route changing means determines the area determining means. If the route of the connection of the neural network is selected in accordance with the region that has been set, it is possible to adopt a configuration in which a specific neuron passes through each region and so on. Of the neural network can be realized.

On the other hand, it is assumed that the neural network has learned by changing the weight of the connection path for each of the areas represented by some of the plurality of input parameters, and the path change means has the area determination means If the weight of the connection path of the neural network is changed in accordance with the determined area, a large weight is applied to the connection path passing through a specific neuron for each area, and a small weight is applied to other neurons. It is possible to realize a configuration in which a very large number of types of neural networks can be connected for each area.

[Brief description of the drawings]

FIG. 1 is a schematic diagram showing an outline of an air-fuel ratio control device according to an embodiment.

FIG. 2 is a diagram showing a hardware configuration of a control unit.

FIG. 3 is a functional block diagram of the air-fuel ratio control device according to the first embodiment.

FIG. 4 is a conceptual diagram of a neural network according to the embodiment.

FIG. 5 is a diagram showing an example of how to divide an area.

6A is a diagram schematically illustrating a part of a route changing unit that changes a route, and FIG. 6B is a diagram schematically illustrating a part of the route changing unit that changes the weight of a route;

FIG. 7 is a conceptual diagram showing an example of a neural network whose route has been changed.

FIG. 8 is a conceptual diagram showing an example of a neural network to which path weights are given.

FIG. 9 is a diagram schematically showing a learning process of the neural network according to the first embodiment.

FIG. 10 is a flowchart showing a control operation of the air-fuel ratio control device according to the first embodiment.

FIG. 11 is a diagram showing an example of area division for applying a fuzzy inference rule.

FIG. 12 is a conceptual diagram showing an example of a configuration of a neural network for applying a fuzzy inference rule.

13A is a diagram showing a membership function of an engine speed, and FIG. 13B is a diagram showing a membership function of an intake air pressure (Pb). FIG.

FIG. 14 is a diagram showing a membership function of an area.

FIG. 15 is a diagram schematically illustrating an example of a method of selecting a route according to a result of fuzzy inference.

FIG. 16 is a functional block diagram of an air-fuel ratio control device according to a second embodiment.

FIG. 17 is a flowchart showing an online learning operation of the air-fuel ratio control device according to the second embodiment.

FIG. 18 is a functional block diagram of an air-fuel ratio control device according to a third embodiment.

FIG. 19 is a flowchart showing the operation of the air-fuel ratio control device according to the third embodiment during online learning.

FIG. 20 is a functional block diagram of an air-fuel ratio control device according to a fourth embodiment.

FIG. 21 is a functional block diagram showing an operation at the time of online learning of the air-fuel ratio control device according to the fourth embodiment.

FIG. 22 is a functional block diagram of an air-fuel ratio control device using a conventional neural network.

FIG. 23 is a conceptual diagram of a neural network used in the air-fuel ratio control device of FIG.

FIG. 24 is a diagram schematically showing a learning process of a conventional neural network.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Neuro calculation part 2 Fuel injection amount calculation part 3 State detection part 11 Neuro estimation part 12 Area judgment part 13 Route change part 14 Reference value acquisition part 15 Estimation data recording part 16 Estimation accuracy judgment part 17a * Temporal generalization judgment part 17a Generalization determining section 17b Learning invalidation section 17c Re-learning section 18 Learning section 18a Learning data generation section 18b Learning data storage section 18c Learning execution section 19a Initial coupling coefficient storage section 19b Initial state return section 20, 20a Neural network update section E engine

Claims (3)

[Claims] 1. A parameter estimating apparatus that receives input of a plurality of input parameter values related to an estimation target parameter and estimates an estimation target parameter value by a neural network from the received input parameter values, wherein the neural network is Learning is performed by changing the combination method as predetermined for each region represented by some of the plurality of input parameters, and the one of the received plurality of input parameter values is Area determining means for determining to which of the areas the parameter belongs from the parameter value of the section; And a parameter estimating device comprising: 2. The method according to claim 1, wherein the neural network performs learning by changing a connection path for each of the regions represented by some of the plurality of input parameters. 2. The parameter estimating apparatus according to claim 1, wherein a path of connection of the neural network is selected according to the region determined by the means. 3. The neural network learns by changing a weight of a connection path for each area represented by a part of the plurality of input parameters. The parameter estimation device according to claim 1, wherein a weight of a connection path of the neural network is changed according to a region determined by the determination unit.