JP3071335B2 - Vehicle driving force control device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a driving force control device for a vehicle, which controls the driving force so that the acceleration of the vehicle becomes the acceleration required by the driver.

[0002]

2. Description of the Related Art Generally, a vehicle is required to run under various environmental conditions, and various driving operations are performed by individual drivers. Then, regarding the behavior of the vehicle, responsiveness and smoothness intended by the driver are required. Conventionally, regarding the behavior related to the driving force of the vehicle,
For example, it is known that in a vehicle equipped with an internal combustion engine, control is performed according to the amount of depression of an accelerator pedal by a driver.

For example, in Japanese Patent Application No. 3-80103 previously proposed by the present applicant, in a vehicle equipped with an internal combustion engine, the opening of a linkless type throttle valve (throttle opening) is determined by the driver. The control is performed according to the depression amount (accelerator stroke) of the accelerator pedal. Here, data for determining the target acceleration corresponding to the accelerator stroke is stored in the backup RAM in advance as a map. Then, the throttle opening is controlled so that the actual acceleration of the vehicle becomes the target acceleration determined from the map, thereby controlling the driving force of the vehicle. Further, in this proposed technique, the change in the accelerator stroke and the actual acceleration are detected as a change in the driver's required acceleration, and the deviation between the detected required acceleration and the target acceleration determined from the above map is determined. Is minimized so that the map data is corrected and stored in the backup RAM. Mathematically, correction (correction) for the target acceleration
Is performed in accordance with the above deviation, thereby rewriting the map. That is, the data of the target acceleration corresponding to the accelerator stroke is learned.

Accordingly, since the target acceleration data is learned so as to match the degree of acceleration required by the driver as described above, the target acceleration suitable for the characteristics of the driver is always determined. . As a result, irrespective of the driver's consciousness or driving environment, a driving force suitable for the driver's characteristics is always obtained, and good driving performance is obtained.

[0005] However, in the above-mentioned proposed technique, learning of the data of the target acceleration simply involves correcting (correcting) the data and rewriting the map. As for the rewriting of the map, only the data of the target acceleration is learned only for a certain point or a certain range of the accelerator stroke at each time.
Therefore, the target acceleration is only corrected (corrected) only in the specific driving region, and the rewritten map is locally biased. As a result, in the rewritten map, the relationship between the target acceleration and the accelerator stroke is partially discontinuous, and the control of the driving force of the vehicle tends to be partially discontinuous with respect to the change in the accelerator stroke. there were.

Therefore, a new technique has been proposed by the present applicant in Japanese Patent Application No. 4-336 with the aim of addressing the above-mentioned problems.
No. 323. In the proposed technology, learning control of a map is performed using a neural network technology. Here, the driver's request for the running of the vehicle is estimated from the actual acceleration at each time as a "required acceleration model", and the "required acceleration model"
The "throttle sensitivity model" is changed based on the throttle sensitivity and the throttle sensitivity is determined. Further, the throttle valve of the engine is controlled to be opened and closed so that the target throttle opening obtained from the product of the determined throttle sensitivity and the accelerator stroke matches the actual throttle opening.

Therefore, since the learning control of the map is performed by using the neural network technology as described above, the entire relationship between the acceleration and the throttle opening with respect to the entire range of the accelerator stroke does not become discontinuous, and the engine drive is performed. Force control can be continuous over the entire range of accelerator strokes by the driver.

[0008]

However, in the latter new proposed technique, the "throttle sensitivity model" is such that the learning control is performed by learning the acceleration deviation, thereby controlling the control amount sensitivity (throttle sensitivity) of the drive source. ) Is to be determined. In this learning, the degree to which the history of past data is not captured, that is, the degree to which new data is captured and updated (learning rate) is always a constant value when the required acceleration increases or decreases. Was held. Accordingly, as shown in FIG. 12, the time from when the throttle sensitivity is lower than the reference value by a certain amount to the reference value, for example, and when the throttle sensitivity is lower than the reference value by a certain amount from the reference value to the reference value, for example, The time to rise was set to be equal.

However, human sensation can generally be more acute when the throttle sensitivity decreases than when it increases. For this reason, in the above proposed technology, the driver could not feel the same when the required acceleration decreases or increases, even though the actual throttle sensitivity increase / decrease speed is the same. In other words, the driver may feel faster when the required acceleration is smaller, that is, when the throttle sensitivity decreases, and may consequently feel that the throttle sensitivity tends to be very insensitive in a short time. Was. As a result, the driver feels a sense of discomfort, and the accelerator operation feeling for the driver becomes insufficient.

SUMMARY OF THE INVENTION The present invention has been made in view of the above-mentioned circumstances, and an object thereof is to realize control of a driving force always suited to the characteristics of a driver irrespective of a driver's consciousness and a driving environment. In addition, it is possible to make the control of the driving force continuous over the entire range of the operation amount of the accelerator pedal or the like by the driver, and at the same time, to appropriately adjust the degree of sensitivity change to the driver's feeling. It is an object of the present invention to provide a possible driving force control device.

[0011]

In order to achieve the above object, according to the present invention, as shown in FIG.
A control amount changing means M3 for changing the control amount of the driving source M2 mounted on the control unit 1; an output operation means M4 operated by a driver to arbitrarily control the output of the driving source M2;
And an operation amount detecting means M5 for detecting an operation amount of the output operation means M4. The output of the driving source M2 is controlled by driving the control amount changing means M3 according to the detection result of the operation amount detecting means M5. A vehicle driving force control device for controlling the driving force of the vehicle M1 by using an acceleration detecting means M6 for detecting the acceleration of the vehicle M1;
1 is used as teacher data to compare acceleration obtained by the detection of the speed detecting means M6 with the speed detecting means M7, and a deviation between the teacher data and the output of the means M8 is used as an error signal. The operation amount of the output operation means M4 and the speed detection means M7 so that the error becomes small.
Required acceleration model learning means M8 for learning the relationship between the acceleration of the vehicle M1 and the speed obtained by the detection of the vehicle as an acceleration model required by the driver, and the acceleration obtained by the detection of the acceleration detecting means M6 and the required acceleration model learning. The deviation from the output of the acceleration model learned by the means M8 is used as an error signal, and the driving amount of the drive source M2 with respect to the operation amount of the output operation means M4 and the speed obtained by the detection of the speed detection means M7 is reduced so that the error is reduced. A control amount sensitivity model learning unit M9 for learning at a predetermined learning rate using a relationship between the control amounts as a control amount sensitivity model, and a control amount sensitivity model learned by the control amount sensitivity model learning unit M9. Using the output as reference data, control is performed in accordance with the operation amount detected by the operation amount detection means M5 based on the reference data. A drive control means M10 for controlling the drive of the amount change means M3, and a predetermined value in which a deviation between the acceleration obtained by the detection of the acceleration detection means M6 and the output of the acceleration model learned by the required acceleration model learning means M8 is a predetermined value. And a sensitivity model for determining whether or not the output of the control amount sensitivity model learned by the control amount sensitivity model learning unit M9 is larger than a predetermined reference value. Output comparison means M
12 and a difference between the acceleration obtained by the detection of the acceleration detecting means M6 and the output of the acceleration model learned by the required acceleration model learning means M8 is smaller than a predetermined value. If the output of the control amount sensitivity model learned by the control amount sensitivity model learning unit M9 is larger than a predetermined reference value in the comparison result of the sensitivity model output comparison unit M12, the control amount sensitivity model learning unit M9 And a learning rate correcting means M13 for correcting the learning rate in the step (b) to a value lower than a predetermined learning rate.

[0012]

According to the above configuration, as shown in FIG. 1, the acceleration obtained by the detection of the acceleration detecting means M6 is used as the teacher data to be compared in the required acceleration model learning means M8. That is, the actual acceleration of the vehicle M1 is used for comparison as the acceleration required by the driver.
In the required acceleration model learning means M8, the deviation between the teacher data and the output of the required means M8 is reduced, that is, the deviation between the acceleration required by the driver and the output of the required acceleration model is reduced. , Output operation means M4
The relationship between the operation amount of the vehicle and the acceleration of the vehicle M1 with respect to the speed obtained by the detection of the speed detecting means M7 is learned as an acceleration model required by the driver. In the control amount sensitivity model learning means M9, the deviation between the acceleration obtained by the detection of the acceleration detection means M6 and the output of the acceleration model learned as described above is used as an error signal. Then, in the control amount sensitivity model learning means M9, the error of the error signal is reduced, that is, the vehicle M
The relationship between the operation amount of the output operation means M4 and the control amount of the drive source M2 with respect to the speed of the vehicle M1 is determined in advance as a control amount sensitivity model so that the deviation between the acceleration 1 and the output of the acceleration model becomes small. Learning is performed at a predetermined learning rate. In the drive control unit M10, the output of the control amount sensitivity model learned as described above is used as reference data, and control is performed based on the reference data in accordance with the operation amount of the output operation unit M4 operated by the driver. Amount changing means M
3 is controlled. As a result, the output of the driving source M2 is controlled, and thus the driving force of the vehicle M1 is controlled.

Therefore, according to the present invention, an acceleration model according to the driver's request is always obtained, and a control amount sensitivity model is obtained based on the acceleration model. Then, the control amount of the drive source M2 is always controlled with the acceleration that matches the driver's request. Further, according to the present invention, the entire relationship between the operation amount of the output operation means M4 and the acceleration of the vehicle M1 with respect to the speed of the vehicle M1 is learned as a continuous model, and the operation amount of the output operation means M4 and the speed of the vehicle M1 are learned. Is learned as a continuous model. Therefore, in the entire range of the speed of the vehicle M1, the relationship between the acceleration and the control amount with respect to the entire range of the operation amount of the output operation unit M4 does not become partially discontinuous.

The deviation comparing means M11 determines whether the deviation between the acceleration obtained by the detection of the acceleration detecting means M6 and the output of the acceleration model learned by the required acceleration model learning means M8 is smaller than a predetermined value. It is determined whether or not. Further, the sensitivity model output comparing means M12 determines whether or not the output of the control amount sensitivity model learned by the control amount sensitivity model learning means M9 is larger than a predetermined reference value. Then, the deviation comparing means M11
Of the acceleration obtained by the detection of the acceleration detecting means M6 and the required acceleration model learning means M8
The deviation from the output of the acceleration model learned by the control model sensitivity model learned by the control quantity sensitivity model learning unit M9 is smaller than a predetermined value, and the comparison result of the sensitivity model output comparison unit M12 is learned by the control amount sensitivity model learning unit M9. Is larger than a predetermined reference value, the learning rate of the control amount sensitivity model learning means M9 is corrected by the learning rate correcting means M13 to a value lower than the predetermined learning rate.

Therefore, when the output of the control amount sensitivity model decreases from a state where the output of the control amount sensitivity model is larger than a predetermined reference value, the learning rate of the control amount sensitivity model learning means M9 is reduced. Is corrected to a value lower than the predetermined learning rate. Therefore, in learning, the degree of taking in the past history of the output of the control amount sensitivity model is larger than in other cases. Therefore, new information that the deviation is smaller than the predetermined value and the output of the control amount sensitivity model is decreasing is less likely to be taken in, and the degree of decrease in the output of the control amount sensitivity model becomes gentle.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of a vehicle driving force control apparatus according to the present invention will be described below in detail with reference to FIGS.

FIG. 2 is a block diagram schematically showing a vehicle driving force control apparatus according to this embodiment. The vehicle 1 is equipped with a gasoline engine (hereinafter simply referred to as “engine”) 2 as a drive source. The engine 2 is an in-line type having a plurality of cylinders. Outside air is taken into the intake passage 3 of the engine 2, and fuel injected from an injector (not shown) is supplied. A mixture of outside air and fuel is taken into each combustion chamber (not shown) of the engine 2 through the intake passage 3. Further, the air-fuel mixture taken into each combustion chamber is exploded and burned by the operation of each of the ignition plugs (not shown), so that the piston and the crankshaft (not shown) operate to obtain the output of the engine 2. Further, the burned gas that has been burned in each combustion chamber of the engine 2 is discharged to the outside through the exhaust passage 4.

In this embodiment, a vehicle 1 is of a front engine / rear drive type (FR type), and a crankshaft of an engine 2 is driven by a transmission, a propeller shaft, a differential gear, a drive shaft, and the like (not shown). The driving wheels are drivingly connected to a pair of left and right rear wheels 5. The pair of left and right front wheels 6 that are driven wheels are steered wheels that are interlocked with the operation of a steering wheel (not shown) provided in a driver's seat.

In this embodiment, a linkless type throttle valve 7 constituting a control amount changing means is provided in the middle of the intake passage 3. That is, the throttle valve 7 is connected to a DC motor 8 provided near the throttle valve 7. The opening of the throttle valve 7 as a control amount of the engine 2, that is, the throttle opening Th, is controlled by the operation of the DC motor 8. Thereby, the amount of air taken into each combustion chamber of the engine 2 through the intake passage 3 is adjusted, and the output of the engine 2 is controlled by adjusting the amount of air.

In the vicinity of the throttle valve 7, a throttle sensor 9 is provided. This throttle sensor 9
Then, the throttle opening Th is detected, and a signal corresponding to the detected throttle opening Th is output. The driver's seat of the vehicle 1 is provided with an accelerator pedal 10 as output operation means. The accelerator pedal 10 is operated by the driver DR to arbitrarily control the output of the engine 2.
In the vicinity of the accelerator pedal 10, an accelerator sensor 11 is provided as operation amount detecting means. The accelerator sensor 11 detects an operation amount of the accelerator pedal 10, that is, an accelerator stroke S, and outputs a signal corresponding thereto. Further, at a substantially center of the vehicle 1, a well-known acceleration sensor 12 constituting an acceleration detecting means is provided. The acceleration sensor 12 detects the acceleration G in the front-rear direction of the vehicle 1 and outputs a signal corresponding thereto.
In addition, the front wheel 6 is provided with a well-known vehicle speed sensor 13 constituting a speed detecting means. In the vehicle speed sensor 13, the speed of the vehicle 1, that is, the vehicle speed V according to the rotation speed of the front wheels 6
Is detected, and a signal corresponding to this is output.

In this embodiment, a throttle computer 21 and a neurocomputer 22 are provided in order to suitably control the opening and closing of the throttle valve 7 in response to a request from the driver DR. The throttle computer 21 constitutes drive control means. The throttle computer 21 includes a DC motor 8 and a throttle sensor 9.
Are electrically connected to each other. Further, the neurocomputer 22 comprises required acceleration model learning means, control amount sensitivity model learning means, deviation comparison means, sensitivity model output comparison means and learning rate correction means, and is configured by applying neural network technology. I have. An accelerator sensor 11, an acceleration sensor 12, and a vehicle speed sensor 13 are electrically connected to the neuro computer 22, respectively. Further, the neurocomputer 22 and the throttle computer 21 are electrically connected to each other.

FIG. 3 is a block diagram showing an electrical configuration of the throttle computer 21 and the neuro computer 22. The neurocomputer 22 includes a central processing unit (CPU) 23 having a counter function, a read-only memory (ROM) 24 in which a predetermined learning control program and the like are stored in advance, and a random access memory (ROM) for temporarily storing the calculation results of the CPU 23 and the like. RAM) 25, a backup RAM 26 for storing data stored in advance, and the like. The neurocomputer 22 is configured as a logical operation circuit in which the units 23 to 26 and the external input / output circuit 27 are connected by a bus 28. The accelerator sensor 11 and the vehicle speed sensor 13 described above are connected to the external input / output circuit 27, respectively. Further, the acceleration sensor 12 is connected to the external input / output circuit 27 via a low-pass filter 29. The low-pass filter 29 allows a signal having a frequency lower than a predetermined reference cutoff frequency among the detection signals of the acceleration sensor 12 to pass freely, and gives a large attenuation to a high frequency. In addition, the throttle computer 21 described above is connected to the external input / output circuit 27. ROM2
4 stores in advance a learning control program using a neural network technology.

The CPU 23 is connected to the external input / output circuit 2
Various signals from each of the sensors 11 to 13 which are input through the interface 7 and the like are read as input values. Based on these input values, the CPU 23 performs learning control of a “requested acceleration model” requested by the driver DR and a “throttle sensitivity model” as a control amount sensitivity model according to the learning control program stored in the ROM 24. Execute And
The CPU 23 outputs the learning result to the throttle computer 21 via the external input / output circuit 27.

On the other hand, the throttle computer 21 has basically the same configuration as the neuro computer 22, and includes a CPU 30, a ROM 31, a RAM 32, a backup RAM 33, an external input / output circuit 34, a bus 35 and the like. The above-described DC motor 8, throttle sensor 9 and neurocomputer 22 are connected to the external input / output circuit 34, respectively. ROM3
1 stores in advance a throttle opening control program for controlling the opening and closing of the throttle valve 7 based on the learning result of the neurocomputer 22 or a separately set value.

Then, the CPU 30 reads, as an input value, data of a learning result input from the neurocomputer 22 via the external input / output circuit 34. Also, CPU3
0 reads a signal from the throttle sensor 9 as an input value. Further, the CPU 30 determines R based on the input values.
The DC motor 8 is suitably controlled according to the throttle opening control program stored in the OM 31.

Here, a conceptual configuration of the neural network technology applied to the neurocomputer 22 will be described with reference to FIGS. 4 (a) and 4 (b). FIGS. 4A and 4B show a neural network in this embodiment.
As shown in FIG. 1, two multilayer neural networks are provided. Each multilayer neural network basically has the same configuration, and includes an “input layer” composed of two neurons n1, an “intermediate layer” composed of 2 to 10 neurons n2, and one neuron. and an “output layer” made of n3. Also, each neuron n between each layer
1, n2 and n3 are connected by a synapse sp.
In each multilayer neural network, signals flow in one direction from the “input layer” to the “intermediate layer” and from the “intermediate layer” to the “output layer”. Neurons n1, n2 in each layer
In n3, the state is determined based on the signal received from the previous layer, and the signal is transmitted to the next layer. The output result of each multilayer neural network is "output layer"
As the state value of the neuron n3.

Here, the multi-layered neural network shown in FIG. 4 (a) is provided for each neuron n1 in the "input layer".
An accelerator stroke S detected by the accelerator sensor 11 and a vehicle speed V detected by the vehicle speed sensor 13 are input. Further, the output obtained from the neuron n3 of the “output layer”, that is, the required acceleration model output Gx is compared with the “teacher data” using the acceleration G of the vehicle 1 obtained by the detection of the acceleration sensor 12 as the “teacher data”. You. Then, the acceleration deviation ΔG (= GG−G)
x) is defined as an “error signal”, and the “weight coefficients” of the synapses sp of all the neurons n1, n2, and n3 are corrected in a direction in which the error decreases. That is, the acceleration G of the vehicle 1
Is the acceleration required by the driver DR, and the acceleration G is “teacher data” to be compared. Then, the relationship between the acceleration G with respect to the accelerator stroke S and the vehicle speed V is learned as a "requested acceleration model" required by the driver DR so that the deviation from the "teacher data" is reduced. The output result of the multilayer neural network is obtained as a required acceleration model output Gx. That is, the “required acceleration model” is learned as a characteristic as shown in FIG. 5 in a direction in which the required acceleration model output Gx approaches the acceleration G.

On the other hand, in the multilayer neural network shown in FIG. 4B, the accelerator stroke S and the vehicle speed V are input to each neuron n1 of the "input layer" in the same manner as described above. The output result obtained from the neuron n3 of the “output layer”, that is, the throttle sensitivity model output Thx is
Learning is performed using the acceleration deviation ΔG between the acceleration G and the required acceleration model output Gx as an “error signal”, and the “weight coefficient” of the synapse sp of all the neurons n1, n2, and n3 is corrected. That is, the relationship between the accelerator stroke S and the throttle opening Th with respect to the vehicle speed V is learned as a "throttle sensitivity model" required by the driver DR so that the acceleration deviation ΔG is set as an “error signal” and the error is reduced. You. However, this learning is performed at a learning rate ε described later. In other words, the degree of taking in new data is changed according to the learning rate ε switched according to various operating states.

The output result of the multilayer neural network is obtained as a throttle sensitivity model output Thx. That is, the "throttle sensitivity model" is learned as a characteristic as shown in FIG.

The conceptual configuration of the neural network as described above has been described for the sake of convenience only, and the actual entity of the neural network is in a learning control program stored in the ROM 24 of the neurocomputer 22 in advance. The neural network is based on a mathematical operation in the learning control program, and a generally known “error back propagation learning algorithm” is applied as a learning method.
In this embodiment, a learning control program is created to finally obtain the characteristic of the relationship between the throttle stroke Thg and the accelerator stroke S as shown in FIG.

Next, a description will be given of a processing operation for learning a "required acceleration model", a "throttle sensitivity model", and the like, which is executed by the neural computer 22 using the above-described neural network technology.
FIG. 8 is a flowchart showing a “learning control routine” of a learning control program executed by the neurocomputer 22. After the processing of this routine is started, it is periodically executed at a constant cycle, for example, at a time interval of “0.1 second”.

When the processing of this routine is started, in step 101, an accelerator stroke S, an acceleration G and a vehicle speed V are read based on various signals from the accelerator sensor 11, the acceleration sensor 12 and the vehicle speed sensor 13, respectively.

Subsequently, in step 102, a "throttle sensitivity model" is executed. That is, the throttle sensitivity model output Thx is obtained from the characteristics of the already learned “throttle sensitivity model” (see FIG. 6) using the accelerator stroke S and the vehicle speed V read this time as input values. Subsequently, the throttle sensitivity Thg is determined from the throttle sensitivity model output Thx according to the following equation (1).

Thg = α + Thx * k1 (1) where “α” is a reference value, and in this embodiment, “α =
1.0 ". “K1” is a positive constant.

Then, at step 103, the throttle sensitivity Thg and the accelerator stroke S determined this time are determined.
Is output to the throttle computer 21. Or,
The product of the throttle sensitivity Thg and the accelerator stroke S,
That is, the target throttle opening Thg · S is obtained, and the target throttle opening Thg · S is determined by the throttle computer 21.
Output to

Next, in step 104, learning of a "required acceleration model" required by the driver DR is performed using the acceleration G of the vehicle 1 as "teacher data". In other words, as the “teacher data” to be compared with the acceleration G of the vehicle 1 detected by the acceleration sensor 12, the relationship between the acceleration G with respect to the accelerator stroke S and the vehicle speed V is set so that the deviation from the “teacher data” is reduced. The learning is performed as the "required acceleration model" required by the driver DR.

For example, it is assumed that the curve shown by the solid line in FIG. 5 is the characteristic of the current "required acceleration model". Then, in order for the driver DR to run faster than at present, the accelerator pedal 10
Is operated to make the acceleration G of the vehicle 1 larger than the current required acceleration model output Gx. The acceleration G at this time is a new required acceleration, and the “required acceleration model”
Is updated from a curve shown by a solid line in FIG. 5 to a characteristic like a curve shown by a broken line. That is, the entire relationship between the accelerator stroke S and the required acceleration model output Gx with respect to the vehicle speed V is learned as a continuous model. This characteristic does not become partially discontinuous.

FIG. 5 shows the characteristics when the vehicle speed V is "0". In the "required acceleration model", the entire range of the accelerator stroke S, the entire range of the vehicle speed V, and the acceleration of the vehicle 1 are shown. The relationship with G is learned.

Subsequently, in step 105, a difference between the acceleration G and the required acceleration model output Gx is obtained as an acceleration deviation ΔG. In step 106, the throttle sensitivity Thg determined in this routine is set.
Is larger than a predetermined reference value Thgstd and whether the acceleration deviation ΔG obtained in this routine is smaller than a predetermined value φ. However, in this embodiment, the reference value Thgstd is “1.
0 ", and the predetermined value φ is" 0 ". That is, in this step 106, the throttle sensitivity Thg becomes "1.
From a relatively sensitive state larger than "0", it is determined whether or not the required acceleration is reduced, the acceleration deviation ΔG becomes negative, and the throttle sensitivity Thg is decreasing. Then, the throttle sensitivity Thg becomes equal to the reference value Thgst.
If it is larger than d and the acceleration deviation ΔG is smaller than the predetermined value φ, the process proceeds to step 107.

In step 107, the basic learning rate ε
A value obtained by multiplying ₀ by a constant k2 is set as a learning rate ε.
However, in this embodiment, the basic learning rate ε ₀ is “0.
2 "and the constant k2 is" 0.5 ". That is, in this step 107, the learning rate ε is “0.2 * 0.5
= 0.1 ".

On the other hand, if the throttle sensitivity Thg is not larger than the reference value Thgstd or the acceleration deviation ΔG is not smaller than the predetermined value φ in step 106, the process proceeds to step 108. Then, in step 108, the basic learning rate ε ₀ is directly used as the learning rate ε
Set as That is, in this step 108, the learning rate ε is set to “0.2”.

Finally, step 107 or step 10
8, and in step 109, the acceleration deviation ΔG obtained in the current routine is used as an error signal.
The learning of the “throttle sensitivity model” is executed at the learning rate ε set in step 107 or step 108. Then, the subsequent processing ends once. That is, the relationship between the accelerator stroke S and the throttle opening Th with respect to the vehicle speed V is learned as a "throttle sensitivity model" so that the acceleration deviation ΔG is used as an error signal so that the value becomes smaller.

For example, it is assumed that the solid line shown in FIG. 6 is the initial value of the "throttle sensitivity model". Then, when the driver DR operates the accelerator pedal 10 to run faster than the current time, the acceleration G of the vehicle 1 increases, and a difference is generated between the acceleration G and the required acceleration model output Gx, and the acceleration deviation ΔG at that time is obtained. Can be When the acceleration deviation Δ is used as an error signal and learning is performed so that the value becomes smaller, the throttle sensitivity model output Thx at that time is a new throttle sensitivity model output Thx, and the “throttle sensitivity model” in FIG. Is updated from the initial value indicated by the solid line to a characteristic such as a curve indicated by the broken line. That is, the entire relationship of the throttle sensitivity model output Thx with respect to the accelerator stroke S and the vehicle speed V is learned as a continuous model. This characteristic does not become partially discontinuous.

However, in the above learning, the throttle sensitivity T
If hg is larger than reference value Thgstd and acceleration deviation ΔG is smaller than predetermined value φ, learning rate ε is set to relatively small “0.1”; otherwise, learning rate ε is set to It is set to “0.2” which is relatively large. Therefore, when the throttle sensitivity Thg is larger than the reference value Thgstd and the acceleration deviation ΔG is smaller than the predetermined value φ, the degree of taking in the past history of the throttle sensitivity Thg becomes larger than in other cases. . Therefore, new information that the acceleration deviation ΔG is smaller than the predetermined value φ and the throttle sensitivity Thg is decreasing is less likely to be captured. That is, it becomes difficult to update the characteristics.

In this way, the learning control process using the neural network technology is executed, and the driver DR
The characteristics of the "required acceleration model" and "throttle sensitivity model" required by the above are learned, respectively. Here, the “weight coefficient” of the synapse sp as a characteristic of the “required acceleration model” and the “throttle sensitivity model” learned at each time is rewritten and stored in the backup RAM 26.

The initial value of the "weighting coefficient" at the time of shipment of the vehicle 1 from the factory is determined as follows. First, in the required acceleration model, the acceleration G when the vehicle 1 is driven by a plurality of drivers DR is learned as “teacher data”, and an average characteristic of the learning result is set as an initial value. And the "weighting factor" of the throttle sensitivity model is
The learning is performed so that the throttle sensitivity model output Thx becomes “0”.

Next, the processing operation of the throttle opening control executed by the throttle computer 21 based on the throttle sensitivity Thg determined by the above processing operation and the accelerator stroke S at that time will be described. FIG. 9 is a flowchart showing a “throttle opening control routine” of the throttle opening control program executed by the throttle computer 21. After the processing of this routine is started, it is periodically executed at predetermined time intervals.

When the processing of this routine is started, first, at step 201, when learning the throttle sensitivity Thg, the throttle opening Th is read based on a signal from the throttle sensor 9. Also, Neuro Computer 2
2 and the latest throttle sensitivity Thg and accelerator stroke S, or the target throttle opening Thg · S
Read. Here, if the reading of the throttle sensitivity Thg and the accelerator stroke S is premised, in step 201, the product of both Thg and S is obtained as the target throttle opening Thg · S.

Subsequently, in step 202, it is determined whether or not the current throttle opening Th is smaller than the target throttle opening Thg · S. Here, if the throttle opening Th is smaller than the target throttle opening Thg · S, in step 203, the DC motor 8 is rotated forward so as to drive the throttle valve 7 in the opening direction.
In step 204, the throttle opening Th is read based on the signal from the throttle sensor 9.

Then, in step 205, it is determined again whether or not the throttle opening Th is smaller than the target throttle opening Thg · S. Here, the throttle opening T
If h is smaller than the target throttle opening Thg · S, the routine jumps to step 203, where the throttle valve 7
Steps 203 and 2 in order to further drive the
Steps 04 and 205 are repeated. On the other hand, if the throttle opening Th is equal to or larger than the target throttle opening Thg · S, the throttle valve 7 is not driven further in the opening direction, and the subsequent processing is temporarily terminated.

On the other hand, if it is determined in step 202 that the current throttle opening Th is not smaller than the target throttle opening Thg · S, the process proceeds to step 206. Then, in step 206, it is determined whether or not the current throttle opening Th is larger than the target throttle opening Thg · S. Here, if the throttle opening Th is not larger than the target throttle opening Thg · S, the subsequent processing is once ended as it is.

If the throttle opening Th is larger than the target throttle opening Thg · S in step 206, the DC motor 8 is reversed in step 207 so as to drive the throttle valve 7 in the closing direction. In step 208, the throttle sensor 9
The throttle opening Th is read based on the signal from the controller.

Then, in step 209, it is determined again whether or not the throttle opening Th is larger than the target throttle opening Thg · S. Here, the throttle opening T
If h is larger than the target throttle opening Thg · S, the routine jumps to step 207, where the throttle valve 7
207 and 20 in order to further drive the
Steps 8 and 209 are repeated. On the other hand, when the throttle opening Th is equal to or smaller than the target throttle opening Thg · S, it is determined that the throttle valve 7 is not driven in the closing direction any more, and the subsequent processing is temporarily terminated.

As described above, the DC motor 8 is controlled so that the throttle opening Th matches the target throttle opening Thg · S.
Is controlled, whereby the throttle valve 7 is controlled to open and close. Thus, the output of the engine 2 is controlled, and as a result, the driving force of the vehicle 1 is controlled.

As described above, in this embodiment, when learning for obtaining the throttle sensitivity Thg, the vehicle 1
The driver DR's request for the running of the vehicle is estimated from the acceleration G at that time as a “required acceleration model”. Further, the "throttle sensitivity model" is changed based on the "required acceleration model" to determine the throttle sensitivity Thg.
Further, a target throttle opening T obtained from the product of the determined throttle sensitivity Thg and the accelerator stroke S.
hg · S matches the throttle opening Th.
The opening and closing of the throttle valve 7 is controlled. In addition, a "required acceleration model" is always obtained according to the request of the driver DR, and a "throttle sensitivity model" is obtained based on the "required acceleration model". Then, the throttle opening Th of the engine 2 is always controlled with the acceleration G that meets the request of the driver DR. Therefore, when the required acceleration of the driver DR for the vehicle 1 is large, the throttle sensitivity T
hg increases. Therefore, the change range of the accelerator stroke S for obtaining the same acceleration G is narrowed, and a large acceleration G can be obtained by operating the accelerator pedal 10 with a small amount, and driving is performed as if the acceleration performance of the vehicle 1 was improved. Person DR can be felt. For example, the driver DR is in a hurry, or the vehicle 1
When the driving environment of the vehicle is on a highway without traffic congestion and the vehicle 1 is to be driven at a high speed, a large acceleration G can be obtained by a small operation of the accelerator pedal 10,
The feeling of acceleration can be improved.

On the other hand, when the required acceleration of the driver DR for the vehicle 1 is small, the throttle sensitivity Thg becomes small. Therefore, the change range of the accelerator stroke S for obtaining the same acceleration G is widened, and the accelerator pedal 1
The acceleration G can be delicately changed by an operation with many zeros, and the operation performance of the accelerator pedal 10 for the driver DR can be improved. For example,
The driver 1 may be in a relaxed state, or the driving environment of the vehicle 1 may be a congested road or a snowy road.
When you want to run slowly, the accelerator pedal 1
The acceleration G can be finely changed by an operation with many zeros, and the operational feeling of the vehicle 1 can be improved.

That is, in this embodiment, since the learning is performed so as to match the degree of demand for the acceleration of the driver DR, the throttle sensitivity Thg that always matches the characteristics of the driver DR is determined. As a result, the driver DR's consciousness (hurrying, leisurely, etc.) and driving environment (road surface, day / night, in a tunnel, rainy road, snowy road,
Regardless of the mountain road, congested road, etc.),
It is possible to always control the driving force that matches the characteristics of the driver DR.

In this embodiment, since the neural network technology is used for learning control in the neurocomputer 22, the accelerator stroke S and the vehicle speed V
Is learned as a continuous model, and the characteristics thereof do not become partially discontinuous. Similarly, the entire relationship of the throttle sensitivity model output Thx with respect to the accelerator stroke S and the vehicle speed V is learned as a continuous model, and its characteristics do not become partially discontinuous. This is because the "required acceleration model" learned between discontinuous points of the accelerator stroke S and the vehicle speed V is interpolated by using the neural network technology. That is, the correction of the required acceleration model output Gx performed for a specific range of the accelerator stroke S and the vehicle speed V is also reflected in the correction of the required acceleration model output Gx corresponding to another range of the accelerator stroke S and the vehicle speed V. . Further, the correction of the throttle sensitivity model output Thx performed for a specific range of the accelerator stroke S and the vehicle speed V is also reflected on the throttle sensitivity model output Thx corresponding to another range of the accelerator stroke S and the vehicle speed V.

As a result, the driving force of the vehicle 1 is controlled by the driver DR over the entire range of the vehicle speed V.
The operation amount is 0, that is, the operation amount can be continuous over the entire operation range of the accelerator stroke S. Therefore, when the accelerator pedal 10 is continuously depressed by the driver DR, the acceleration G of the vehicle 1 does not suddenly change, and the increase in the vehicle speed V can always be made smooth.

Further, in this embodiment, a "requested acceleration model" is estimated from the acceleration G, and the "required acceleration model" is obtained.
Learning is performed to change the "throttle sensitivity model" based on
As compared with the case where the map is rewritten, the interpolation calculation of the map becomes unnecessary, and the calculation time can be further shortened.

The feature of this embodiment is that the learning of the "throttle sensitivity model" is executed at one of the two learning rates ε using the acceleration deviation ΔG as an error signal. did. That is, at the time of learning, when the throttle sensitivity Thg is larger than the reference value Thgstd and the acceleration deviation ΔG is smaller than the predetermined value φ, the learning rate ε is set to a relatively small value “0.1”. In this case, the learning rate ε is relatively large “0.
2 ". Then, the learning of the "throttle sensitivity model" is executed at the set learning rate ε. Therefore, when the throttle sensitivity Thg is larger than the reference value Thgstd and the acceleration deviation ΔG is smaller than the predetermined value φ,
When hg is decreasing from a relatively large state,
The degree to which the past history of the throttle sensitivity Thg is captured is larger than in other cases. Therefore, it becomes difficult to capture new information that the throttle sensitivity Thg is decreasing from a relatively large state. That is, it becomes difficult to update the characteristics. Therefore, as shown in FIG. 10, when the throttle sensitivity Thg is decreasing from a relatively large state, the throttle sensitivity Thg
Is more gradual compared to the case where is increasing from a relatively small state. That is, the throttle sensitivity Thg takes a relatively long time and becomes insensitive. As a result, the driver does not feel that the throttle sensitivity Thg becomes very insensitive in a short time, and the driver does not feel uncomfortable. That is,
The degree of change in sensitivity can be appropriately adapted to the driver's feeling.

In addition, in this embodiment, in the neurocomputer 22, the acceleration sensor 12 is connected to an external input / output circuit 27 via a low-pass filter 29. Therefore, when noise occurs in the detection signal of the acceleration sensor 12 due to harshness when the vehicle 1 is traveling on an uneven road, a high-frequency component correlated with the noise is attenuated by the low-pass filter 29. You. That is,
The signal of the acceleration G from the acceleration sensor 12 is shown in FIG.
Even when a large noise due to harshness is included as shown in FIG. 11A, the signal passes through the low-pass filter 29, and is converted into a signal of the acceleration G with little noise as shown in FIG. Will be adjusted.

As a result, in the neurocomputer 22, the signal of the acceleration G used for learning can be obtained by removing noise caused by harshness. Therefore, it is possible to prevent erroneous learning about the “required acceleration model” and “throttle sensitivity model”, and to prevent the throttle sensitivity Thg from being adjusted in the wrong direction.

It should be noted that the present invention is not limited to the above-described embodiment, and may be carried out as follows by appropriately changing a part of the configuration without departing from the spirit of the invention. (1) In the above embodiment, when the throttle sensitivity Thg is larger than the reference value Thgstd and the acceleration deviation ΔG is smaller than the predetermined value φ, the learning rate ε is set to “0.1”, and otherwise it is not. In this case, the learning rate ε is set to “0.2”, but the learning rate ε is not limited to the above numerical value. That is, the throttle sensitivity Thg is larger than the reference value Thgstd and the acceleration deviation ΔG
Is smaller than the case where the learning rate ε is smaller than the predetermined value φ.

(2) In the above embodiment, the gasoline engine 2 is used as the drive source and the linkless type throttle valve 7 is used as the control amount changing means. However, the present invention is embodied in other drive sources and control amount change means. You can also. For example,
In an electric vehicle, a motor such as a DC motor may be used as a drive source, and a current control circuit or the like for controlling a current to the motor may be used as a control amount changing unit.

(3) In the above embodiment, the accelerator pedal 10 is used as the output operation means operated by the driver DR. However, an accelerator lever or another operation member may be used as the output operation means.

(4) In the above embodiment, the accelerator sensor 1
Although the accelerator stroke S is detected by using 1 as the operation amount detecting means, the following method can be used. That is, instead of the accelerator stroke S, a sensor for detecting the accelerator pedal force may be used, or an accelerator sensor for detecting the accelerator stroke S and a sensor for detecting the accelerator pedal force may be used in combination.

(5) In the above embodiment, the acceleration sensor 12 is used as the acceleration detecting means.
Instead of 2, the vehicle speed sensor 13 may be employed as the acceleration detecting means, and the acceleration G may be obtained by differentiating the vehicle speed V detected by the vehicle speed sensor 13 with time.

(6) In the above embodiment, the neural network technology in the neurocomputer 22 is as follows.
Although a multilayer neural network is employed, an interconnected neural network may be employed.

(7) In the above embodiment, the throttle sensitivity Thg is obtained by the equation (1) using the throttle sensitivity model output Thx as the throttle sensitivity correction amount. However, the throttle sensitivity Thg may be directly output. In this case, the initial value at the time of factory shipment of the vehicle 1 is learned so as to output the reference value α.

[0071]

As described in detail above, according to the present invention, vehicle acceleration is used as teacher data to be compared,
The relationship of the vehicle acceleration to the operation amount of the output operation means,
It learns as the acceleration model required by the driver.
Further, a deviation between the learned output of the acceleration model and the acceleration of the vehicle is calculated, and the relationship between the operation amount of the output operation means and the control amount of the drive source is controlled as a control amount sensitivity model so as to reduce the deviation. doing. Then, using the output of the learned control amount sensitivity model as reference data,
The control amount of the drive source is controlled according to the operation amount of the output operation means.

Therefore, an acceleration model corresponding to the driver's request is obtained, and a control amount sensitivity model is obtained corresponding to the acceleration model, and the control amount is always controlled with the acceleration that matches the driver's request. . In addition, the entire relationship between the acceleration and the control amount with respect to the operation amount is learned as a continuous model, and the entire relationship between the acceleration and the control amount with respect to the entire operation range of the output operation means becomes discontinuous. There is no.

As a result, it is possible to always control the driving force according to the characteristics of the driver regardless of the driver's consciousness or driving environment. Moreover, there is an excellent effect that the control of the driving force can be made continuous over the entire range of the operation amount of the accelerator pedal or the like by the driver.

Further, according to the present invention, when learning as a control amount sensitivity model, when the output of the control amount sensitivity model is larger than a predetermined reference value and the output is decreasing, the learning rate Is corrected to a value lower than the predetermined learning rate. Therefore, the output of the control amount sensitivity model decreases over a long period of time. As a result, the driver does not feel that the control amount sensitivity has become extremely insensitive in a short time, the driver does not feel uncomfortable, and the degree of sensitivity change can be improved in the driver's sense. It has an excellent effect that it can be adapted.

[Brief description of the drawings]

FIG. 1 is a conceptual configuration diagram showing a basic conceptual configuration of the present invention.

FIG. 2 is a schematic configuration diagram showing a schematic configuration of a vehicle driving force control device according to an embodiment of the present invention;

FIG. 3 is a block diagram showing an electrical configuration of a throttle computer and a neuro computer in one embodiment.

FIG. 4 is a configuration diagram showing a conceptual configuration of a multilayer neural network applied to a neurocomputer in one embodiment.

FIG. 5 is a characteristic diagram showing characteristics of a “demanded acceleration model” in one embodiment.

FIG. 6 shows a “throttle sensitivity model” in one embodiment.
FIG. 4 is a characteristic diagram showing the characteristics of FIG.

FIG. 7 is a characteristic diagram showing characteristics of “throttle sensitivity” in one embodiment.

FIG. 8 is a flowchart showing a “learning control routine” executed by the neurocomputer in one embodiment.

FIG. 9 is a flowchart showing a “throttle opening control routine” executed by a throttle computer in one embodiment.

FIG. 10 is a graph showing a relationship between a throttle sensitivity and a time in one embodiment.

FIG. 11 is a time chart showing a change in an acceleration signal related to the operation of a low-pass filter in one embodiment.

FIG. 12 is a graph showing the relationship between throttle sensitivity and time in the related art.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Vehicle, 2 ... Engine as a drive source, 7 ... Throttle valve, 8 ... DC motor (7 and 8 constitute control amount change means), 10 ... Accelerator pedal as output operation means, 11 ... Operation Accelerator sensor as amount detecting means, 12... Acceleration sensor as acceleration detecting means, 13
... Vehicle speed sensor as speed detecting means, 21 ... Throttle computer constituting drive control means, 22 ... Required acceleration model learning means, control amount sensitivity model learning means, deviation comparison means, sensitivity model output comparison means, and learning rate correction means Neuro computer to configure.

Continuation of the front page (51) Int.Cl. ⁷ Identification code FI F02D 29/02 301 F02D 29/02 301Z 41/04 310 41/04 310C G05B 13/02 G05B 13/02 L 13/04 13/04 (72 ) Inventor ▲ Yoshi ▼ Hiroyuki 41, 41, Chuchu-ji, Yokomichi, Nagakute-cho, Aichi-gun, Aichi Prefecture Inside Toyota Central R & D Laboratories Co., Ltd. 78539 (JP, A) JP-A-4-112935 (JP, A) JP-A-61-8434 (JP, A) JP-A-5-92731 (JP, A) JP-A-4-314940 (JP, A) JP-A-4-71933 (JP, A) JP-A-3-96636 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) F02D 45/00 370 F02D 45/00 340 F02D 29 / 02 301 F02D 41/04 310

Claims (1)

(57) [Claims] A control amount changing unit configured to change a control amount of a drive source mounted on a vehicle; an output operation unit operated by a driver to arbitrarily control an output of the drive source; Operating amount detecting means for detecting the operating amount of the output operating means, and controlling the output of the drive source by driving the control amount changing means in accordance with the detection result of the operating amount detecting means. In a vehicle driving force control device configured to control a driving force of a vehicle, an acceleration detecting unit for detecting an acceleration of the vehicle, a speed detecting unit for detecting a speed of the vehicle, and the acceleration detecting unit As the teacher data to be compared with the acceleration obtained by the detection of the error, the deviation between the teacher data and the output of the means is used as an error signal, and the operation amount of the output operation means is reduced so that the error is reduced. Required acceleration model learning means for learning, as an acceleration model required by the driver, the relationship between the acceleration of the vehicle and the speed obtained by the detection of the speed detection means; and the acceleration obtained by the detection of the acceleration detection means And an output of the acceleration model learned by the required acceleration model learning means as an error signal, and an operation amount of the output operation means and a speed obtained by detection of the speed detection means such that the error is reduced. Control amount sensitivity model learning means for learning at a predetermined learning rate using the relationship of the control amount of the drive source to the control amount as a control amount sensitivity model; and control learned by the control amount sensitivity model learning means. The output of the mass sensitivity model is used as reference data, and is detected by the manipulated variable detection means based on the reference data. Drive control means for controlling the drive of the control amount changing means in accordance with the operation amount; and a deviation between the acceleration obtained by the detection of the acceleration detection means and the output of the acceleration model learned by the required acceleration model learning means. Deviation comparing means for determining whether or not the control amount sensitivity model learning means is smaller than a predetermined value; determining whether or not the output of the control amount sensitivity model learned by the control amount sensitivity model learning means is larger than a predetermined reference value. A sensitivity model output comparing means for determining, and a deviation between an acceleration obtained by the detection of the acceleration detecting means and an output of an acceleration model learned by the required acceleration model learning means is determined in advance in a comparison result of the deviation comparing means. Smaller than the predetermined value, and
When the output of the control amount sensitivity model learned by the control amount sensitivity model learning unit is larger than a predetermined reference value in the comparison result of the sensitivity model output comparison unit, the learning by the control amount sensitivity model learning unit is performed. And a learning rate correcting means for correcting the rate to a value lower than the predetermined learning rate.