JPH06286630A - Road surface frictional coefficient estimating device - Google Patents

Abstract

PURPOSE:To estimate the road surface frictional. coefficient which can not be measured directly, with high precision, by detecting the vehicle traveling state and inputting the state to a neural network, as for a road surface frictional coefficient estimating device during the traveling of a vehicle. CONSTITUTION:Into a controller 2, each information representing the traveling state is supplied from a steering angle sensor 21, car speed sensor 22, yaw rate sensor 23, lateral slip angle sensor 24, and a rear wheel steering angle sensor 25. The controller reads these information in each prescribed interruption timing by using a microcomputer and executes the calculation processing based on the neural network, and calculates the road surface frictional efficient. The neural network allows plural numerical values to be outputted in parallel, and consists of the equal number of elements to a plurality of outputs, and learning is carried out, having the numerical series set in the pattern corresponding to the well-know road surface frictional coefficient, as learning signal, and calculation processing is carried out.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device for estimating a friction coefficient of a road surface of a vehicle while the vehicle is traveling, and more particularly, to a vehicle in which information about a traveling state of the vehicle which is relatively easy to detect by using a neural network. Based on this, it is possible to estimate the road surface friction coefficient, which is difficult to measure directly during running.

[0002]

2. Description of the Related Art Conventionally, 4WS (four-wheel steering system), AB
There are various technologies such as S (anti-lock brake system), TCS (traction control system), 4WD (four-wheel drive) that enhance the steering stability of the vehicle, and these technologies are actually used. It is installed in the vehicle and contributes greatly to the improvement of steering stability.

However, none of the conventional devices for enhancing the steering stability as described above measures or estimates the friction coefficient of the road surface itself and controls it in consideration of the result. . This is because it is very difficult to directly measure the friction coefficient of the road surface while the vehicle is traveling. Therefore, control is performed assuming that the friction coefficient of the traveling road surface is a constant value, but naturally the friction coefficient of the actual traveling road surface is not a constant value. Therefore, the rear wheel steering angle control in 4WS and the driving force distribution control in 4WD must be performed so as not to impair the steering stability of the vehicle regardless of the value of the friction coefficient. As a result, it is the current situation that the capabilities possessed by these technologies that enhance steering stability are not fully utilized.

[0004] Although there are some types which change the control contents by indirectly judging the slip state of the road surface from the wheel speed difference between the front and rear wheels, raindrop sensors, etc. (for example, Japanese Patent Laid-Open No. 1-959).
No. 68, etc.), the accuracy of detecting the slipping condition of the road surface is not so high because the judgment is made from the wheel speed difference, the raindrop sensor and the like. Further, although a technique for reading the roughness of a road surface using an ultrasonic sensor or the like has been studied, it has not reached a practical level.

On the other hand, there is a technique of estimating a difficult-to-measure parameter on the basis of a vehicle parameter that is easy to measure (see Japanese Patent Laid-Open No. 4-138970). Was using a neural network for. That is, the conventional technique disclosed in the above publication is to estimate a vehicle parameter by using a neural network based on other vehicle parameters. Specifically, it is disclosed in the vehicle. A technology that estimates the sideslip angle and yaw rate that may occur in a vehicle based on the longitudinal acceleration, lateral acceleration, vertical acceleration, steering torque, front wheel steering angle, vehicle speed, rear wheel steering angle, etc. that are occurring. However, in any case, only the technique of estimating another vehicle parameter based on the vehicle parameter is disclosed.

[0006]

However, the sideslip angle and yaw rate are currently parameters that can be easily measured directly by using an inexpensive gyro, and the need for using a neural network to estimate them is diminished. ing. In other words, it can be said that there is currently no need to use a neural network to estimate the vehicle parameters, which are the vehicle behavior itself. Rather, the vehicle behavior based on such vehicle behavior that is easy to detect is used. It is desirable to estimate the numerical value, etc., which has a great influence on the steering stability of the vehicle but which is difficult to measure directly, using a neural network.

The present invention has been made from such a point of view, and provides an apparatus for estimating the friction coefficient of a road surface, which has a great influence on the steering stability of a vehicle, based on information that is actually easy to measure. Is intended.

[0008]

In order to achieve the above object, a road surface friction coefficient estimating device according to the invention of claim 1 is a vehicle running state detecting means for detecting a running state of a vehicle, and this vehicle running state. And a neural network that outputs the road friction coefficient as an input and the traveling state of the vehicle detected by the detection means as an input.

According to a second aspect of the present invention, there is provided a road surface friction coefficient estimating device according to the first aspect of the invention, wherein the neural network uses a neural network for outputting a plurality of numerical values in parallel. The learning is performed by using, as a teacher signal, a number sequence that is composed of the same number of elements as a plurality of outputs and is set in a pattern according to a known road surface friction coefficient.

Further, in the road surface friction coefficient estimating device according to the invention of claim 3, in the invention of claim 1 or 2, the neural network uses a neural network having an autoregressive model, and An online learning means for learning the neural network by using the output of the autoregressive model is provided.

[0011]

According to the first aspect of the invention, the vehicle running state detecting means has the running state of the vehicle (for example, the operating state such as the steering angle and the throttle opening, the vehicle behavior state such as the longitudinal acceleration and the yaw rate). Are detected, they are input to the neural network. Then, the neural network performs an operation according to the connection state of each neuron, the weighting coefficient for the input, etc., and the estimation result of the road surface friction coefficient is output from the output layer of the neural network.

The running state of the vehicle used as input is
It must be changed under the influence of the road surface friction coefficient, but most of the detectable running states of the vehicle are affected by the road surface friction coefficient. The output of the sensor used for the control (for example, rear wheel steering control, driving force distribution control, etc.) may be used. Also, the neural network is
Before being mounted on a vehicle, it is necessary to complete learning based on many cases using such a traveling state of the vehicle as an input and a known road surface friction coefficient as a teacher signal.

According to the second aspect of the present invention, since the neural network outputs a plurality of (here, n) numerical values in parallel, the respective values of the outputs are O _i (i = 1). , ..., n), T _j (j = 1, ..., N) is used as a teacher signal used in learning of the neural network.
n), the t _ji each element of the teacher signal T _j, teacher signal T _j
If the known road surface friction coefficient corresponding to the above is μ _j ^* , learning is performed according to an algorithm such as back propagation so that the output O _i matches the corresponding element t _ji of the teacher signal T _j. System identification will be performed.

When a road surface friction coefficient is estimated by a neural network, one road surface friction coefficient is represented by a plurality of outputs O _i , so that one output represents a road surface friction coefficient. Highly accurate estimation is performed. Further, in the invention according to claim 3,
Since the neural network having the autoregressive model is used and the online learning means is provided, the road friction coefficient can be estimated with high accuracy even during the transient response.

[0015]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a schematic configuration diagram of a vehicle according to a first embodiment of the present invention. In this embodiment, a road surface friction coefficient estimating device according to the present invention is applied to a vehicle having a 4WS function.

First, the structure will be described. In FIG.
The front wheels 1L, 1R are connected to the outer ends of the tie rods 3L, 3R via knuckles (not shown).
The inner ends of L and 3R are connected to the rack shaft 4a of the rack and pinion type steering device 4. A pinion shaft (not shown) of the rack-and-pinion steering device 4 and a steering wheel 6 are connected via a steering shaft 5 so that rotation can be transmitted.
That is, the front wheels 1L and 1R are steered to the left or right when the operator steers the steering wheel 6.

On the other hand, the rear wheels 2L and 2R are connected to the outer ends of the axles 10L and 10R which rotate by receiving the driving force distributed by a differential gear box (not shown) so as to be able to transmit the rotational force. Therefore, this vehicle has rear wheels 2L, 2
It is a rear-wheel drive vehicle in which R is the drive wheel. Further, this vehicle has a rear wheel steering device 12 driven by the rotational force of a motor 11, and outer ends of tie rods 13L and 13R for rear wheel steering, the inner ends of which are connected to the rear wheel steering device 12. Is connected to the rear wheels 2L and 2R. Therefore, the rear wheel 2
By controlling the current supplied to the motor 11, the L and 2R are steered in any direction and with any size.

A motor driver 15 is connected to the motor 11, and the motor driver 15 and a relay switch 16a of a battery 16 for supplying electric power to the motor driver 15 are connected to each other.
It is controlled by a controller 20 including a microcomputer and necessary interface circuits. The controller 20 calculates the target rear wheel steering angle δ _r ^* that improves the steering stability of the vehicle by executing the arithmetic processing described later, and the actual rear steering angle that is the actual steering angle of the rear wheels 2L, 2R. The wheel steering angle δ _r is the target rear wheel steering angle δ _r
Control signal C to the motor driver 15 so that it matches ^*.
It outputs S to drive the motor 11, and the controller 20 is supplied with various signals necessary for such control.

In this embodiment, the controller 20 includes
For example, a steering angle sensor 21 configured of a known potentiometer for detecting a steering angle of the steering wheel 6, a vehicle speed sensor 22 configured for example of a tachometer attached to an output shaft of a transmission for detecting a vehicle speed, for example, a known gyro. A yaw rate sensor 23 configured to detect a yaw rate occurring in the vehicle, a sideslip angle sensor 24 configured to detect the sideslip angle of the vehicle configured by using, for example, a known gyro, and a rotation angle of a gear of the rear wheel steering device 12, for example. A rear-wheel steering angle sensor 25 for detecting the actual rear-wheel steering angle δ _r of the rear wheels 2L, 2R is connected thereto, and the steering angle θ, the vehicle speed v, and the yaw rate supplied from the respective sensors 21 to 25 are connected. Predetermined control is executed based on φ, sideslip angle β, and actual rear wheel steering angle δ _r .

FIG. 2 is a flow chart showing the outline of the processing relating to the rear wheel steering control executed in the controller 10. This processing is executed by an operating system (not shown) as interrupt processing at every predetermined cycle (for example, 5 msec). It The process flow will be described with reference to the flowchart. First, at step 101, the steering angle θ supplied from the steering angle sensor 21 and the vehicle speed sensor 22 are supplied.
The vehicle speed v supplied from the yaw rate sensor, the yaw rate φ supplied from the yaw rate sensor, the sideslip angle β supplied from the sideslip angle sensor 24, and the actual rear wheel steering angle δ _r supplied from the rear wheel steering angle sensor 25 are read.

Next, the routine proceeds to step 102, where the steering angle θ, vehicle speed v, yaw rate φ read in step 101 are read.
And the sideslip angle β as input, the arithmetic processing based on the neural network NN as shown in FIG. 3 is executed to calculate the road surface friction coefficient μ. Neural network NN
The contents of will be described in detail later. Then, the process proceeds to step 103, and the rear wheel steering angle determining coefficient C ₁ used in the next step 104 is set according to the following equation (1).

C ₁ = C ₀ · μ (1) The road surface friction coefficient μ calculated in step 102 is 0.0 ≦
It is a dimensionless coefficient that falls within the range of μ ≤ 1.0. It takes a value close to 1.0 in a slippery condition like a dry paved road, and in a slippery condition like a frozen road surface. 0.0
It takes a value close to. Therefore, the coefficient C ₁ for determining the rear wheel steering angle is a constant C ₀ when the traveling road surface is difficult to slip.
A value close to is taken, and a smaller value is taken as the road surface becomes slippery.

After the calculation of step 103 is performed,
The process proceeds to step 104, where the yaw rate φ and the front wheel steering angle δ when the vehicle is approximated by a two-wheel model and the sideslip angle β = 0
_The target yaw rate φ ^* is calculated according to the following equation (2) representing the relationship with _f . φ ^* = {G ₀ / (1 + τ ₀ s)} δ _f (2) The front wheel steering angle δ _f is calculated from the steering gear ratio N on the front wheels 1L and 1R side as δ _f = θ / N. . Further, G ₀ = C ₁ VL / (ML _r V ² + LL _f C ₁ ) τ ₀ = VI / (ML _r V ² + LL _f C ₁ ), and M, I, L _f , L _r , L (= L _f + L _r ) is a constant determined by vehicle specifications and the like, and s is a Laplace operator.

When the target yaw rate φ ^* is obtained, the routine proceeds to step 105, where the target rear wheel steering angle δ _r ^* is calculated according to the following equation (3). δ _r ^* = K _p (φ ^* −φ) (3) K _p is a proportional constant. Thus, the target yaw rate φ ^*
Actual yaw rate since it is the ^* target rear wheel steering angle [delta] _r is determined by multiplying the deviation in the proportional constant K _p and phi, the target yaw rate of the vehicle even when there is disturbance by increasing the proportional constant K _p When Although it is possible to make it close to φ ^* , the proportional constant K _p cannot be increased so much because the response delay of the motor 11 and the like actually exists. Here, it is assumed that K _p = −1 (sec).

Then, the routine proceeds to step 106, where a control signal CS for the motor driver 15 is determined so that the actual rear wheel steering angle δ _r matches the target rear wheel steering angle δ _r ^* , and this control signal CS is set to step 107. At the motor driver 15
Supply to. In step 107, if the relay switch 16a is off, a control signal for turning it on is also output. When the control signal is supplied to the motor driver 15, the interrupt processing this time is terminated, and at the next interrupt timing, the process returns to step 101 and the above-described processing is repeatedly executed.

Here, the neural network NN used to calculate the road surface friction coefficient μ in step 102.
Is a network in which the steering angle θ, the vehicle speed v, the yaw rate φ and the sideslip angle β are input and the road surface friction coefficient μ is output, as shown in FIG. 3. In this example, the input layer, the intermediate layer and the output Neural network NN with three-layer structure
Is used.

Each neuron nr constituting the neural network NN is a processing element that produces one output y for a prescribed number of inputs x _k , and produces an output y according to a prescribed transfer function. For example, if a weighting coefficient for each input to each neuron nr is w _k and a constant relating to its own state is ε, a transfer function as in the following equation (4) can be used.

[0028] The neural network NN needs to be learned before being mounted on the vehicle. Specifically, a known road surface friction coefficient μ ^* for a large number of patterns of steering angle θ, vehicle speed v, yaw rate φ and sideslip angle β is used as a teacher signal, and the known road surface friction coefficient μ ^* and the output layer neuron nr are output. The weight coefficient and the constant ε of each neuron nr are updated so that the error with the road surface friction coefficient μ is reduced. As such a learning algorithm, for example, known back propagation or the like is applied.

4 (b) to 4 (g) show a vehicle speed v = 100 km.
4 / h, the changes in yaw rate φ and sideslip angle β when the steering angle θ is changed in a sinusoidal manner as shown in FIG. 4 (a) are calculated using a plurality of road surface friction coefficients μ (= 1.0, 0.4, 0). .1) is a waveform diagram showing the result of the vehicle body motion simulation obtained for each. As can be seen from these waveform diagrams, the yaw rate φ and the sideslip angle β as responses to steering are affected by the road surface friction coefficient μ. It changes a lot. In addition,
Although not shown here, if the vehicle speed v is different, the yaw rate φ
It is also known that the waveform of the sideslip angle β also changes.

Therefore, it is sufficiently possible to estimate the road surface friction coefficient μ by using the neural network NN as shown in FIG. Next, the operation of this embodiment will be described. Since the microcomputer shown in FIG. 2 executes the process shown in FIG. 2 at each predetermined interrupt timing, the steering angle θ, the vehicle speed v, the yaw rate φ, the sideslip angle β, and the actual rear wheel steering angle, which are the output values of the sensors, are output. δ _r is read, and the steering angle θ, the vehicle speed v, the yaw rate φ and the sideslip angle β are input, and the arithmetic processing based on the neural network NN shown in FIG. 3 is executed to calculate the road surface friction coefficient μ. However, since the neural network NN inputs numerical values in different units and also calculates the road surface friction coefficient μ whose units are different from those inputs, the input values are made dimensionless (range of 0.0 to 1.0). The output value is obtained as a dimensionless value.

Table 1 shows the results of the road surface friction coefficient μ calculated by the neural network NN shown in FIG. 3 for various vehicle speeds v, and also shows the true values of the road surface friction coefficient μ. The neural network NN that obtained this result has three road surface friction coefficients μ (= 1.0,0.
4, 0.1) was learned.

[0032]

[Table 1]

As can be seen from the results shown in Table 1, by using the neural network NN, the road surface friction coefficient μ, which is very difficult to measure directly, can be estimated with a certain degree of accuracy. . In addition, since the neural network NN that obtained the results shown in Table 1 above is learning by using the response of the vehicle using the rigid body mass spring model, the estimation accuracy may be a little low. The target parameters are run on roads with different friction coefficients at different speeds and the parameters (here, steering angle θ, yaw rate φ, sideslip angle β) are recorded, and based on the recorded parameters. If the neural network NN is learned in this way, it is possible to perform estimation with higher accuracy. Further, here, learning is performed for three road surface friction coefficients μ (= 1.0, 0.4, 0.1), but if the types of teacher signals are further increased, the estimation accuracy of the neural network NN is estimated. Can be further increased.

Then, the microcomputer in the controller 10 calculates the target rear wheel steering angle δ _r ^* in consideration of the estimated road surface friction coefficient μ, and calculates the actual rear wheel steering angle δ.
_The control signal CS is output to the motor driver 15 so that _r matches the target rear wheel steering angle δ _r ^* . That is,
In the present embodiment, it is possible to estimate the road surface friction coefficient μ, which is practically impossible to measure directly, based on the steering angle θ, the vehicle speed v, the yaw rate φ and the sideslip angle β, which are easy to measure. Further, since the rear wheel steering angle control of the 4WS vehicle is performed by using the estimated road surface friction coefficient μ, the rear wheel steering angle control is performed without considering the road surface friction coefficient μ. The steering stability will be further improved compared to the conventional 4WS vehicle.

FIG. 5 shows a vehicle speed v = 120 km / on a dry road surface.
FIG. 3 is a waveform diagram showing changes in the yaw rate φ, the lateral velocity v _y, and the rear wheel steering angle δ _r when the steering angle θ is changed stepwise by 30 degrees while the vehicle is traveling straight at h. The response of the conventional 4WS vehicle without considering the road surface friction coefficient μ, and the waveform B shows the response of the normal 2WS vehicle. As described above, even in the conventional 4WS vehicle in which the rear wheels are steered without considering the road surface friction coefficient μ, the hunting of the yaw rate φ can be suppressed and the sideslip angle (= v _y
/ V) also approaches zero, so it can be said that it is easier to drive.

FIG. 6 shows a vehicle speed v = 120 on a road surface with snow.
FIG. 3 is a waveform diagram showing changes in the yaw rate φ, the lateral velocity v _y, and the rear wheel steering angle δ _r when the steering angle θ is changed stepwise by 30 degrees during straight traveling at km / h. A is the response of the conventional 4WS vehicle that does not consider the road surface friction coefficient μ,
The waveform B is the response of a normal 2WS vehicle, the waveform C is the response of a 4WS vehicle in which the road surface friction coefficient μ is detected by an ultrasonic sensor or the like, and the rear wheel steering control is performed based on the detected value, and the waveform D is of the present embodiment. 4 shows the response of a 4WS vehicle.

According to this, on a road surface with snow, the sideslip angle cannot be suppressed so much even with a conventional 4WS vehicle, whereas on a 4WS vehicle in which rear wheel steering control is performed in consideration of the road surface friction coefficient μ, sideslip The corners are kept small to make driving easier. In particular, the 4WS vehicle of the present embodiment can reduce the sideslip angle even on snowy road surfaces to the same extent as can be achieved by conventional 4WS vehicles on dry road surfaces. That is, with the configuration of this embodiment, good steering stability can be obtained regardless of the change in the road surface friction coefficient μ.

Here, in the present embodiment, the steering angle sensor 21, the vehicle speed sensor 22, the yaw rate sensor 23, and the sideslip angle sensor 24 correspond to the vehicle traveling state detecting means, and the steering angle θ, the vehicle speed v, Each of the yaw rate φ and the sideslip angle β corresponds to the running state of the vehicle. FIG. 7 is a diagram showing a second embodiment of the present invention. This embodiment is characterized in that a neural network NN based on a statistical mathematical model is used.

Specifically, the neural network NN
By providing three neurons nr in the output layer, the three outputs O ₁ , O ₂ and O ₃ are output in parallel. Also, this neural network NN needs to be appropriately learned before being mounted on the vehicle as in the case of the first embodiment, but the outputs of the neural network NN are O ₁ , O ₂ , O _3. Since it is obtained as three numerical values, it is necessary to perform learning corresponding to this.

Here, as shown in FIG. 7, three teacher signals T corresponding to the outputs of the neural network NN.
₁ , T ₂ , T ₃ are used, and each of them is
Known road friction coefficient μ ₁ ^* (= 1.0), μ ₂ ^* (= 0.
4), μ ₃ ^* (= 0.1), and each teacher signal T ₁ , T ₂ , T ₃ has three elements t _.1 , t _. , Corresponding to the number of outputs of the neural network NN _{. 2} , t _.3 (・
= 1, 2 or 3). In this embodiment, T ₁ = (1,0,0) T ₂ = (0,1,0) T ₃ = (0,0,1) and these teacher signals T ₁ , T ₂ , T ₃ Each element t of
Learning is performed by using a learning algorithm such as back propagation so that the error between _.1 , t _.2 , t _.3 and the outputs O ₁ , O ₂ , O ₃ becomes small.

Similar to Table 1, Table 2 shows the road surface friction coefficient μ calculated by the neural network NN shown in FIG.
6 is a table showing the results of the above for various vehicle speeds v.

[0042]

[Table 2]

However, since the output of the neural network NN is three numerical values of O _{1 to} O ₃ and the above-mentioned teacher signals T _{1 to} T ₃ are used for learning, the road surface friction coefficient μ is , Is calculated based on the following equation (5). Comparing Tables 1 and 2, it can be seen that the road surface friction coefficient μ can be estimated with higher accuracy by using the neural network NN based on the statistical mathematical model as in this embodiment. This is because three neurons nr are provided in the output layer so that one road surface friction coefficient μ is represented by _three outputs O _{1 to} O ₃ , so that the number of weighting factors that affect the estimation accuracy increases and the number of neurons increases. The main reason is that the bond relationship between nr has become complicated.

That is, in the present embodiment, the road surface friction coefficient μ, which is practically impossible to directly measure, is calculated as follows:
Since it is possible to perform estimation with higher accuracy than in the first embodiment, if the rear wheel steering angle control of the 4WS vehicle is performed using the estimated road surface friction coefficient μ, steering stability is further improved than in the first embodiment. Will improve the sex. Other functions and effects are the same as those in the first embodiment, and the description thereof will be omitted.

In this embodiment, the neural network NN has three outputs, but the number of outputs is arbitrary and can be increased or decreased depending on the computing capability of the microcomputer. The more the number of outputs of the neural network NN, the more accurately the road friction coefficient μ
Therefore, it is preferable to increase the number of outputs as much as possible within the range of the computing power of the microcomputer.

FIGS. 8 and 9 are views showing a third embodiment of the present invention. This embodiment uses a neural network NN having an ARMA (Auto Regressive Moving Average) model which is one of autoregressive models. It is characterized by the fact that it was there. Here, as shown in FIG. 4 described in the first embodiment, the characteristic of the response of the yaw rate φ and the sideslip angle β appears when the road surface friction coefficient μ is different, but immediately after the steering state is changed. , The yaw rate φ and the sideslip angle β are in a transient response state, and the transient response lasts about 1 to 4 seconds although it depends on the road surface friction coefficient μ, so after the steering situation changes, the yaw rate φ and the sideslip angle β Until the β response becomes a steady response, the estimation accuracy of the road friction coefficient μ may be extremely reduced.

In other words, there is no particular problem in a traveling condition in which the same steering state continues for a relatively long time, such as when traveling on an expressway. However, steering is performed as in the case of traveling in a normal city. In situations where the state changes frequently,
Since the estimation accuracy of the road surface friction coefficient μ may be low,
Even if the road surface friction coefficient μ is estimated, it cannot be effectively used for rear wheel steering control or the like.

On the other hand, in this embodiment, as shown in FIG. 7, inputs I ₁ (t), ..., I _m (t) (these inputs are, for example, as in the first and second embodiments described above). , Steering angle θ,
The vehicle speed v, the yaw rate φ, and the sideslip angle β may be used, or other sensor outputs may be used as in other embodiments described later. ) With its time delay value I ₁ (t
−1), ..., I ₁ (t−k), ..., I _m (t−1),
..., and supplies I _m a (t-k) to the input layer of the neural network NN, further, the output O ₁ of the neural network NN (t), ..., O _n (t) is the time lag calculating section 31
And to the input layer of the neural network NN via the ARMA model 30.

An example of the basic concept of the ARMA model 30 is shown in FIG. The example of FIG. 9 is composed of one main memory 30a and two auxiliary memories 30b and 30c, and the relationship between the input x (k) and the output y (k + 1) is within the ARMA model. The weighting factors of α _i (t), β
_j (t), γ _k (t), main memory 30a and auxiliary memory 30
Assuming that the numbers of memories of b and 30c are n, l and m, the following equation (6) is obtained.

[0050] Then, the output y (t + 1) of the ARMA model 30 and the input x (t +) to the ARMA model 30 in the next process.
1) error e (t + 1) {= x (t + 1) -y (t +
1)} becomes small, the weighting factors α _i (t), β _j (t), γ _k (t) and the number of memories n, 1, and m are appropriately updated by using the least squares method. The processing is executed together with the processing shown in FIG. 2 described in the first embodiment.

That is, in this embodiment, not only the current input value as in the first and second embodiments but also the past input and output are multiplied by the corresponding weighting factors and taken into the neural network NN. Since the output of the ARMA model 30 is used to update the weighting coefficient and the number of memories in the ARMA model 30 forming a part of the neural network NN (that is, the online learning function),
The road surface friction coefficient μ can be estimated with high accuracy not only during steady response, but also during transient response such as immediately after the steering state changes.

Therefore, even when driving in a normal city, the rear wheel steering control can be performed by incorporating the road surface friction coefficient μ, so that the overall steering stability can be improved. become. Incidentally, a neural network N having an autoregressive model as in this embodiment
In order to obtain the same effect without using N, the sensor input is decomposed into a number of frequency components, the neural network provided for each frequency component performs an operation, and the output of each neural network is calculated as a time axis. Although it has to be configured to calculate the road surface friction coefficient μ by returning it to the upper part, in order to obtain sufficient accuracy, it is necessary to decompose the sensor input into many order components (for example, 100th order or more). Practically impossible.

Here, in the present embodiment, the online learning means is constituted by the processing of updating the weighting coefficient and the number of memories using the above-mentioned least squares method. In the present embodiment, the case where the ARMA model is applied as the autoregressive model has been described, but another autoregressive model, for example, ARMAX (ARMA + ARX (Auto Regressive
Exogenous)) model or extended Kalman filter may be used.

FIG. 10 is a diagram showing a fourth embodiment of the present invention. In this embodiment, three sensor outputs of steering angle θ, vehicle speed v and yaw rate φ are used as inputs to the neural network NN. Is. That is, the first to the first
In the third embodiment, the steering angle θ, the vehicle speed v, the yaw rate φ, and the sideslip angle β are input to the neural network NN, whereas in the present embodiment, the sideslip angle β is omitted. , Because the sideslip angle sensor is unnecessary,
An inexpensive configuration can be made. That is, there is an advantage that it can be easily applied even to a vehicle that does not have a sideslip angle sensor. In addition, the configuration of the present embodiment can be realized at low cost because the necessary sensors are already provided especially for a 4WS vehicle.

FIG. 11 is a diagram showing a fifth embodiment of the present invention, in which the longitudinal acceleration Xg is added as an input to the neural network NN to the configuration of the fourth embodiment.
That is, the yaw rate φ generated in the vehicle body is the road surface friction coefficient μ
However, since it also greatly changes under the influence of the braking force, the longitudinal acceleration Xg changes depending on the braking force.
Is input to the neural network NN to improve the estimation accuracy of the road surface friction coefficient μ.

For example, when the steering wheel is steered while a light brake operation is performed while traveling on a dry road surface having a road surface friction coefficient μ close to 1.0, a large yaw rate is generated.
In the same situation, a strong brake operation produces a slightly lower yaw rate. Therefore, if the longitudinal acceleration Xg is also input to the neural network NN as in this embodiment, the road surface friction coefficient μ can be estimated with higher accuracy. In addition, the configuration of the present embodiment can be realized at low cost because the necessary sensors are already provided especially for a 4WS vehicle.

FIG. 12 is a diagram showing a sixth embodiment of the present invention, in which the actual rear wheel steering angle δ _r is added as an input to the neural network NN in the configuration of the fifth embodiment.
That is, in the case of a 4WS vehicle, the yaw rate generated also changes depending on the rear wheel steering angle. Therefore, if the actual rear wheel steering angle δ _r is also input to the neural network NN, the road surface friction coefficient μ will become even more accurate. Will be able to estimate.

FIG. 13 is a diagram showing a seventh embodiment of the present invention, in which the wheel rotation speed ω, the brake pressure P and the brake torque T _B are input to the neural network NN. That is, if the road surface friction coefficient μ is different, the wheel rotation speed ω and the brake torque T _{B with} respect to the same brake pressure P.
Therefore, the road surface friction coefficient μ can be estimated even with the configuration of this embodiment.

In particular, the configuration of this embodiment is advantageous in that the road surface friction coefficient μ can be estimated if the brake operation is performed even when the vehicle is traveling straight. Further, the configuration of the present embodiment can be realized at low cost because the necessary sensors are already provided in a vehicle having an ABS. FIG. 14 is a diagram showing an eighth embodiment of the present invention, in which the wheel rotational speed ω, the engine rotational speed N _E , the throttle opening κ and the drive torque T _E are input to the neural network NN.

With such a configuration, the road surface friction coefficient μ can be estimated from the change in the wheel rotation speed ω when the accelerator pedal is depressed, so that the road surface friction coefficient μ can be estimated even when accelerating during straight running. You will be able to. In addition, the configuration of the present embodiment can be realized at low cost because the necessary sensors are already provided in a vehicle having a TCS.

FIG. 15 is a diagram showing a ninth embodiment of the present invention, in which the steering angle θ and the hydraulic pressure P _S of the power steering device are input to the neural network NN. That is, in a vehicle having a hydraulic power steering device, if the road surface friction coefficient μ is different, the power steering oil pressure P _S generated with respect to the steering angle θ is different. It is possible to estimate the coefficient μ.

With the structure of this embodiment, 4 W
Even a vehicle that does not have a special function such as S, ABS, or TCS can be realized at low cost. FIG. 16 is a diagram showing a tenth embodiment of the present invention, in which the vehicle speed v is added to the configuration of the ninth embodiment as an input to the neural network NN. With such a configuration, there are advantages similar to those of the ninth embodiment, and even the vehicle speed-sensitive hydraulic power steering device can estimate the road surface friction coefficient μ with high accuracy.

FIG. 17 is a diagram showing an eleventh embodiment of the present invention, in which a plurality of neural networks NN having the characteristics described in each of the above embodiments are provided and the outputs of these neural networks NN are provided. The road friction coefficient μ is set to another neural network NN.
The input is ₀ , and the neural network NN ₀ tries to further calculate the road surface friction coefficient μ ₀ .

With such a structure, the road surface friction coefficient μ _{0 is} calculated based on the road surface friction coefficient μ obtained in each of the above embodiments.
Is required, the estimation accuracy will be dramatically improved. FIG. 18 is a diagram showing a twelfth embodiment of the present invention. This embodiment integrates the functions of the fourth to eleventh embodiments.

That is, all the sensor outputs required for the above fourth to eleventh embodiments can be read, and the sensor outputs can be supplied to the neural network NN through the data loader 35 and the switch 36 in a required combination. It is what Then, for example, when the vehicle is functioning as a 4WS vehicle and steering is being performed, the sensor output necessary for functioning as the fourth, fifth or sixth embodiment is supplied to the neural network NN, For example, during braking, the sensor output necessary for it to function as the seventh embodiment is supplied to the neural network NN, and during acceleration, for example, the sensor output necessary for functioning as the eighth embodiment is supplied to the neural network NN. Particularly, when highly accurate estimation of the road surface friction coefficient μ is required, all sensor outputs are supplied to the neural network NN so as to function as the eleventh embodiment. In the case of the present embodiment, the number of inputs of the neural network NN varies depending on the selected function, but "0" is input to the input layer in the non-supply state, and learning is appropriately performed corresponding to that. You just have to go.

With such a configuration, the advantages of each of the above-described embodiments can be utilized, so it is possible to estimate the road surface friction coefficient μ and effectively use it for various controls in any situation. become able to. In each of the above embodiments, the case where the present invention is applied to a 4WS vehicle has been described, but the vehicle to which the present invention can be applied is not limited to this, and various functions such as 4WD, ABS, and TCS can be applied. The vehicle may be provided, or it may be applied to engine control, power steering control, or the like.

The running state of the vehicle as an input to the neural network NN is not limited to the one described in each of the above-mentioned embodiments. In short, the running state of the vehicle changes due to the influence of the road surface friction coefficient and detection is not possible. Anything is possible as long as it is possible.

[0068]

As described above, according to the present invention,
Since the vehicle running state is detected and the road surface friction coefficient is calculated by using this as input to the neural network, it is possible to accurately estimate the road surface friction coefficient, which is difficult to measure directly, based on the parameters that can be detected directly. The effect is that you can do it.

In particular, the invention according to claim 2 has an effect that the road surface friction coefficient can be estimated with higher accuracy. Furthermore, the invention according to claim 3 has an effect that the road surface friction coefficient can be estimated with high accuracy even during a transient response.

[Brief description of drawings]

FIG. 1 is a conceptual diagram showing a configuration of a vehicle in a first embodiment of the present invention.

FIG. 2 is a flowchart showing an outline of processing executed in a controller.

FIG. 3 is a conceptual diagram showing a configuration of a neural network in the first embodiment.

FIG. 4 is a waveform diagram showing an example of response of yaw rate and sideslip angle to steering.

FIG. 5 is a waveform diagram illustrating a control effect of a conventional 4WS vehicle.

FIG. 6 is a waveform diagram illustrating the effect of the present embodiment.

FIG. 7 is a conceptual diagram showing a configuration of a neural network in a second embodiment.

FIG. 8 is a conceptual diagram showing a configuration of a neural network in a third embodiment.

FIG. 9 is a conceptual diagram showing an example of an ARMA model.

FIG. 10 is a configuration diagram of a fourth embodiment.

FIG. 11 is a configuration diagram of a fifth embodiment.

FIG. 12 is a configuration diagram of a sixth embodiment.

FIG. 13 is a configuration diagram of a seventh embodiment.

FIG. 14 is a configuration diagram of an eighth embodiment.

FIG. 15 is a configuration diagram of a ninth embodiment.

FIG. 16 is a configuration diagram of a tenth embodiment.

FIG. 17 is a configuration diagram of an eleventh embodiment.

FIG. 18 is a configuration diagram of a twelfth embodiment.

[Explanation of symbols]

 20 Controller 21 Steering Angle Sensor 22 Vehicle Speed Sensor 23 Yaw Rate Sensor 24 Sideslip Angle Sensor 25 Rear Wheel Steering Angle Sensor NN Neural Network

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁵ Identification code Office reference number FI technical display location B62D 113: 00 123: 00 127: 00 137: 00

Claims (3)

[Claims] 1. A vehicle running state detecting means for detecting a running state of a vehicle, and a neural network for inputting a running state of a vehicle detected by the vehicle running state detecting means and outputting a road surface friction coefficient. An apparatus for estimating a road surface friction coefficient. 2. The neural network uses a neural network that outputs a plurality of numerical values in parallel, and is composed of the same number of elements as the plurality of outputs and is set in a pattern according to a known road surface friction coefficient. The road surface friction coefficient estimating device according to claim 1, wherein learning is performed using a sequence of numbers as a teacher signal. 3. The neural network having an autoregressive model is used as the neural network, and online learning means for learning the neural network using the output of the autoregressive model is provided. Road surface friction coefficient estimation device.