JP3123430B2 - Activation control device for occupant protection device, product-sum operation circuit learning method and device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an occupant protection device such as an airbag device for protecting an occupant in a vehicle when the vehicle collides, and more particularly to an occupant protection device for controlling activation of such an occupant protection device. The present invention relates to a start-up control device, a product-sum operation circuit learning method for determining, by learning, an operation constant of each product-sum operation element in a product-sum operation circuit used in the start control device, and an apparatus therefor.

[0002]

2. Description of the Related Art As a device for controlling activation of an occupant protection device, there is, for example, a device for controlling ignition of a squib in an airbag device. In an airbag device, a gas generating agent is ignited by a squib in an inflator, a gas is generated from the inflator, and the gas is used to inflate a bag to protect an occupant from collision with indoor components.

A device for controlling the ignition of a squib of such an airbag device usually includes a deceleration sensor (G sensor) for detecting a deceleration applied to the vehicle (particularly, a deceleration in the longitudinal direction of the vehicle). The ignition control is performed based on the deceleration detection signal obtained by the G sensor. For example, the signal value of the deceleration detection signal or its integral value is compared with an appropriate threshold value, and when the signal value or the integral value exceeds the threshold value, control is performed to ignite by squib.

[0004]

However, in the case where the activation of the occupant protection device is controlled by comparing the signal value or the integral value of the deceleration detection signal with the threshold value in this way, the collision type of the vehicle (that is, the collision type). Depending on the mode, the vehicle speed at the time of collision, the collision object that is the collision partner, and the like), it is necessary to perform a complicated calculation to reliably determine whether or not to activate the occupant protection device. For example, in the collision mode of the vehicle, if the collision mode is other than a head-on collision (for example, an oblique collision, an offset collision, an underride collision, or a pole collision) and the vehicle speed at the time of the collision is relatively high, the occupant protection device is provided. Should start,
When the collision mode is a head-on collision and the vehicle speed at the time of the collision is relatively low, it is not sufficient to activate the occupant protection device. However, comparing the two cases with the deceleration detection signal obtained from the G sensor, there is not much difference between the two within a predetermined time (required time) from the occurrence of the vehicle collision. As described above, it is necessary to distinguish between the two by a complicated operation, and there is a problem that a cost is required due to the construction of an operation system and the like.

Accordingly, an object of the present invention is to solve the above-mentioned problems of the prior art, and to reliably determine whether or not to activate the occupant protection device regardless of the type of vehicle collision. To provide an activation control device.

[0006]

In order to achieve at least a part of the above object, a first aspect of the present invention is to provide a vehicle shock corresponding to the magnitude of an impact applied to a vehicle.
A signal obtained from a signal is input, and a plurality of
Performs predetermined operations including product-sum operation using operation constants
To obtain an activation control signal for controlling activation of the occupant protection device.
-Sum operation circuit for start-up control device that outputs signals for
In order to determine the plurality of arithmetic constants by learning,
A sum of product arithmetic circuit learning method, comprising: (a) a plurality of product-sum operation device, having a plurality of inputs
And adjusting the operation constants of the plurality of product-sum operation elements.
Preparing a learning product-sum operation circuit capable Rukoto, (b) a plurality of places from the previously prepared learning vehicle impact signal
Extracting a constant frequency component; and (c) extracting the plurality of extracted predetermined frequency components for learning.
The parallel inputs to the input terminals of the product-sum operation circuit
By the plurality of product-sum operation elements in the product-sum operation circuit,
Each of the predetermined operations including the product-sum operation using the operation constants
And (d) outputting from the learning product-sum operation circuit as an operation result.
Output signal corresponding to the learning vehicle impact signal
The error from the ideal output signal prepared in advance becomes a predetermined value
The plurality of products in the learning product-sum operation circuit.
Adjusting the operation constant of the sum operation element.
The operation constants of the plurality of product-sum operation elements obtained as a result of
The duplication in the product-sum operation circuit for the activation control device
The gist is that it is determined as a numerical operation constant.

[0007]

[0008]

[0009]

[0010]

[0011]

[0012]

[0013]

[0014]

In the first invention, first, a plurality of predetermined frequency components are extracted from the learning vehicle impact signal. Next, the plurality of predetermined frequency components are input in parallel to the input terminal of the learning product-sum operation circuit, and the predetermined operation including the product-sum operation is performed by the plurality of product-sum operation elements using respective operation constants. A plurality of sums of products in the learning sum-of-products arithmetic circuit are set such that an error between an output signal output from the learning sum-of-products arithmetic circuit as an operation result and an ideal output signal corresponding to the learning vehicle impact signal has a predetermined value. The arithmetic constant of the arithmetic element is adjusted, and the learning product-sum arithmetic circuit is made to learn. The operation constants obtained by the learning are determined as a plurality of operation constants in the product-sum operation circuit for the activation control device.

As described above, according to the first aspect, a plurality of predetermined frequency components are preliminarily extracted from the learning vehicle impact signal as a pre-process when learning is performed by the learning product-sum operation circuit. As described above, since the frequency component of the vehicle impact signal differs depending on the type of collision of the vehicle, the learning product-sum operation circuit learns using the predetermined frequency component extracted from the vehicle impact signal, so that the vehicle collision signal is obtained. Learning in a form that reflects the form can be performed. Therefore, the operation constant of each product-sum operation element in the learning product-sum operation circuit is:
By repeating the learning, the value converges to an optimal value, and finally, an optimal value such that an error between the output signal and the ideal output signal becomes a predetermined value can be obtained.

In the method for learning a product-sum operation circuit according to the first invention, it is preferable that the step (a) includes a step of preparing a neural network as the learning product-sum operation circuit.

By using a neural network as the learning product-sum operation circuit, learning of the product-sum operation circuit can be easily performed.

In the method for learning a product-sum operation circuit according to a first aspect of the present invention, in the step (b), a vehicle impact signal other than a vehicle impact signal obtained while the vehicle is traveling on a rough road is prepared as the learning vehicle impact signal. Preferably, the method includes a step.

Since the peak waveform of the vehicle impact signal on a rough road and the peak waveform of the vehicle impact signal at the time of a high-speed head-on collision are very similar to each other and it is difficult to discriminate between the two, the learning product-sum operation circuit learns the two. When making the learning vehicle impact signal,
If a vehicle impact signal during running on a rough road is also used, the calculation constant does not easily converge. Thus, by excluding the vehicle impact signal during traveling on a rough road and causing the learning product-sum operation circuit to learn, an optimum value can be derived as an operation constant in a short time.

According to a second aspect of the present invention, a signal obtained from a vehicle impact signal corresponding to the magnitude of an impact applied to a vehicle is inputted, and a predetermined operation including a sum of products operation is performed on the signal by using a plurality of operation constants. A product-sum operation circuit for the activation control device for outputting a signal for obtaining an activation control signal for controlling activation of the occupant protection device, the product-sum operation circuit for determining the plurality of operation constants by learning A learning device, comprising: frequency component extraction means for extracting a plurality of predetermined frequency components from a learning vehicle impact signal prepared in advance; and a plurality of multiply-accumulation elements, having a plurality of input terminals, and A plurality of predetermined frequency components are input in parallel from the input terminal, a predetermined operation including a product-sum operation is performed by the plurality of product-sum operation elements using respective operation constants, and an operation constant of the plurality of product-sum operation elements is provided. A learning product-sum calculation circuit that can be adjusted; an output signal output from the learning product-sum calculation circuit as a calculation result; and an ideal output signal prepared in advance corresponding to the learning vehicle impact signal. Operation constant adjusting means for adjusting the operation constants of the plurality of sum-of-products elements in the learning sum-of-products operation circuit so that the error has a predetermined value, wherein the plurality of products obtained as a result of the adjustment are provided. The gist of the invention is to determine an operation constant of the sum operation element as the plurality of operation constants in the product-sum operation circuit for the activation control device.

In the second invention, a plurality of predetermined frequency components are extracted from the learning vehicle impact signal by the frequency component extracting means. Then, the plurality of predetermined frequency components are input in parallel from an input terminal by a learning product-sum operation circuit, and a predetermined operation including a product-sum operation is performed by a plurality of product-sum operation elements using respective operation constants. At this time, the learning constant is adjusted by the calculation constant adjusting means so that the error between the output signal of the learning product-sum calculation circuit, which is the calculation result, and the ideal output signal corresponding to the learning vehicle impact signal becomes a predetermined value. The operation constant of each product-sum operation element in the sum operation circuit is adjusted, and the learning product-sum operation circuit learns. The operation constants obtained by the learning are determined as a plurality of operation constants in the product-sum operation circuit for the activation control device.

Therefore, according to the second aspect of the present invention, the learning product-sum operation circuit is learned by using the predetermined frequency component extracted from the vehicle impact signal, so that the learning in a form that reflects the collision mode of the vehicle is performed. In the end, an optimal value such that the error between the output signal and the ideal output signal becomes a predetermined value is calculated as the operation constant of each product-sum operation element in the learning product-sum operation circuit. Obtainable.

[0024] In the product-sum operation circuit learning apparatus according to the second invention, it is preferable that the learning product-sum operation circuit comprises a neural network.

By configuring the learning product-sum operation circuit with a neural network, learning of the product-sum operation circuit can be easily performed.

[0026] In the product-sum operation circuit learning apparatus according to the second invention, the learning vehicle impact signal prepared in advance is:
It is preferable that the signal is a vehicle impact signal other than the vehicle impact signal obtained when the vehicle is traveling on a rough road.

As described above, by excluding the vehicle impact signal during traveling on a rough road and causing the learning product-sum operation circuit to learn, the operation constant of each product-sum operation element in the learning product-sum operation circuit is determined. As a result, an optimum value can be derived.

According to a third aspect of the invention, the magnitude of the impact applied to the vehicle
Of occupant protection system based on vehicle impact signal corresponding to vehicle
Control of an occupant protection device that obtains a start control signal for controlling the vehicle
A plurality of predetermined frequencies from the vehicle impact signal.
Frequency component extracting means for extracting components, and a plurality of input terminals.
Having the plurality of extracted predetermined frequency components at the input terminal.
Input in parallel, and for the plurality of predetermined frequency components,
Includes multiply-accumulate operation using multiple set operation constants
A signal obtained by performing a predetermined operation and obtaining the result of the operation
And a product-sum operation circuit that outputs
When the activation control signal is obtained from the signal output from the
Are used as the plurality of operation constants in the product-sum operation circuit.
The product-sum operation circuit learning method according to the first invention
The gist is to use the determined operation constant.

[0029] In the first invention, it is possible to derive an optimum value as a processing constant for each processing element in the learning product-sum operation circuit, such optimum operation constant, and the
By using the product-sum operation circuit as a plurality of operation constants, it is possible to derive an optimal start control signal according to the vehicle impact signal.

According to a fourth aspect of the present invention, the magnitude of the impact applied to the vehicle
Of occupant protection system based on vehicle impact signal corresponding to vehicle
Control of an occupant protection device that obtains a start control signal for controlling the vehicle
A plurality of predetermined frequencies from the vehicle impact signal.
Frequency component extracting means for extracting components, and a plurality of input terminals.
Having the plurality of extracted predetermined frequency components at the input terminal.
Input in parallel, and for the plurality of predetermined frequency components,
Includes multiply-accumulate operation using multiple set operation constants
A signal obtained by performing a predetermined operation and obtaining the result of the operation
And a product-sum operation circuit that outputs
To determine whether it is a vehicle collision or running on a rough road.
Discriminating means for deriving a discriminating signal for calculating the product-sum operation circuit
The signal output from the vehicle should activate the occupant protection device.
And the information is derived by the determination means.
Said discrimination signal indicates a vehicle collision
In this case, the occupant protection device may be used as the activation control signal.
Starting control signal output means for outputting a starting signal.
And the plurality of operation constants in the product-sum operation circuit.
The product-sum operation circuit learning method according to the first invention
The gist is to use the determined operation constant. No.
The invention according to claim 5 is a vehicle corresponding to the magnitude of the impact applied to the vehicle.
Activation that controls activation of the occupant protection device based on the impact signal
An activation control device for an occupant protection device that obtains a control signal,
Extracting a plurality of predetermined frequency components from the vehicle impact signal
Frequency component extracting means, having a plurality of input terminals,
The plurality of predetermined frequency components are input in parallel from the input terminal.
And, for the plurality of predetermined frequency components, a preset
Using a plurality of operation constants, perform predetermined operations including product-sum operation
Product that outputs the signal obtained as the result of the operation
A vehicle collision based on the sum calculation circuit and the vehicle impact signal;
Signal for determining whether the vehicle is traveling on rough roads
Discriminating means for deriving
Said signal should trigger said occupant protection device
And the determination derived by the determination means.
If another signal indicates a vehicle collision,
A signal for activating the occupant protection device as a start control signal
Starting control signal output means for outputting
An arithmetic circuit configured to determine whether the value obtained by the learning is the plurality of arithmetic operations;
From a neural network set as a constant
And the plurality of operations in the product-sum operation circuit
As a constant, the product-sum operation circuit learning method according to claim 3 is used.
Therefore, the gist is to use the determined operation constant.
You.

[0031]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described below based on examples. FIG. 1 is a block diagram showing a configuration of an activation control device of an airbag device as one embodiment of the present invention.

As shown in FIG. 1, the start control device 40 of this embodiment mainly includes a deceleration sensor (G sensor) 41,
An analog / digital converter (A / D converter) 42, a frequency component extraction circuit 44, and a product-sum operation circuit 46 are provided. Then, the activation control device 40 includes a drive circuit 52
Is connected to the squib 51 in the airbag device 50 through the squib 51. These components are mounted at predetermined positions in the vehicle 32, respectively.

First, the G sensor 41 detects the deceleration applied to the vehicle 32 in the front-rear direction, and outputs the detected deceleration as an analog deceleration detection signal (G signal). The G signal thus obtained is a signal corresponding to the magnitude of the impact applied to the vehicle 32. The G sensor 41 may be a sensor that detects only deceleration, or may be a sensor that detects acceleration in a broad sense including deceleration.

Next, the A / D converter 42 receives the G signal output from the G sensor 41, converts the analog signal into a digital signal, and outputs the digital signal to the frequency component extracting circuit 44. When the G signal output from the G sensor 41 is a digital signal, the A / D converter 42 can be omitted.

Next, the frequency component extraction circuit 44 extracts a plurality of predetermined frequency components from the input G signal. Specifically, from the G signal, 50 Hz to 55
Each frequency component up to 0 Hz is extracted. Thus, G
By extracting a plurality of predetermined frequency components from the signal, even when a vehicle collision occurs, the collision mode can be further clarified. Since the configuration and function of the frequency component extraction circuit 44 are the same as those of the frequency component extraction circuit 24 in the product-sum operation circuit learning device 20 described later, a detailed description will be made instead of the description of the frequency component extraction circuit 24.

Next, the product-sum operation circuit 46 includes a plurality of product-sum operation elements (not shown), and has a single output terminal for a plurality of input terminals. The product-sum operation circuit 46 inputs a plurality of predetermined frequency components extracted by the frequency component extraction circuit 44 in parallel from an input terminal, and performs a product-sum operation using respective operation constants by each internal product-sum operation element. Then, a binary signal (“0”, “1”) obtained as a result of the calculation is output from the output terminal as a start control signal of the airbag device 50. The start control signal is
When the signal value is “1”, it indicates that the airbag device 50 should be activated, and when the signal value is “0”, it indicates that the airbag device 50 cannot be activated.

Next, the drive circuit 52 controls the activation control device 40
While the activation control signal output from is input and the signal value is “0”, the airbag device 50 is not activated. However, when the signal value becomes “1”, the airbag device 50 is activated. The squib 51 in the airbag device 50 is energized.

The airbag device 50 includes a squib 51, a gas generating agent (not shown), a bag (not shown), and the like. Therefore, when the squib 51 is energized by the drive circuit 52, the squib 51 ignites the gas generating agent, whereby gas is generated from the gas generating agent. The gas inflates the bag to protect the occupant from collision with interior components in the vehicle 32.

In this embodiment, in order to derive a start control signal from a plurality of predetermined frequency components extracted by the frequency component extraction circuit 44, a product-sum operation circuit 46 having a plurality of product-sum operation elements is used. . This product-sum operation circuit 4
The operation constant of each product-sum operation element in 6 is determined by learning using a neural network as described later so as to derive an optimal activation control signal.

FIG. 2 is a block diagram showing the configuration of a product-sum operation circuit learning device for determining the operation constant of the product-sum operation element in the product-sum operation circuit 46 of FIG. 1 by learning. FIG.
As shown in (1), the product-sum operation circuit learning device 20 includes an A / D converter 22, a frequency component extraction circuit 24, a neural network 26, and an adjustment circuit 28.

Of these, the A / D converter 22 is the one shown in FIG.
The A / D converter 42 has the same configuration and the same function as the A / D converter 42 shown in FIG. 1 and inputs a G signal for learning prepared in advance and converts the G signal from an analog signal to a digital signal. As the G signal for learning, a G signal that would be obtained at the time of a vehicle collision is used when the airbag device is to be started (in the case of the squib ignition ON requirement) and when it is not necessary to start the airbag device (squib ignition). (In the case of the OFF requirement), and further prepared for each type of collision. That is, for example, in each of the frontal collision, the oblique collision, the offset collision, the underride collision, and the pole collision, a G signal when the airbag device is to be activated and a G signal when the activation is insignificant are prepared. .

The frequency component extraction circuit 24 has the same configuration and the same function as the frequency component extraction circuit 44 shown in FIG. 1, and inputs the learning G signal converted by the A / D converter 22. Then, a plurality of predetermined frequency components are extracted from the G signal. Specifically, each frequency component from 50 Hz to 550 Hz is extracted from the G signal in steps of 50 Hz.

The neural network 26 has a plurality of neuron models, and has a single output terminal for a plurality of input terminals. The neural network 26 is substantially the same as the product-sum operation circuit 46 shown in FIG.
A plurality of predetermined frequency components extracted by the frequency component extraction circuit 24 are input in parallel as learning data from an input terminal,
The product-sum operation using the operation constants is performed by each internal neuron model, and binary signals (“0”, “1”) obtained as the operation result are output. However, in the neural network 26, it is possible to change and adjust the operation constant in each neuron model.

The adjustment circuit 28 receives the output signal output from the neural network 26, inputs the ideal output signal prepared in advance as teacher data, finds an error between both signals, and reduces the error to almost zero. The arithmetic constant of each neuron model in the neural network 26 is changed and adjusted so that the neural network 26 learns. Note that, as an ideal output signal, when a learning G signal is input to the A / D converter 22, it is an ideal binary signal that is output from the neural network 26 corresponding to the G signal. “0”, “1”) are prepared. More specifically, if the learning G signal is a signal when the airbag device is to be activated (in the case of a squib ignition ON requirement), a predetermined discrimination request time (when the airbag device is activated from the time of a vehicle collision). Until the elapse of the maximum time required for issuing a signal to be activated), a signal having a value “0” indicating that the airbag device is not activated is prepared as an ideal output signal, and the determination request time elapses. After that, a signal having a value “1” indicating that the airbag device should be activated is prepared as an ideal output signal. In addition, if the learning G signal is a signal that is not enough to activate the airbag device (in the case of the squib ignition off requirement), a value indicating that the airbag device is not activated as an ideal output signal. A signal of “0” is prepared.

FIG. 3 shows the frequency component extracting circuit 24 shown in FIG.
FIG. 2 is a configuration diagram showing each internal configuration of a neural network 26; As shown in FIG. 3, the frequency component extraction circuit 24
Has eleven frequency component extraction units E. These frequency component extraction units E1 to E11 extract each frequency component from 50 Hz to 550 Hz from the input G signal in increments of 50 Hz.

FIG. 4 is a block diagram showing a specific example of the frequency component extraction unit E shown in FIG. As shown in FIG. 4, each frequency component extraction unit E shown in FIG. 2 includes multiplication processing units 150 and 152, low-pass filter (LPF) processing units 154 and 156, square processing units 158 and 160, and an addition unit. A processing unit 162, and an LPF processing unit 164, and the G signal G output from the A / D converter 22.
From (t), a frequency component Gf ₀ ′ (t) including a preset frequency f ₀ is extracted.

First, in the multiplication processing section 150, the G signal G
(T) is a cosine function cos2πf₀multiplied by t and the product signal G
c (t) is obtained, and the multiplication processing unit 152 calculates the sine function sin
2πf ₀The product signal Gs (t) is obtained by multiplying by t. That is,
The product signals Gc (t) and Gs (t) are as shown in Expression (1).
You.

[0048] Gc (t) = G (t ) · cos2πf 0 t Gs (t) = G (t) · sin2πf 0 t ......... (1)

[0049] FIG. 5 is a waveform diagram showing G signal G (t) and the trigonometric function cos2πf ₀ t in FIG. 4, an example of the change over time in sin2πf ₀ t. That is, when the G signal G (t) as shown in FIG. 5A is multiplied by the cosine function cos2πf ₀ t of the frequency f ₀ as shown in FIG. The portion where the frequency and phase match is emphasized to the positive and negative sides, but the portion where the frequency or phase does not match is attenuated, so that the emphasized portion is obtained as the product signal Gc (t). become. Similarly, the G signal G (t) shown in FIG.
The sine function sin2πf _{0 of the} frequency f ₀ as shown in FIG.
When multiplied by t, the portion of the G signal whose frequency and phase match the sine function is emphasized to the positive and negative sides, but the portion whose frequency and phase do not match is attenuated, so that the product signal Gs
As (t), the emphasized portion is obtained.

Here, the phases of the cosine function cos2πf ₀ t and the sine function sin2πf ₀ t are different by π / 4 as shown in FIGS. 5 (b) and 5 (c). Therefore, even in a portion of the G signal that is out of phase with the cosine function (that is, a portion that is not emphasized in the multiplication processing section 150), it is in phase with the sine function, and the multiplication processing section 152 has a positive and negative side. Conversely, even if the phase is not in phase with the sine function (ie, the part is not emphasized in the multiplication unit 152), the phase is in phase with the cosine function, and At 150, the emphasis is on the positive and negative sides. Therefore, of the G signal, a portion where the frequency matches the cosine function and the sine function (that is, a portion corresponding to the frequency f ₀ ) is emphasized and obtained as either the product signal Gc (t) or Gs (t). .

Next, the LPF processing units 154 and 156
The product signals Gc (t) and Gs (t) are each subjected to LPF processing to obtain signals Gc ′ (t) and Gs ′ (t).
That is, the LPF processing units 154 and 156 respectively function as digital LPFs, and the product signals Gc (t) and Gs (t)
Of the signal Gc ′ (t),
Gs' (t) is derived. At this time, the cutoff frequency f _{C of the} LPF is set to a frequency higher than the frequency f ₀ .

Here, the reason why the LPF processing is performed is as follows. That is, in the subsequent square processing units 158 and 160 and the addition processing unit 162, the sum of squares is obtained for the signal. However, when the sum of squares is directly obtained for the product signals Gc (t) and Gs (t), the following equation is obtained. 2), and the square of the G signal G (t) is simply obtained.

{Gc (t)} ² + {Gs (t)} ² = {G (t) · cos 2πf ₀ t} ² + {G (t) · sin 2πf ₀ t} ² = {G (t)} ²・ {Cos ² 2πf ₀ t + sin ² 2πf ₀ t} = {G (t)} ² (2)

Next, in the square processing units 158 and 160 and the addition processing unit 162, the product signal Gc ′ subjected to the LPF processing is output.
(T), by adding Gs' a (t) in terms of the square, respectively, calculates the square sum to obtain a signal Gf ₀ (t). That is, the signal Gf ₀ (t) is as shown in Expression (3).

Gf ₀ (t) = {Gc ′ (t)} ² + {Gs ′ (t)} ² (3)

As described above, the product signal Gc (t) is
A portion of the signal whose frequency and phase match the cosine function is a signal emphasized on the positive and negative sides, and the product signal Gs (t)
Is a signal in which the portion where the frequency and phase match the sine function are emphasized on the positive and negative sides. Therefore, each product signal Gc
By squaring (t) and Gs (t) (actually, the product signals Gc ′ (t) and Gs ′ (t) subjected to the LPF processing), all the parts emphasized on the negative side are positive. Move to the side,
Further, by calculating the sum thereof, a portion where the frequency matches the cosine function and the sine function regardless of the phase (that is,
The portion corresponding to the frequency f ₀ ) is all emphasized on the positive side. Therefore, the signal Gf obtained as a sum of squares
_{It can be said that 0} (t) is a portion corresponding to the frequency f ₀ in the G signal G (t), that is, a frequency component of the frequency f ₀ .

Next, in the LPF processing section 164, the frequency f
LPF processing is performed on the signal Gf ₀ (t), which is a frequency component of ₀ , to remove unnecessary components, thereby obtaining a signal Gf ₀ ′ (t).
When unnecessary components are unnecessary, such as when there are few unnecessary components,
The LPF processing section 164 may be omitted.

As described above, the frequency component extraction unit E
Extracts a frequency component Gf ₀ ′ (t) including a preset frequency f ₀ from the G signal G (t). That is, the 50 Hz extraction unit E1 shown in FIG.
The 100 Hz extraction unit E2 extracts a 100 Hz frequency component included in the G signal, and the... 100 Hz extraction unit E11 extracts a 550 Hz frequency component included in the G signal.

Incidentally, a neural network (neural network) is an information processing system developed by imitating a neural network having a multi-layered structure formed by a large number of nerve cells in the brain. As shown in FIG. 3, the neural network 26 used in the present embodiment includes an input layer,
It is divided into three layers, an intermediate layer and an output layer, and each layer is constituted by a neuron model 30 simulating nerve cells in the brain.

FIG. 6 is an explanatory diagram for explaining the neuron model 30 shown in FIG. 3 and the sigmoid function fs (X) used in the neuron model 30. FIG.
As shown in (a), the neuron model 30 is modeled as a weighted multi-input / one-output system, and is considerably simplified as compared with actual nerve cells in the brain.

The neuron model 30 first receives input signals x1 and x1 from the other i neuron models.
2, ..., xi are received. Then, as shown in equation (4), the input signals x1, x2,..., Xi are multiplied by weight coefficients w1, w2,. The state quantity X is obtained.

[0062]

(Equation 1)

Next, this state quantity X is converted by a transfer function f (X) as shown in equation (5) to obtain an output signal y. y = f (X) (5)

Here, the transfer function f (X) is expressed as shown in equation (6). f (X) = fs (X) + α (6) In equation (6), fs (X) is a sigmoid function, and α is one of the operation constants and is a bias value.

The sigmoid function fs (X) is a function as shown in equation (7), and changes with respect to the state quantity X as shown in FIG. fs (X) = 1 / {1 + exp (−X)} (7)

In the neural network 26,
The weighting coefficients w1, w2,..., Wi and the bias value α, which are operation constants, can be changed within a predetermined range.

In the neural network 26 shown in FIG. 3, the input layer is constituted by eleven neuron models 30. Each neuron model 30 receives each frequency component extraction unit E1 in the frequency component extraction circuit 24 as an input signal. The frequency components extracted by E11 are input. Also, the intermediate layer is about 30 at the initial stage.
Although it is composed of zero neuron models 30, it is eliminated by learning and finally composed of about 100 neuron models 30. The output layer is constituted by one neuron model 30,
An output signal is output to the adjustment circuit 28. Among the neuron models 30, the neuron model whose output line branches into two or more outputs the same output signal y to each output line.

On the other hand, as described above, the adjustment circuit 28
An error between the output signal output from the neural network 26 and an ideal output signal prepared in advance is obtained, and the neural network 26 is set so that the error becomes zero.
Are adjusted by changing the operation constants (ie, weighting factors w1, w2,..., Wi and the bias value α) of each neuron model 30.
Let 6 learn. Such learning of the neural network 26 is sequentially performed on all the G signals for learning prepared in advance. That is, learning is performed on the G signal when the airbag device is to be activated and the G signal when it is not necessary to activate the airbag device in each of the frontal collision, the oblique collision, the offset collision, the underride collision, and the pole collision. Then, when all learning is completed, each neuron model 30 in the neural network 26
Are determined respectively.

In this manner, the product-sum operation circuit learning device 20
The calculation constant of each neuron model 30 determined by the learning by is used as a calculation constant of each product-sum operation element in the product-sum operation circuit 46 shown in FIG.

That is, in the product-sum operation circuit 46 shown in FIG. 1, each product-sum operation element is arranged so as to have the same arrangement as each neuron model 30 in the neural network 26 shown in FIG. Each of the product-sum operation elements is configured to perform the same operation as the neuron model 30 shown in FIG. At this time, the operation constants (that is, the weighting factors w1, w2,..., Wi and the bias value α) of each of the product-sum operation elements are described above in the neuron model arranged at the same position as the product-sum operation element. Is set to the same value as the operation constant determined by learning.

By doing so, the product-sum operation circuit 46 has the same configuration and function as the neural network 26 after the learning is completed.

Therefore, as shown in FIG.
When a predetermined frequency component extracted from the G signal actually obtained by the above is input to the product-sum operation circuit 46, at the time of a vehicle collision, a signal close to an ideal output signal according to the mode of the collision is activated from the product-sum operation circuit 46. It can be output as a signal. Therefore, it is possible to reliably determine whether or not to activate the airbag device 50 regardless of the type of collision of the vehicle.

In the present embodiment, as described above, only the G signal that would be obtained at the time of a vehicle collision is prepared as the learning G signal, and when the vehicle is traveling on a rough road ( In this case, the G signal that would be obtained when the airbag device does not need to be activated) is excluded. The reason for excluding the G signal during traveling on a rough road is as follows. In other words, the peak waveform of the G signal during traveling on a rough road is very similar to the peak waveform of the G signal at the beginning of collision at the time of a high-speed head-on collision, and it is difficult to discriminate the G signal. This is because when the signal is input as a signal and the neural network 26 learns, the error between the output signal and the ideal output signal may not easily converge to zero.

However, when the learning of the neural network 26 is performed by excluding the G signal on a rough road as described above, each neuron model 3 finally obtained is obtained.
The calculation constant of 0 naturally does not include information about running on a rough road. Therefore, if these operation constants are used as the operation constants of the product-sum operation elements in the product-sum operation circuit 46 shown in FIG. 1, the operation constants of these product-sum operation elements also include information about traveling on a rough road. Therefore, in the activation control device 20, it may be difficult to reliably determine whether the vehicle is traveling on a rough road and a vehicle collision.

An embodiment in which a rough road determination circuit 62 is provided separately from the product-sum operation circuit 46 in order to reliably determine whether the vehicle is running on a rough road and a vehicle collision will be described below.

FIG. 7 is a block diagram showing the configuration of an activation control device as another embodiment of the present invention. In FIG.
The same components as those in FIG. 1 are denoted by the same reference numerals. Therefore, description of those components will be omitted.

The rough road determination circuit 62 which is a feature of the present embodiment.
Receives the G signal output from the A / D converter 42,
Based on the G signal, it is determined whether the vehicle is in a collision or traveling on a bad road, and the result of the determination is output as a determination signal. That is, when it is determined that the vehicle is a collision, a signal of a value “1” is output as a determination signal, and when it is determined that the vehicle is not a vehicle collision (including driving on a rough road), a value of the determination signal is output. The signal of "0" is output.

Further, an AND circuit 64 receives the output signal from the product-sum operation circuit 46 and the discrimination signal from the rough road judgment circuit 62, obtains the logical product of the two, and starts the operation result. Output as a signal. The activation control signal output at this time indicates that the airbag device 50 should be activated when the signal value is "1", and when the signal value is "0", as in the case of FIG. This indicates that it is not enough to activate the airbag device 50.

Therefore, for example, even if the value of the output signal from the product-sum operation circuit 46 becomes “1” during running on a rough road, indicating that the airbag device 50 should be started, the rough road determination is performed. Since the value of the determination signal from the circuit 62 remains "0", indicating that the vehicle is not in a vehicle collision (that is, traveling on a rough road), the value of the output signal from the AND circuit 64 is "0". To indicate that it is not necessary to activate the airbag device 50. Therefore, there is little possibility that the airbag device 50 will be activated during traveling on a rough road.

The rough road determination circuit 62 can be constituted by, for example, an integration circuit and a comparison circuit. Generally, the integrated value of the G signal is relatively large during a vehicle collision, but relatively small during traveling on a rough road. Therefore, the input G signal is integrated by an integrating circuit, and the obtained integrated value is compared with a predetermined threshold value by a comparing circuit. When the integrated value is larger than the threshold value, it is determined that the vehicle is in a collision, and a determination signal of a value “1” is output. And outputs a determination signal having a value “0”. The above-described integration circuit may be a section integration circuit that performs section integration for a predetermined time (for example, 10 ms).

The rough road determination circuit 62 is, for example, LP
It is also possible to configure with an F circuit and a comparison circuit.
In general, the G signal does not contain much high frequency components at the time of a vehicle collision, while the G signal contains many high frequency components during running on a rough road. Therefore, from the input G signal, the LPF
A high frequency component is removed by a circuit, and the value of the G signal is compared with a predetermined threshold value by a comparison circuit. When the value of the G signal is larger than the threshold value, it is determined that the vehicle is in a collision, and a determination signal of a value “1” is output. (Including during road driving)
And outputs a determination signal having a value “0”.

In the present embodiment, the G signal input to the rough road determination circuit 62 is a G signal corresponding to the deceleration applied to the vehicle 32 in the front-rear direction. The applied deceleration may be detected, and the G signal corresponding to the deceleration may be used.

As described above, according to the present embodiment, the provision of the rough road determination circuit 62 makes it possible to reliably determine whether the vehicle is traveling on a rough road or at the time of a vehicle collision. It is unlikely that the airbag device 50 will be activated.

The present invention is not limited to the examples and embodiments described above, but can be implemented in various modes without departing from the scope of the invention.

In the above-described embodiment, the description has been made on the assumption that a general product-sum operation circuit is used as the product-sum operation circuit 46 used for the activation control device 40. Since this is an arithmetic circuit and is a kind of the product-sum operation circuit, the product-sum operation circuit 46
It is possible to use as. In this case, the neural network 26 after learning used in the product-sum operation circuit learning device 20 may be used as it is, but another neural network is prepared, and an arithmetic constant obtained by learning is set in the neural network. It may be used.

In the above-described embodiment, as shown in FIG. 4, the frequency component extraction units E in the frequency component extraction circuits 24 and 44 are respectively replaced by multiplication processing units 150 and 152,
Although it is configured by the square processing units 158 and 160, the addition processing unit 162, etc., the frequency component extraction unit E may be configured by a neural network instead.

In the above embodiment, as shown in FIG. 4, the frequency component extraction circuits 24 and 44 perform processing by a relatively simple calculation using a multiplication processing unit, a square processing unit, an addition processing unit, and the like. However, if it is permissible that the processing becomes complicated and the configuration becomes complicated, it is also possible to use, for example, a fast Fourier transform circuit.

In the above-described embodiment, the G signal is directly used as the vehicle impact signal corresponding to the magnitude of the impact applied to the vehicle and is input to the frequency component extraction circuit. Not limited thereto, a signal obtained by performing a predetermined operation (for example, integration or differentiation) on the G signal is
The vehicle impact signal may be input to a frequency component extraction circuit.

In the above embodiment, the activation control device 4
The frequency component extraction circuit 44 and the product-sum operation circuit 46 in 0, or the frequency component extraction circuit 24 and the neural network 26 in the product-sum operation circuit learning device 20 can be configured by hardware, but according to a program or the like. It can also be realized by software using an operating CPU or the like.

In the above-described embodiment, the airbag device is used as the occupant protection device. However, in addition to the airbag device, for example, an auto pretensioner may be used. The auto pretensioner operates, for example, to wind up a seat belt with gas generated from an inflator and further enhance the function of the seat belt.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an activation control device of an airbag device as one embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a product-sum operation circuit learning device for determining, by learning, an operation constant of a product-sum operation element in the product-sum operation circuit 46 of FIG. 1;

3 is a configuration diagram showing each internal configuration of a frequency component extraction circuit 24 and a neural network 26 shown in FIG.

FIG. 4 is a block diagram illustrating a specific example of a frequency component extraction unit E illustrated in FIG. 2;

FIG. 5 shows a G signal G (t) and a trigonometric function cos in FIG.
FIG. 7 is a waveform chart showing an example of a temporal change of 2πf ₀ t and sin2πf ₀ t.

6 shows a neuron model 30 shown in FIG. 3 and a sigmoid function fs used in the neuron model 30.
It is explanatory drawing for demonstrating (X).

FIG. 7 is a block diagram showing a configuration of an activation control device as another embodiment of the present invention.

[Explanation of symbols]

 Reference Signs List 20: Product-sum operation circuit learning device 22: A / D converter 24: Frequency component extraction circuit 26: Neural network 28: Adjustment circuit 30: Neuron model 32: Vehicle 40: Activation controller 41: G sensor 42: A / D Converter 44 frequency component extraction circuit 46 product-sum operation circuit 50 airbag device 51 squib 52 drive circuit 62 rough road determination circuit 64 AND circuit 150, 152 multiplication processing unit 154, 156 LPF processing unit 158, 160: square processing unit 162: addition processing unit 164: LPF processing unit

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-4-224439 (JP, A) JP-A-6-92199 (JP, A) JP-A-6-263000 (JP, A) (58) Field (Int.Cl. ⁷ , DB name) B60R 21/16-21/32

Claims (9)

(57) [Claims] 1. A signal obtained from a vehicle impact signal corresponding to the magnitude of an impact applied to a vehicle is input, and a predetermined operation including a product-sum operation is performed on the signal by using a plurality of operation constants. A product-sum operation circuit learning method for determining a plurality of operation constants by learning, for a product-sum operation circuit for an activation control device that outputs a signal for obtaining an activation control signal for controlling activation of an occupant protection device. (A) preparing a learning product-sum operation circuit having a plurality of product-sum operation elements, having a plurality of input terminals, and capable of adjusting the operation constants of the plurality of product-sum operation elements; (B) extracting a plurality of predetermined frequency components from a learning vehicle impact signal prepared in advance; and (c) paralleling the extracted plurality of predetermined frequency components to input terminals of the learning product-sum operation circuit. Enter and learn By the plurality of product-sum operation elements in the product-sum operation circuit,
Performing a predetermined operation including a product-sum operation using respective operation constants; (d) an output signal output from the learning product-sum operation circuit as an operation result; Adjusting the operation constants of the plurality of sum-of-products elements in the learning sum-of-products circuit so that an error from the prepared ideal output signal becomes a predetermined value. Determining a calculation constant of the plurality of product-sum calculation elements as the plurality of calculation constants in the product-sum calculation circuit for the activation control device. 2. The product-sum operation circuit learning method according to claim 1 , wherein the step (a) includes a step of preparing a neural network as the learning product-sum operation circuit. Method. 3. A product-sum operation circuit learning method according to claim 1 or 2, wherein step (b), other than a vehicle impact signal obtained during rough road running of the vehicle as the learning vehicle impact signal And a product-sum operation circuit learning method comprising the step of preparing a vehicle impact signal. 4. A signal obtained from a vehicle impact signal corresponding to the magnitude of an impact applied to the vehicle is input, and a predetermined operation including a product-sum operation is performed on the signal by using a plurality of operation constants. A product-sum operation circuit learning device for determining, by learning, the plurality of operation constants, for a product-sum operation circuit for an activation control device that outputs a signal for obtaining an activation control signal for controlling activation of the occupant protection device. A frequency component extracting means for extracting a plurality of predetermined frequency components from a learning vehicle impact signal prepared in advance; and a plurality of sum-of-products computing elements, having a plurality of input terminals, and extracting the plurality of extracted predetermined frequencies. The components are input in parallel from the input terminal, a predetermined operation including a product-sum operation is performed by each of the plurality of product-sum operation elements using an operation constant, and an operation constant of the plurality of product-sum operation elements is adjusted. A learning sum-of-products arithmetic circuit capable of performing the following operations: an error between an output signal output from the learning sum-of-products arithmetic circuit as an operation result and an ideal output signal prepared in advance corresponding to the learning vehicle shock signal is determined. And a calculation constant adjusting means for adjusting calculation constants of the plurality of sum-of-products elements in the learning sum-of-products circuit so as to obtain a value of the plurality of sum-of-products calculation elements obtained as a result of adjustment. Wherein the arithmetic constant is determined as the plurality of arithmetic constants in the product-sum arithmetic circuit for the activation control device. 5. The learning device according to claim 4 , wherein the learning product-sum circuit comprises a neural network. 6. A product-sum calculation circuit learning device according to claim 4 or 5, is prepared in advance the learning vehicle impact signal, the vehicle other than a vehicle impact signal obtained during rough road running of the vehicle A product-sum operation circuit learning device, which is an impact signal. 7. A vehicle corresponding to the magnitude of an impact applied to the vehicle.
A start control that controls the start of the occupant protection device based on both impact signals.
The starting control device of the occupant protection device that obtains the dynamic control signal
Te, extracting a plurality of predetermined frequency components from the vehicle impact signal
Frequency component extracting means, and a plurality of input terminals , the plurality of extracted predetermined frequency components
Input from the input terminal in parallel, and
For a minute, using a plurality of preset operation constants,
Performs a predetermined operation including a product-sum operation, and calculates
And a product-sum operation circuit for outputting a signal obtained Te, the activation system from the signal output from the integrating sum calculation circuit
An activation control device for an occupant protection device that obtains a control signal and uses an operation constant determined by the product-sum operation circuit learning method according to claim 1 or 2 as the plurality of operation constants in the product-sum operation circuit. 8. A vehicle corresponding to the magnitude of an impact applied to the vehicle.
A start control that controls the start of the occupant protection device based on both impact signals.
The starting control device of the occupant protection device that obtains the dynamic control signal
Te, extracting a plurality of predetermined frequency components from the vehicle impact signal
Frequency component extracting means, and a plurality of input terminals , the plurality of extracted predetermined frequency components
Input from the input terminal in parallel, and
For a minute, using a plurality of preset operation constants,
Performs a predetermined operation including a product-sum operation, and calculates
A product-sum operation circuit for outputting a signal obtained from the vehicle, and based on the vehicle impact signal,
Discrimination that derives a discrimination signal for discriminating whether the line is running
Means, and the signal output from the product-sum operation circuit is
Protection device should be activated, and
The discrimination signal derived by the means is a vehicle collision
The start control signal,
A start control signal for outputting a signal for starting the occupant protection device
And an output unit, as the plurality of operation constants in the product-sum operation circuit,
An activation control device for an occupant protection device using an operation constant determined by the product-sum operation circuit learning method according to claim 3 . 9. A vehicle corresponding to the magnitude of an impact applied to the vehicle.
A start control that controls the start of the occupant protection device based on both impact signals.
The starting control device of the occupant protection device that obtains the dynamic control signal
Te, extracting a plurality of predetermined frequency components from the vehicle impact signal
Frequency component extracting means, and a plurality of input terminals , the plurality of extracted predetermined frequency components
Input from the input terminal in parallel, and
For a minute, using a plurality of preset operation constants,
Performs a predetermined operation including a product-sum operation, and calculates
A product-sum operation circuit for outputting a signal obtained from the vehicle, and based on the vehicle impact signal,
Discrimination that derives a discrimination signal for discriminating whether the line is running
Means, and the signal output from the product-sum operation circuit is
Protection device should be activated, and
The discrimination signal derived by the means is a vehicle collision
The start control signal,
A start control signal for outputting a signal for starting the occupant protection device
Output means , wherein the product-sum operation circuit outputs the value obtained by learning to the
Neural network set as a numerical operation constant
And as the plurality of operation constants in the product-sum operation circuit,
The product-sum operation circuit learning method according to claim 3,
The starting control device of the occupant protection device using the calculated operation constant.