JP3387004B2 - 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 control device, and more particularly to a control device to which neuro theory is applied.

[0002]

2. Description of the Related Art In a currently applied automobile control device, a program control using a digital computer, as represented by fuel control, is the mainstream.

An example of such program control is, for example, US Pat. No. 4,542,730 or US Pat.
Those described in the specification of 785,783 are known.

[0004]

In such a conventional technique, the operating condition of the vehicle (engine, transmission, brake, vehicle height, suspension, etc.).
Many sensors are used to detect the.

These sensor outputs are individually reflected in the control amount, or individually reflected in the learning control correction amount.

However, there is a problem that the control amount may not be reflected correctly in the physical amount.

An object of the present invention is to control an actuator by inputting an output target value of a sensor so that a desired control can be performed.
It is to provide a control device that can obtain a result .

[0008]

SUMMARY OF THE INVENTION The above object is actuated
And the actuator that changes depending on the operation of the actuator.
Output of the first sensor that detects the state quantity
An output of a second sensor for detecting a state quantity different from the state quantity,
Inputting the output target value of the first sensor to the input layer
And a neuro computer.
Actuator based on signals from the output layer of the data
It is achieved by outputting a control signal to.

[0009]

[0010]

According to the above solving means, the actuator can be controlled in consideration of the target value of the physical quantity and the detected value (control result) of the physical quantity, so that the control is properly performed.

[0012]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the present invention is shown in FIG. The engine 1 has an intake pipe 20 and an exhaust pipe 21.
The fuel injection valve 6 is provided in a part of the intake pipe 20 to supply fuel. The intake air amount is measured by the air amount sensor 22 and input to the control circuit 2. Signals from the crankshaft rotation angle detector 3, the oxygen sensor 4 and the like are input to the control circuit 2 from the engine, and an ignition signal is input to the ignition plug 7 via the ignition coil 5. A knock sensor 8 for detecting mechanical vibration of the engine is attached to the engine block. The signal from the knock sensor 8 is input to the sample hold circuit 11 via the amplifier 9. This input signal 10
Is a time series signal. The sample hold circuit 11
The time series signal 10 is sampled and held at a constant time period, and is input to the input phase of the neurocomputer 13 as spatial multivariable signals 12a, 12b, 12c, ..., 12n in the order of input. The output 14 of the neurocomputer is output as a voltage proportional to the strength of knocking of the engine, and is converted from an analog signal to a digital signal by the A / D converter of the control circuit 2 to control the ignition timing. On the other hand, the signal 15 from the control circuit 2 controls the sample hold circuit 11.

FIG. 2 shows a control flow of the ignition timing θ _ig .
When the control flow starts, in step 25 the basic ignition timing θ _O is read from the memory map. θ _O is determined by the engine speed and the load (fuel injection time, throttle valve opening, etc.). In step 26, the ignition timing θ is calculated as θ _ig = θ _O ± Δθ. Here, Δθ is a value calculated for correction. In step 27, θ _ig is set and ignited. In step 28, the knock intensity is measured from the output 14 of the neuro computer 13. In step 29, the correction amount Δθ is calculated. Here, Δθ is − depending on the size of the output 14.
Δθ (retard angle amount) to + Δθ (advance angle amount) are calculated.

In-cylinder natural frequency F, which is generally generated by knocking of the engine, is proposed by the following equation by the draper.

[0015]

[Equation 1]

Here, B: cylinder bore diameter C: sound velocity U _um : eigenvalue according to vibration mode According to the equation (1), knocking is determined by the relationship between B and C. Therefore, the knocking frequency changes depending on the bore diameter of the engine.

FIG. 3 shows the relationship between the frequency and the output of the knock sensor 8. In the non-resonant type, the output is at the same level over the entire frequency range, but the sensor output is small. On the other hand, the resonance type has a large sensor output at a specific frequency. Further, the wide-range resonance type has an output at the resonance point smaller than that of the resonance type, but the resonance range is wide. The wide-range resonance type is suitable for the present invention in relation to the knocking frequency of FIG. However, even the resonance type can be used by designing the resonance point according to the cylinder bore diameter.

FIG. 4 shows the relationship between the frequency of the pressure ratio of the in-cylinder pressure when knock occurs and when there is no knock.
In FIG. 4, 7.9 KHz, 13.8 KHz, 18.5
There is a point where the pressure ratio becomes large at KHz and 22 KHz, and it can be seen that the pressure at the time of knocking becomes large in a wide frequency range. Therefore, the wide-range resonance type, which can detect many knocking frequencies, is advantageous.

FIG. 5 shows the relationship between the crank angle and the sensor output. FIG. 5A shows the case of the resonance type. The knocking phenomenon occurs at the pressure when the pressure wave reciprocates in the cylinder due to the fact that the air-fuel mixture in the cylinder rises in the fuel and the pressure in the cylinder rises to cause the mixture to ignite spontaneously. For this reason, knocking occurs later than ignition. In the case of the resonance type, the sensor output is generated at the resonance frequency. Meanwhile, FIG.
(B) is a wide-range resonance type. As shown, it is output as a composite wave of many frequencies. In either case, knocking occurs with ignition as the starting point and thereafter.
Therefore, when the sample-hold circuit 11 is controlled by the control circuit 2 with the signal 15, it is sufficient to generate the sample start signal with ignition as the base point.

On the other hand, the sample period of the sample hold circuit 11 is preferably about 1/10 of the maximum measurement frequency in the case of the wide range resonance type. That is, the sampling period is required to intermittently measure the knocking signal and restore the frequency change of the original signal. Further, the sampling period may be a period in which the minimum measurement frequency is at least one cycle or more.

In the case of a resonance type sensor, 1 / of the resonance frequency
It is good to sample for 10 cycles or more for 1 cycle or more.

FIG. 6 shows a neuro element 23 which is a component of the neuro computer 13. Inputs O _k , O _j , O _l
And weighting factors W _ik , W _ij , and W _il are attached to
The output O _i is

[0023]

[Equation 2]

Where θ: threshold value Is.

The output O _i is output to the next neuro element.

On the other hand, for the output O _i , as shown in FIG. 7, the output changes stepwise to 1 when it becomes larger than the threshold value θ, or as shown in FIG.

[0027]

[Equation 3]

And a sigmoid function is output. The sigmoid function is advantageous when the knocking intensity and the output are proportional to each other as in the present invention.

FIG. 9 shows a part of the configuration of the neuro element in the neuro computer. 3 as input layer
0a, 30b, 30c and sample hold circuit 11
Signals are sequentially input to 12a, 12b, and 12c. The output of the neuro element 30a is input with weighting factors added to the neuro elements 31a, 31b, 31c of the intermediate layer. Similarly, the output signals of the neuro elements 30b and 30c are each added with a weighting coefficient, and the neuro element 3 of the intermediate layer is added.
It is input to 1a, 31b and 31c. The outputs of the neuro elements 31a, 31b, 31c in the intermediate layer are input with a threshold value θ ₁ added to each neuro element 32 in the output layer with a weighting coefficient added thereto, and output at the threshold value θ ₂ of the neuro element 32. It

FIG. 10 shows a flow of a learning method of each weighting coefficient and threshold value of the neuro computer 13. In step 33, model data is input to the input layer. Therefore, in step 34, it is determined whether or not the data is correctly output to the output layer. In the case of Yes, it progresses to step 35 and the weighting coefficient of a correct solution path is enlarged. Similarly, reduce the threshold value.
In step 37, the difference from the correct answer is determined, and if the error amount is small, the process ends. If the error amount is large, the process returns to step 34, the weighting coefficient and the threshold value are adjusted again, and the process is repeated until the error amount becomes small. By learning about several types of model data by the above method, the accuracy rate increases.

An example of the configuration of the neurocomputer when the knock sensor 8 is of the wide-range resonance type will be shown. Assuming that the knocking frequencies resonate at 7.9 KHz, 13.8 KHz and 18.5 KHz, the sample period is

[0032]

[Equation 4]

The sample period is

[0034]

[Equation 5]

Or

[0036]

## EQU00006 ## 5.4 .mu.s.times.30 = 162 .mu.s (6) Input layer 30 intermediate layers 3 output layers 1 total 34 elements. FIG. 11 shows the configuration. The number of input layers is 30, and the output of each neuro element is input to the three intermediate layers.
The three middle layers are input to one output layer.

The configuration of FIG. 11 is an embodiment of the present invention, and the detection accuracy is improved by increasing the number of elements in the input layer and the intermediate layer.

12 to 14 show the neuro element 23 of FIG.
It shows the engineering structure of. FIG. 12 shows an operational amplifier. 13 shows the case of a transconductance amplifier, and FIG. 14 shows the case of an inverter. In either case, v _i ⁺ , v _i ⁻ : input v _O ⁺ , v _O ⁻ : output.

FIGS. 15 and 16 show the construction of the transistors of the operating pair of the input stage in the case of FIGS. Figure 1
5 is for bipolar, and FIG. 16 is for CMOS.

17 to 19 show a method for engineeringly realizing the weighting factor of the neuro element 23. 17 shows the case where the triode region of the FET is used, and FIG. 18 shows the case where the transconductance amplifier is used. FIG. 19 shows the case where a 4-bit capacitor array is used.

FIG. 20 shows a method for storing the learned weighting factors. D / A converter 41 built in the memory 40 of the computer in the control circuit
To convert to analog and operate the FET.

FIG. 21 shows another embodiment. This is a case where a resonance type is used for the knock sensor 8. The outputs obtained by converting the time series signals of the knock sensor 8 into spatial multivariable signals by the sample hold circuit are designated as 12a to 12f. This signal 12a
.About.12f is input to a neuro computer composed of an input layer and an output layer. Here, the same number of neuro elements are used in the input layer and the output layer. The sampling period is set to 1/4 of the resonance frequency, and the sampling time is also the same as the resonance frequency. Therefore, the signal of 43a has no resonance frequency component and becomes a noise component. Therefore, if the original signal input to the input layer and the signals of 43a to 43f are subtracted, the noise signal is removed and the signal of the resonance frequency remains. 44a in circuit 42
The difference between the maximum value and the minimum value of the signals of ~ 44f is calculated and used as the knocking detection signal.

FIG. 22 shows the signals of the respective parts, (a) is an input signal, (b) is a noise signal, and (c) is a resonance frequency signal. In this embodiment, the sampling period is set to 1 of the resonance frequency.
Although it has been set to / 4, the measurement accuracy is improved by reducing the sampling period.

FIG. 23 shows the configuration of an engine control device using a neural network. The target value is input to the neural controller, and the result is output to the engine 51 that is the control target. The state quantity from the engine 51
Feedback to the neural controller 50.
In this device, in addition to this, error feedback 52
There is a routine by. This routine detects the difference between the target value and the actual state quantity, and when the difference is large, changes the weighting coefficient w of the neural net portion of the neural controller 50 or the conversion function of the neuro element.

The embodiment shown in FIG. 23 is a system for controlling the air-fuel ratio of the engine. The target value r of the air-fuel ratio is given from the controller 53. In addition, other state quantities such as the water temperature T _w , the rotation speed N, and the air quantity (load) Q _q are input to the neural controller 50. The output of the neural controller 50 is the valve opening time T of the injection valve. This T is given to the engine 51, and as a result, the actual air-fuel ratio (A / F) λ is output. This λ
Is input to the neural controller 50 and also to the error feedback unit 52. The target value r is also input to the error feedback unit 52. Here, the deviation (error) between λ and r is detected. If this error is large, the configuration of the neural controller 50 is determined to be inappropriate and is reconfigured. This reconstruction is executed by rewriting the weighting coefficient of the neural network or rewriting the conversion function (sigmoid function or the like).

The structure of the neural network is shown in FIG. It is an input layer λ, an intermediate layer j, and an output layer k. The output O _{j of the} middle layer is

[0047]

[Equation 7]

It becomes This f is a conversion function, for example, a sigmoid function. The output T is

[0049]

[Equation 8]

It becomes

The operation state is shown in FIG. Figure 25 (a)
Is a state of function conversion in the intermediate layer j. θ
_j ⁽¹⁾ is the state in the first neuro element of the intermediate layer. The sum of the inputs to this element is

[0052]

[Outer 1]

The output of this function is subjected to a non-linear function conversion to become θ _j ⁽¹⁾ . The outputs of the following intermediate layers are also determined in the same manner, and the outputs are θ _j ⁽²⁾ , θ _j ⁽³⁾ , ... Next, the sum of the outputs of these middle layers is shown in FIG.
It becomes an input to the element of the output layer shown in (b). Here, the function conversion is similarly performed to obtain the output T. This T is the engine 5
Given to 1.

Here, for example, when it is desired to determine T in consideration of the transportation delay of the supplied fuel during the transient operation of the engine, the function of FIG. 25 (a) may be determined to a desired value.
The neural controller 50 can be considered as a model of the internal state of the engine 51. Therefore, as will be described later, if the reconstruction by learning is repeated,
Phenomena that are difficult to describe in mathematical formulas such as fuel transportation delay will be compensated.

The state of reconstruction of the neural network is shown. FIG. 26A shows a method of rewriting the weighting factor of the neuro element. Before learning, in the state of (a), the output is θ
_j ⁽¹⁾ . Here, the weight is rewritten by learning,
When W _1i ′, the state of (b) is reached and the output is
θ _j ⁽¹⁾ ′. In this way, by rewriting the weighting coefficient, an output different from that before can be obtained.

FIG. 26 (b) shows another method. Here, the conversion function is a method of rewriting from (a) to (b) before learning. This way, the input

[0057]

[Outside 2]

Even if the same as the above, the output is from θ _j ⁽¹⁾ to O _j ⁽¹⁾ '
It changes like. In other words, it is possible to learn by modifying the shape of the function.

The learning flow chart of the neural controller 50 is shown in FIG. This learning is activated at every sampling period of λ. In step 54, the target value r is read. In step 55, λ is read and then the error variable e is obtained. e becomes the formula (9).

[0060]

## EQU9 ## e = k (λ-r) (9) Based on this e, the correction amount ΔT of T is calculated in step 57. Then, in step 58, the correction program is started.

FIG. 28 shows a modification program. When this program is started, the current ΔT _(n) and the previous ΔT _(n-1) are read in steps 591 and 592, and the difference ΔE between them is obtained in step 593. The magnitude of this ΔE is determined in steps 594 to 597 by the reference values T _{1 to} T ₄ , and step 598 is performed according to the magnitude of each.
At 602, the weight group W is finely modified.

FIG. 29 shows another modification program. Here, steps 591 to 597 are the same, and in steps 603 to 607, according to the magnitude of ΔE, the function f ₁ to
Fix f ₅ .

[0063] Figure 28, selection of the weight group W _a to _W-e shown in FIG. 29, the selection of the function f ₁ ~f ₅ is previously programmed, whether rewriting it has been determined theoretically . FIG. 30 shows a method of storing the correction value. Figure 30
As shown in (a), when the modification program is completed,
Immediately store the modified value (weight or function) in battery-backable memory. This is done to preserve the learning effect. This modified value is shown in FIG.
As shown in (b), the memory is read immediately after the key switch is turned on.

FIG. 31 shows the effect of learning. FIG. 6 is a diagram showing a change in air-fuel ratio λ with respect to accelerator opening θ _ac .
The variation of λ becomes smaller every time learning is performed in the order of (a) → (b) → (c).

As described above, the learning using the neural network is achieved by rewriting the weighting coefficient, so that a great effect can be obtained without explicitly knowing the engine model.

FIG. 32 shows an example of multi-variable control. In this embodiment, the air-fuel ratio A / F, the torque T, and the acceleration g are input as target values from the controller 53 to the neural controller 50. From the neural controller 50, the valve opening time T _inj of the injection valve and the ignition timing T
_ig , throttle disclosure θ _th , transmission shift position P _tr , and line pressure OP _l are output. These values are given to the automobile 60. Furthermore, the actual A / F _r , T _r , g _{r of the} automobile
Is measured and the error feedback unit 5 with the target value
Entered in 2. Here, as shown in FIGS. 27 to 30, the error and the correction amount are calculated, and the neural controller is reconfigured. In this way, the neural control can be applied to control of multiple variables.

FIG. 33 shows the structure of the neural network. In addition to the target values A / F, T, and g, various state quantities such as water temperature, intake hot water, rotation speed, and load are input to the input layer i for determination.

FIG. 34 shows the hardware structure. 61 is
It is a controller unit including a neural controller.

Inputs to the control unit 61 are an access opening θ _ac 62, an air amount signal 63, a torque signal 64, an acceleration signal 65, etc., and an output is a signal to an actuator 67 which electrically operates a throttle 66. 68, an ignition signal 70 to the ignition coil 69, a valve opening signal 72 to the injection valve 71, a shift position signal 74 to the transmission 73, and a hydraulic line pressure control signal OP (not shown) for controlling the transmission.
_{is l} . These variables are controlled by the neural controller.

FIG. 35 shows the target values A / F, T, g.
An example of the determination method of Here, since the driver's intention is determined to determine the torque T and the acceleration g of the vehicle, the accelerator opening θ _ac is used as an input. As shown in Fig. 35, θ _ac and θ
A / F by the change in θ _'ac of the _ac, T, retrieved from three types of map of g.

FIG. 36 shows another method for determining the target value. This method uses two neural controllers 50. Neural controller 50
In (A), the accelerator opening θ _ac , the brake depression angle θ _br , and the vehicle speed V are input, and the environment in which the vehicle is placed is determined. In other words, it is determined whether the environment is an urban area, a highway, a traffic jam, a bad road (cold, water, ice), an uphill road, a downhill road, etc. decide. This target value is input to the neural controller 50 (B) and controls each part of the automobile as described above.

That is, in the example of FIG. 32, the driver's intention is emphasized. FIG. 36 shows a control method for an automobile, in which the vehicle runs optimally according to the environment.

Next, another embodiment of the present invention is shown in FIG. It is composed of an automobile 115, an automobile controller 116, and an engine shaft torque detector 117. The engine shaft torque detection device 117 detects the shaft torque of the engine, and the vehicle control device 116 determines whether the target torque is reached. If the target torque is not the target torque, the parameter for changing the shaft torque of the engine 115 is changed to obtain the target torque. Control so that.

FIG. 37 (b) shows the control flow of FIG. 37 (a). The target torque map is stored in the memory 118. The memory 118 is, for example, a ROM or a RAM backed up by a battery. The target torque is, for example, a map of engine speed and load (see FIG. 38).
Instead of the load, the basic fuel injection pulse width, the intake pressure, the throttle valve opening, and the accelerator opening may be selected. The target torque is compared with the shaft torque of the engine by the comparison unit 119. If the target torque does not match the target torque, the parameter for changing the shaft torque is corrected by the torque correction unit 122 to change the shaft torque of the engine 115. Specifically, the torque correction method is performed by changing the throttle valve opening, the fuel amount, and the ignition timing.

The shaft torque of the engine is detected by the torque detecting means 121.

FIG. 39 shows a method of detecting the engine shaft torque. Four pistons 100 are connected to the crankshaft 103. In this example, it is a 4-cylinder engine. Sensors 101 and 102 for detecting the rotation angles of both ends of the crank shaft are attached to both ends of the crank shaft. The signals of the sensors 101 and 102 are processed by the shaft torque detecting means 108. As is well known, the engine shaft output can be detected by detecting the twist angle of the crankshaft.

FIG. 40 shows an example of the measurement result of the shaft torque of a single cylinder engine. Combustion pressure and inertial forces are combined to produce complex changes in shaft torque with crank angle. In a multi-cylinder engine, it is difficult to detect the axial torque because the combined axial torque between the cylinders is detected.

FIG. 41 shows an example of the sensor signal. The sensors 101 and 102 are, for example, gears and electromagnetic pickups. When the clan shaft rotates, Fig. 41 (a),
A signal as shown in (b) is obtained. When the phase difference between these two signals is obtained, a signal as shown in FIG. 41 (c) is obtained. This phase difference corresponds to the twist angle of the crankshaft. That is, the twist angle of the crack axis can be provided by the phase difference between the two sensor signals.

[0079]

[Equation 10] Te = f (Δθ _cr ) ... (10) If the twist angle Δθ is obtained as in the equation (10), the axial torque can be obtained. However, as described above, the axial torque is a function f (Δ
It is difficult to find θ _cr ). Especially at high rotation, the processing time is short, the torsion angle of the shaft is large, and it is difficult to obtain the torque. Therefore, an approximate model of f (Δθ _cr ) is used, but the accuracy is poor. Further, if the accuracy of the model is improved, it takes time to calculate the shaft torque, which is not practical.

FIG. 42 shows a torque detection method using a neuro computer. Phase differences of signals temporally captured in the input layer of the neurocomputer Δθ ₁ , Δθ ₂ , Δ
Input θ ₃ , ..., Δθ _n . In this case, the output layer outputs the shaft torque. As shown in FIG. 42 (a), the connection of each layer is uniform in the initial stage, and the shaft torque cannot be obtained with high accuracy. As shown in FIG. 42 (b), when the neurocomputer learns the correct axial torque, the connection of each layer is optimized, and the correct axial torque can be obtained. That is, the axial torque can be detected without previously _obtaining the relation f (Δθ _cr ) of the axial torque with respect to the phase difference by giving the axial computer to the neurocomputer for learning. Moreover, since the calculation of each signal is performed in parallel, the calculation of the shaft torque is completed in a short time.

FIG. 43 shows the configuration of the present invention. The signals θ _a and θ _b of the two sensors are input to the circuit 120, and the phase difference Δθ is obtained. The phase difference Δθ is input to the peak hold circuit 106, and the peak-held signal is input to the neuro computer 107. One peak hold circuit 106 is provided for each input layer of the neurocomputer. The signals input to the input layers of the neurocomputer are signals that are shifted in time series. The sampling period is in the range of 1 to 30 degrees of crank angle. In the present invention, in other configurations, a CPU 108 for performing another calculation, a ROM 109 for inputting a program, a battery backup RAM 110 for storing learning parameters of a neuro computer, a RAM 111 for executing a program, and an A for inputting analog data.
It is composed of a / D converter 121.

FIG. 44 shows another embodiment of the present invention. The phase difference signal is converted into a digital signal by the A / D converter 108. After that, the digital signal is stored in the RAM 112. RAM1
In 12, the A / D converted values are sequentially stored in the memory of the RAM in time series. The data stored in the RAM 112 is input to the neuro computer 107. The CPU 108 controls the time-series data acquisition into the RAM 112. In such a configuration, without using a peak hold circuit,
Since data can be imported, it can be made compact.
Further, since the signal has a digital value, the signal processing is easy.

FIG. 45 shows another embodiment of the present invention. RAM1
The data acquisition to 12 is different from the main CPU 108
(Smaller than the main CPU). For example, 8 bits are enough for 32 bits of the main CPU. In this case, since the data can be taken in without the intervention of the main CPU, the load of the main CPU such as interrupt processing can be greatly reduced.

FIG. 46 shows the basic operation of the neurocomputer.

[0085] x _1, x ₂ to the outside, ..., if there are k signals _{x k, S = {x 1} , x 2, ..., x k} and expressed.

Information S For example, while contacting the shaft twist angle,
The neurocomputer internally modifies the parameters to fit the structure of S. This automatically forms a model of S inside the neurocomputer and uses it to create S
Processing information. The nerve field F that receives a signal from the outside has a spatially uniform structure, and the mutual result system is the mutual inhibition type. That is, 1 in the place
When one neuron emits a signal, the output of this neuron also tries to generate a signal for other neurons in the immediate vicinity, but acts to suppress the neurons at a distance.

When each neuron receives the input signal x, it receives it using its own joint weighted belt S,
Dot product as sum

Receive the stimulation of strength. In addition to this, an inhibitory signal x _O is received with a strength of −S _O x _O , and if another neuron outputs an output, it receives a mutual inhibition type action from that. If the sum of all these inputs exceeds a certain value, this neuron will output.

U _i is the average potential of the i-th neuron,
Let Z _i be its output, W _{ij be} the strength of the connection from the j-th neuron to the i-th neuron, and S _{i be the} connection weight for the external signal x of this neuron. Let S _{Oi be the} load on the suppression input x _O. If the input signal x is input, the excited state of the neuron is

It becomes

If a neuron is excited now, the excitatory coupling S _Oi is also increased in proportion to the strength of the input at that time.

[0092]

[Equation 11]

[0093]

[Equation 12]

That is, if the constants C and C'are properly selected, it is numerically proved that the neuron connection corresponding to S can be constructed in the neurocomputer.

From the above, for example, if the axial torque obtained by another method is given in advance while the engine is operating and W _ij of the neurocomputer is optimized, the axial torque T _e can be obtained from the phase difference Δθ _cr. It will be possible. W learned
_ij is stored in the battery backup memory.

FIG. 47 shows another embodiment of the present invention. The automobile 146 is provided with a front light emitting element 141 and a light receiving element 142. As the light emitting element, for example, a laser or an infrared diode is used. As the light receiving element, for example, a phototransistor or a photodiode is used. Light emitting element 141
The light generated by the light is reflected and scattered on the road surface, and the light receiving element 1
Reach 42.

As shown in FIG. 48, the light receiving element 144 has two
A plurality are arranged in a dimensional array. Light receiving element 14
4 signals are introduced to the input layer of the neuro computer 107.

As shown in FIG. 49, the signal strength varies depending on the position of each light receiving element. Road conditions (dry, rain,
The signal strength pattern differs depending on the size and shape of the obstacle (freezing, etc.). That is, by processing these signals in the neurocomputer as described above,
The shape of the obstacle and the road surface condition can be detected at high speed.

For learning of the neuro computer, the road surface condition is given in advance to optimize the weighting coefficient and the threshold value in the neuro computer. As described above, the presence / absence of the obstacle and the road surface condition can be accurately grasped, so that the driver can be warned or the driving parameter can be optimized.

FIG. 50 shows another method of detecting the road surface condition. The laser light emitted from the front of the vehicle is applied to the road surface, and the reflected light is Fourier transformed by the action of the lens and the filter.
By examining the shape and intensity distribution of light, it is possible to determine the road surface condition in front of the track, the inclination of the road, and the curvature.

The above is based on the fact that the light from the light-emitting element is reflected and scattered on the road surface is detected by the light-receiving element. For example, if a high-sensitivity CCD array is used as the light-receiving element, only the light from the outside is detected. can do.

FIG. 51 shows another embodiment. Adder 200
Signal 198 (eg knocking signal) and signal 19
9 (for example, the relative twist angle before and after the crankshaft) is input. The signal 199 is the angle signal 19 at the front end of the crankshaft.
6, the angle signal 197 of the crankshaft rear end (ring gear) is subtracted by the subtractor 195 (relative angle difference). Therefore, the signal 199 becomes the twist angle of the crankshaft, and a signal proportional to the engine output is generated. The output 201 of the adder 200 is input to the sample hold circuit 210, the time series signal is converted into a spatial multivariable sequence, and the sequence is input to the input layer of the neurocomputer 220. As a result, a knocking signal and an output 2 are output to the output 221 of the neuro computer 220.
An engine output signal is generated at 22. When the neurocomputer 220 is trained, each signal 19
6, the signal 197 and the signal 198 may be input, and the combined signal may be used as in the measurement.

In this embodiment, two outputs are detected with three signals, but two or more outputs can be obtained with two or three or more inputs.

FIG. 52 shows an embodiment for measuring the degree of deterioration of the catalyst 230. The O ₂ sensor 231 on the upstream side of the catalyst 230 and the O ₂ sensor 232 on the downstream side are input to the subtractor 233, the output 237 thereof is input to the sample hold circuit 234, and the neuro computer 235 calculates the catalyst deterioration degree output 236.

FIG. 53 shows another embodiment for detecting the degree of catalyst deterioration. As an input of the neuro computer 240, a signal 246 from an O ₂ sensor 231 on the upstream side of the catalyst, a signal 247 from a downstream temperature sensor 245 on the catalyst, and an engine speed signal 24
2. Using the throttle opening signal 244, the catalyst deterioration degree signal 241 is obtained. Although the throttle opening signal 244 is used here, the intake air amount signal of the engine or the intake pipe pressure may be used.

FIG. 54 shows another embodiment of the present invention. The neural network 300 includes the number of revolutions, knock signal, water temperature, oil temperature,
Vehicle information such as air volume, gear position, vehicle speed, throttle valve opening, suspension damping rate, and vehicle vibration are input. Further, the neural network 300 has a final output, and if a difference between this final output and the comfort index (which the driver freely sets) occurs, the weighting factors W ₁ , W ₂ , W ₃ are changed, and the final output and the comfort level are changed. The degree index matches. Meanwhile, intermediate outputs 301 to 305
Is branched into two directions a and b, and 301a is input to the final stage of the neural network. Further, 301b is input as a target value of the air-fuel ratio control system. Further, the intermediate output 302b is a target value of the ignition timing control system, 303b is a target value of the transmission control system, 304b is a target value of the suspension control system, and 305.
b is used as a target value of the throttle valve control system. As described above, when the driver inputs the comfort index that suits his / her preference, the target value of the control system of each engine is changed, and the drivability suitable for the taste is obtained.

FIG. 55 shows another embodiment, which can be used as a comfort index detector. The rotational speed, knock signal, water temperature, oil temperature, air amount, gear position, vehicle speed, throttle valve opening, suspension damping rate, and vehicle body vibration are input to the neural network 300.
The output of the neural network 300 is learned by changing the weighting coefficient using the comfort index as teacher data. When the learning is completed, the output 310 becomes an output representing the comfort index. If another driver shows a comfort index suitable for himself as teacher data, the weight coefficient is changed again as a generator of a comfort index suitable for the driver.

[0108]

According to the present invention, by inputting the output target value of the sensor, the actuator is controlled and the desired control is achieved.
You can get the result .

[Brief description of drawings]

FIG. 1 is a schematic diagram of an embodiment of the present invention.

FIG. 2 is a control flow chart.

FIG. 3 is an output characteristic diagram of a knock sensor.

FIG. 4 is an output characteristic diagram of a knock sensor.

FIG. 5 is an output characteristic diagram of a knock sensor.

FIG. 6 is a diagram showing a coupling of neuro elements.

FIG. 7 is a diagram showing an excited state of a neuro element.

FIG. 8 is a diagram showing an excited state of a neuro element.

FIG. 9 is a diagram showing coupling of neuro elements.

FIG. 10 is a control flow chart.

FIG. 11 is a diagram showing the coupling of neuro elements.

FIG. 12 is a diagram showing a structural unit of a neuro element.

FIG. 13 is a diagram showing a structural unit of a neuro element.

FIG. 14 is a diagram showing a structural unit of a neuro element.

FIG. 15 is a diagram showing a structural unit of a neuro element.

FIG. 16 is a diagram showing a structural unit of a neuro element.

FIG. 17 is a diagram showing a structural unit of a neuro element.

FIG. 18 is a diagram showing a structural unit of a neuro element.

FIG. 19 is a diagram showing a structural unit of a neuro element.

FIG. 20 is a diagram showing a structural unit of a neuro element.

FIG. 21 is a diagram showing the coupling of neuro elements.

FIG. 22 is a diagram showing an excited state of a neuro element.

FIG. 23 is a schematic diagram of an example of the present invention.

FIG. 24 is a view showing the coupling of neuro elements.

FIG. 25 is a diagram showing an excited state of a neuro element.

FIG. 26 is a diagram showing an excited state of a neuro element.

FIG. 27 is a control flow chart.

FIG. 28 is a control flow chart.

FIG. 29 is a control flow chart.

FIG. 30 is a control flow chart.

FIG. 31 is a view showing a change in air-fuel ratio with respect to opening of a throttle valve.

FIG. 32 is a schematic diagram of an example of the present invention.

FIG. 33 is a view showing a coupling of neuro elements.

FIG. 34 is a schematic diagram of an example of the present invention.

FIG. 35 is an air-fuel ratio determined by the opening and opening speed of the throttle valve,
The figure which shows the map which memorized torque and acceleration.

FIG. 36 is a schematic diagram of an example of the present invention.

FIG. 37 is a schematic view of an example of the present invention.

FIG. 38 is a diagram showing a map in which a torque determined from a load and a rotation speed is stored in a memory.

FIG. 39 is a diagram showing a torque detection device.

FIG. 40 is a diagram showing torque characteristics.

FIG. 41 is a diagram showing torque characteristics.

FIG. 42 is a view showing a coupling of neuro elements.

FIG. 43 is a schematic view of an embodiment of the present invention.

FIG. 44 is a schematic view of an example of the present invention.

FIG. 45 is a schematic view of an example of the present invention.

FIG. 46 is a view showing the coupling of neuro elements.

FIG. 47 is a configuration diagram for detecting a road surface condition.

FIG. 48 is a diagram showing the detection principle.

FIG. 49 is a signal characteristic diagram.

FIG. 50 is a diagram showing another detection method.

FIG. 51 is a schematic view of an example of the present invention.

52 is a schematic view of an example of the present invention. FIG.

FIG. 53 is a schematic view of an example of the present invention.

FIG. 54 is a schematic view of an example of the present invention.

FIG. 55 is a schematic view of an example of the present invention.

[Explanation of symbols]

2 ... Control circuit, 11 ... Sample and hold circuit, 13 ... Neurocomputer.

─────────────────────────────────────────────────── ─── Continued Front Page (51) Int.Cl. ⁷ Identification Code FI G05B 13/02 G05B 13/02 L 15/02 15/02 A 21/02 21/02 A G06F 15/18 550 G06F 15/18 550E (72) Inventor Toshiharu Nogi 4026 Kuji Town, Hitachi City, Hitachi, Ibaraki Prefecture Hitachi, Ltd. Hitachi Research Laboratory (72) Inventor Yoshishige Oyama 4026 Kuji Town, Hitachi City, Ibaraki Prefecture Hitachi Laboratory, Hitachi Research Institute (56) Reference Documents JP-A-2-37015 (JP, A) JP-A-2-33655 (JP, A) JP-A-3-96636 (JP, A) JP-A-2-287860 (JP, A)

Claims (4)

(57) [Claims] An actuator [1 claim], the output of the first sensor for detecting a condition amount which changes by the operation of the actuator, and the output of the second sensor for detecting another state quantity of the state quantity in which the certain, A neurocomputer for inputting the output target value of the first sensor into an input layer;
A control device which outputs a control signal to the actuator based on a signal from an output layer of the neurocomputer. 2. The control device according to claim 1, further comprising changing means for changing weighting of the neurocomputer so that a difference between the first sensor output and the target value is reduced. 3. The control device according to claim 1, further comprising changing means for changing a conversion function of the neurocomputer so that a difference between the first sensor output and the target value is reduced. . 4. The method according to any one of claims 1 to 3, A control device characterized by being used for controlling an automobile.