JP2540431B2 - Collision state prediction system by neural network - Google Patents

Abstract

PURPOSE:To promptly predict the conditions which may occur in the collision, and rapidly and surely secure the safety of an occupant by providing a displacement predicting means to quantitatively predict the conditions which may occur subsequently for the data inputted by applying the parallel processing algorithm. CONSTITUTION:The preliminarily specified collision wave is inputted in an input layer 1 of a first neural network and a second neural network having an intermediate layer as the learning data, and the desired output data are given to an output layer 4. The learning operation of a displacement predicting circuit and an air bag unfolding discriminating circuit is executed by these data. In addition, the collision wave pattern data detected by an acceleration sensor S during the collision based on the neural network in the learning completed condition are inputted, and the specified threshold distance reaching time is extrapolatedly predicted by means of a displacement predicting circuit based on the learning results of the neural network. The air bag unfolding action according to the collided conditions is determined by the logical product with the air bag unfolding signal by the air bag unfolding discriminating circuit.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a collision state prediction system using a neural network, and more particularly to predicting the time until the occupant displacement threshold is reached by inputting a learning collision waveform into the neural network in advance. The present invention relates to a collision state prediction system in which the protection means is operated early and surely for the passenger represented by an airbag based on the collision waveform obtained at the time of a collision based on the learning result.

[0002]

2. Description of the Related Art In recent years, various techniques have been developed to maintain the safety of passengers in the event of an automobile collision.
Among them, the airbag system has recently been widely installed in commercial vehicles. The airbag system currently being developed is a bag-like cushion that inflates in the event of a collision and receives the upper body of the passenger, and the cushion gas is generated almost instantaneously by generating an inert gas such as nitrogen gas due to the combustion reaction of the stored drug. An inflator that can be inflated, an in-vehicle sensor that detects a collision, and a determination circuit that determines whether or not a drug in the inflator can be ignited based on a detection signal from the in-vehicle sensor.

In most cases, acceleration sensors are used for the above-mentioned vehicle-mounted sensors. Generally, an acceleration sensor has a mass supported by an elastic body, and has a structure that detects acceleration acting on the mass and detects a collision. Further, the acceleration sensor mounted on the vehicle needs to have its natural period set sufficiently short in response to the impact acceleration that lasts for a relatively long time.

It is known that the acceleration data of the collision waveform at the time of a sudden impact has a very complicated waveform distribution and it is difficult to grasp the waveform. In FIG. 8, acceleration data and
A velocity curve obtained by integrating this acceleration data and a displacement curve obtained by integrating this velocity curve are shown at the same time. In the figure (a), the car is 14MPH (Miles pe
r Hour) = 22.5 km / h, which is the collision waveform obtained when the electric pole collided. In the figure, α is an acceleration curve, v is a velocity curve, and d is a displacement curve. FIG. 6B shows each waveform at low speed (8 MPH = 12.8 km / h). It can be seen from all the waveforms that the acceleration data has a complicated time-series amplitude characteristic and low systematicity. Therefore, development has been advanced to improve the accuracy of collected data and to improve the sensitivity of the sensor so that a collision can be detected at an early timing.

However, the high-sensitivity electronic sensor developed for such a purpose has a simple structure in which the mass in the sensor is braked by viscous resistance or moment of inertia, as is the case with conventional acceleration sensors. Absent. Further, the sensor sensitivity may be too sensitive and the airbag may malfunction. Therefore, multiple sensors are arranged in each part of the vehicle body to determine the logical product of the output signals from each sensor, and a large number of collision waveforms are collected by actual vehicle experiments, and the prediction parameters are set based on the measured values. A decision circuit including a prediction algorithm for calculation has also been proposed. As an example of this type of prediction algorithm, there is an integration-type prediction algorithm in which the acceleration waveform at the initial stage of collision is filtered by an analog filter and predetermined integration is performed. This prediction algorithm utilizes that the filtered acceleration waveform is approximated by a straight line having a constant gradient, and the gradient value is used as a parameter to determine whether a collision occurs.

By the way, the current airbag device is a seat belt device which prevents the occupant from moving forward and hitting a steering wheel or a shield window by inflating a cushion of the airbag at a position in front of the occupant at the time of a collision. The main purpose is to prevent it. That is, the cushion of the airbag device needs to be fully deployed before the occupant excessively moves forward in a collision.

Therefore, in relation to the displacement of the occupant after the collision, the TTF (Time) is used as the nominal reference value of the time required to fully deploy the airbag at a predetermined timing.
toFire) has been proposed. The meaning of this TTF will be described with reference to FIG. Here, it is assumed that the required deployment time from the start of deployment of the airbag B (which may be considered to coincide with the start of operation of the inflator) to the full deployment is 30 ms from experimental values and actual results. At this time, the TTF can be defined as the latest possible time to effectively fully deploy the airbag B until the head of the occupant M moves 5 inches forward from the normal position at the time of a collision. . That is, the above relationship can be expressed by the following equation. TTF = T5 ″ −30 (ms) (Equation 1) where T5 ″: the time required from the moment of the collision until the occupant's head moves forward 5 inches. Therefore, at the latest, at this TTF timing, the airbag When the deployment of No. 1 is started, the passenger who has moved 5 inches forward can be reliably received by the fully deployed airbag.

[0008]

By the way, this TTF
In order to obtain, the conventional prediction algorithm uses the acceleration as it is as a reference parameter. The acceleration data obtained from the acceleration sensor has very strong temporal autocorrelation as a property of time series data. Therefore, when the acceleration waveform is treated as a waveform pattern for analysis, the algorithm of the feature extraction function required for recognition of this waveform pattern becomes considerably complicated. That is, in order to perform this kind of data pattern extraction with a conventional sequential program, the algorithm becomes complicated, and even if the program is completed, it has flexibility and generalization ability for changing the specifications such as adding functions of the program. Is low.

Furthermore, the conventional prediction algorithm does not have a quantitative time-dependent prediction function of predicting a state change after a collision. For this reason, the actual time of T5 ″ after the collision cannot be known in advance. Therefore, the actual collision waveform and the reference waveform are compared with each other, and the analog determination is made to determine whether the ignition of the airbag is 0.
N, OFF is determined. The ignition timing corresponds to the time corresponding to TTF set in advance in a predetermined range.

Therefore, an object of the present invention is to solve the above-mentioned problems of the prior art, process complicated and enormous data obtained at the time of collision of an automobile, extract feature of input data pattern and predict time series. The algorithm and self-organization and learning are performed in the parallel processing process by the neural network, and based on the input data, the data of the state that will continue to occur is quantitatively predicted as the output value, and the boarding mounted on the car It is an object of the present invention to provide a collision state prediction system using a neural network that can accurately determine whether or not a safety device for protecting persons is activated.

[0011]

In order to achieve the above object, the present invention comprises a detection means which is mounted on a vehicle and which can collect the impact force acting on an occupant during a collision of the vehicle as a continuous collision waveform. A plurality of learning collision waveforms are input in advance as a first parallel learning data group, the learning collision waveforms are converted into displacement data of a predetermined part of the occupant's body, and learning is performed. The collision waveform data obtained by the detection means when a collision occurs is input to the first neural network as a data set having a predetermined time width, and the occupant's body An occupant displacement predicting means for extrapolating an arrival time at which a predetermined portion reaches a displacement threshold and predicting a latest threshold time for deploying the occupant's body to be protected by an airbag, and the airbag. Unfolded Is input to the second learning data group as the learning collision waveform, and the collision waveform data at the time of the collision is input to the second neural network that has undergone the learning calculation. Input at the same time as input to the net,
Based on the learning result of the second neural network, the airbag deployment identifying means for identifying whether or not the airbag needs to be deployed and outputting an identification signal, the latest threshold time and the identification signal are input, A logic circuit that determines whether or not the airbag needs to be deployed with respect to the collision state at the latest threshold time, and outputs an airbag deployment operation command signal based on the determination, and an airbag deployment from the logic circuit. An airbag deployment operation means for inflating and deploying the airbag based on the operation command signal is provided. At this time, the first
In the neural network and the second neural network, it is preferable to construct a two-dimensional self-organizing competitive learning layer in the intermediate layer.
It is preferable that the learning calculation is performed using a back propagation method.

In the first parallel learning data group, the input data input to the input layer of the neural network by performing parallel conversion of the continuous displacement data in the predetermined range and the predetermined range connected to the continuous displacement data. It is preferable that the displacement data is converted into parallel data and applied to the output layer of the neural network. Further, it is preferable that the first neural network and the second neural network are compiled in a learning completed state to generate a code and to be incorporated in a substrate of a control circuit of the airbag system.

[0013]

According to the present invention, the vehicle is equipped with a detecting means capable of collecting the impact force acting on the occupant in the event of a collision as a continuous collision waveform, and a plurality of learning collision waveforms are previously stored in the first parallel learning data group. Is input as the learning collision waveform to the displacement data of a predetermined part of the occupant's body and learning is performed on the first neural network having the learned intermediate layer, and when the collision means is used, The obtained collision waveform data is input as a data set having a predetermined time width, and the arrival time at which a predetermined part of the occupant's body reaches a displacement threshold is extrapolated and predicted based on the learning result in the first neural network. However, the occupant displacement predicting means predicts the latest threshold time for deploying the occupant's body so as to be protected by the airbag, and at the same time, it is recognized that the airbag needs to be deployed. A second neural network learning operation performed was learned by entering a learning collision waveform to a second training data set, the collision waveform data at the time of the collision first
Input to the neural network at the same time, based on the learning result of the second neural network, the necessity of deploying the airbag is identified and an identification signal is output by the airbag deployment identifying means.
Determining the logical product of the latest threshold time and the identification signal,
The necessity of the airbag deployment operation according to the collision state is determined from the logical product, and the airbag deployment operation command signal based on the determination is output to the airbag deployment operation means. An appropriate weighting coefficient is set in the intermediate layer by learning the first neural network using the collision waveform, and the learned neural network is processed with some initial data of the actual collision waveform. Input with a predetermined width, predict the subsequent changes in data,
The time to reach the set threshold value can be calculated from the input value. In addition, using the airbag deployment identification means,
By taking the logical product, a signal that does not deploy the airbag in the case of a slight collision can be obtained, so whether or not the airbag should be deployed in accordance with the collision state and whether the deployment start time should be accurate. The prediction can be determined.

The first learning data group includes input data input into the input layer of the neural network by parallel conversion of continuous displacement data in a predetermined range, and a predetermined range connected to the continuous displacement data. Acceleration data having a strong autocorrelation by being composed of desired output data to be applied to the output layer of the neural network by performing parallel conversion of displacement data,
It can be simplified considerably by performing appropriate pre-processing such as integration calculation, weakening the strong autocorrelation of the acceleration waveform, facilitating feature extraction and self-organization in the neural network, The calculation efficiency can be significantly improved.

Further, by compiling the first neural network and the second neural network in a learning completed state to generate a code and incorporating the code into a substrate of a control circuit of an airbag system, a small and high performance airbag is provided. The operation determination circuit can be constructed and the manufacturing cost thereof can be significantly reduced.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention inputs a collision waveform at the time of collision into a neural network as learning input data, uses the learned neural network as a judgment circuit, and uses an airbag at an appropriate timing during an actual collision. The prediction is made for deployment. A preferred embodiment will be described below by taking a collision state of an automobile as an example. First, a known general neural network will be outlined, and an embodiment of a collision state prediction system using the neural network according to the present invention will be described with reference to the accompanying drawings.

[Neural Network] The neural network is an information processing system developed by imitating the information processing mechanism of the neuro network that constitutes the cranial nervous system. The neural network has a parallel processing function in which a large number of processing elements (hereinafter referred to as PEs) corresponding to neurons operate at the same time. The structure is P
It has a hierarchical structure in which Es are connected to each other, and parallel distributed processing can be realized through each hierarchy.

Here, the function of the PE will be briefly described. As shown in FIG. 2, the PE has a multi-input / single-output element and has a considerably simplified structure as compared with an actual neuron in the brain. In the function calculation of the element, first, the plurality of input values xi are multiplied by the weighting coefficient wi. Then, an output value f (X) of a predetermined transfer function corresponding to the total sum (X = Σxi · wi) is output. This output value f (X) is for each PE
The energy values of the above are shown, and the output value for each input value is fed back to the input side again to form a stable state as the whole system. The method for this stabilization depends on the theory of minimizing the energy of the system. Here, an example of a prediction system using two different learning algorithms will be described.

[First Embodiment] In the first embodiment shown below, in addition to the input value input to the input side of the neural network,
A back propagation (Back Propagation, hereinafter referred to as BP) method for performing learning with a desired learning output value output to the output side is applied. This BP method is used to predict the airbag deployment determination during a vehicle collision. Here, the concept of the BP method will be briefly described. The neural network used in the BP method is composed of a multi-layer network having several hidden layers as an intermediate layer in addition to PEs constituting an input layer and an output layer. The learning algorithm is as follows. First, when the desired output value (teaching data) is fed back to the input value (learning data), the differential coefficient of the transfer function is multiplied by the difference (error) between the actual output value of each PE and the desired output value. Then, the above-mentioned weighting factors wi are updated one after another to minimize the error function between the desired output and the actual output. That is, the energy state is changed so as to minimize the error in the direction of decreasing the energy of the system, and the learning is repeated until the system finally reaches the equilibrium state. If the output error becomes sufficiently small and the continuity between the input data string and the data string that will be output near the boundary between the input value and the output value is obtained, the learning is completed. Optimization methods for advancing this learning include generalized δ rule, steepest descent method, and the like. In these methods, random numbers are initially generated to determine learning data to be input. After that, the learning is performed only on the portion other than that where the energy of the system is in the constant state. Thereby, the calculation can be converged at an early stage.

In the actual learning by the BP method, predetermined input learning data is input to the PE of the input layer, and each PE connected to the PE of the input layer performs an operation. Then, the PE output value obtained in the output layer is compared with the desired output value.
The error at that time is obtained, and the calculation is repeated until the error converges. This repeated calculation is taken as one cycle,
Data to be input to the input layer is randomly extracted from the input data set and the same repeated calculation is executed. Then, when the threshold value within the preset allowable error is obtained, the weighting coefficient of each PE at that time is obtained and the learning is completed. In this embodiment, there are two systems of an occupant displacement predicting circuit for predicting the TTF and an identifying circuit for outputting ON / OFF (1/0) data for identifying whether or not the airbag inflator is turned ON. It has a neural network.
The contents of each neural network circuit will be described below.

(Learning Displacement Prediction Circuit Learning Plan) A displacement prediction circuit learning plan to which the above-mentioned BP method is applied will be described. (1) Collection of Collision Waveform from Acceleration Sensor FIG. 1 shows a state in which the automobile C has collided with an object Ob such as a barrier. At this time, the analog-type collision waveform detected by the acceleration sensor SF mounted on the front part of the automobile C is noise-removed by a low-pass filter or the like, and a signal converter 20 including a predetermined sample-hold circuit and an AD converter. Convert to digital data. Then, this digital data is collected at a sampling interval (Δt) over a sampling range (ΔT) to collect collision waveform data that has acted on the occupant's body, particularly the head. (In the present embodiment, Δt
= 0.1 ms, ΔT = 200 ms. It should be noted that the acceleration sensor for collecting the collision waveform used for learning has different waveform characteristics at the mounting position thereof, and therefore it is preferable to perform waveform characteristic extraction in order to remove an error due to a difference in sensor position.

(2) Processing of Collision Waveform Data (Preprocessing) The collision waveform data transferred and input through the parallel input interface 21 is integrated twice, and data indicating the amount of forward movement of the occupant's head (hereinafter , Displacement data. In the graph, the curve indicated by the symbol d) is obtained. In this embodiment, some layers of the above-mentioned neural network are used to obtain the displacement data. That is, as shown in FIG. 1, the weighting factor wi =
The intermediate layers corresponding to two layers, which are set to 1, are used as the integration layers 2a and 2b. The collision input data is input to the input layer 1 as parallel data by the parallel input interface 21, and 2
Iterative addition corresponding to the round integration is executed by the neural network. As a result, a series of displacement data strings is obtained as an output value from the integration layer 2b. At this time, since the collision waveform data sequence and the displacement data sequence uniquely correspond to each other (that is, they have a mapping relationship), if the above-described sampling interval Δt is used, the amount of forward displacement of the occupant's head per predetermined time width is increased. Can be easily grasped. It is also possible to improve the calculation speed by handling the collision waveform data as parallel data of a predetermined time width and directly performing the calculation by the circuit of the neural network mounted on the hardware.

(3) Structure of neural network The neural network in this embodiment has a hierarchical structure as shown in FIG. There are n input layers 1 (n = 64 in this embodiment).
(Set to 1.) are arranged side by side in a row. Furthermore, on the output side of this input layer 1, the above-mentioned integration layer 2
PEs a and 2b are connected in series. This integration layer 2
The PE values of the displacement data obtained in a and b pass through the n hidden layers 3 and m (PE in this embodiment is set to m = 300) PEs are arranged side by side in a displacement prediction. It is output to the output layer 4 of the circuit. Further, the output value of the input layer 1 is also output to the other hidden layer 5 as it is. Then, an identification signal to be described later is output to the output layer 6 of the airbag deployment identification circuit of one PE connected to the other hidden layer 5. Note that the weighting factor wi between the PE layers before and after mutual input / output is performed.
Is considered, and all PEs are connected to each other.

(4) Setting of Transfer Function In this embodiment, the sigmoid function shown in FIG. 3 is adopted as the transfer function of each PE. This sigmoid function is a function having a quasi-linear saturated transfer characteristic. By using this function, a neural network model with good calculation efficiency can be set. A y-direction scale and a y-coordinate offset are set as parameters of the sigmoid function in each layer. In this embodiment, the values shown in the following table are adopted. As shown in the table, the y-direction scale of the function is defined so as to show a predetermined exponential change for each layer. This improves the convergence efficiency of the calculation.

[0025]

[Table 1]

Various other functions can be applied as the transfer function. For example, the sine function has a wide variation range in which the differential coefficient with respect to the total input at each PE is equivalent to that of the original function. The use of this sine function makes the hardware configuration a little complicated, but the learning convergence is significantly improved.

(5) Learning Method, Number of Learning Data In the present embodiment, a series of discrete displacement data which maintains continuity with each other is provided for each number of data equal to the number (n) of PEs in the input layer of the neural network. Be divided. A series of discrete displacement data in time series is continuously extracted in the form of n serial data. Then, n pieces of data are handled as one data set. This data set is first converted into parallel data, and the input learning data of the neural network is n
It is input to each PE. At the same time, the above n
The m pieces of displacement data generated subsequently to the pieces of displacement data are converted into parallel data and input to the output layer as desired output data. In this embodiment, n = 64 and m = 300 are set. 4 to 6, the horizontal axis represents a state in which parallel data is divided into data sets of a predetermined width and input. The vertical axis conceptually shows displacement data corresponding to the forward displacement amount (movement amount) of the occupant's head at that time.

FIG. 4 schematically shows the relationship between the converted parallel data and the original displacement data curve d. That is, if the learned input data and the desired output data are arranged side by side in a line, they can be displayed like a horizontally long strip graph in the lower part of the figure. Here, the time width between the input data and the desired output data will be described with reference to FIG. Since the data sampling interval of this embodiment is Δt, the time width of the input data is (Δt × n) ms. The time width of the desired output data corresponds to (ΔT × m) ms. Since Δt = 0.1 ms in this embodiment,
The input data is set to 6.4 ms width and the output data is set to 30 ms width. The value of 30 ms of this desired output data corresponds to the time taken for the airbag to fully deploy. In this way, the total time required to deploy the airbag can be represented by the product of the sampling interval and the number of data. Therefore, if the sampling interval is made large, a compact neural network with the same accuracy can be constructed with a small number of data.

By thus setting the predetermined time width in the data set of the parallel data, the parallel data can be considered to be equivalent to the time series displacement data. If the final value of the desired output data is the threshold dT for obtaining T5 ″,
After that, the final value of the input data that is input one after another is set to TTF.
Can be made equivalent to.

Further, in the iterative calculation by the actual computer according to the present embodiment, the data sets of 80 sets of parallel data are shifted by 1 ms and input to the input layer one after another (see FIG. 5). It should be noted that the number of data sets to be input may be set to a number that can cover the range in which the feature extraction is easy in the displacement curve.

FIG. 5 shows the parallel data input procedure in relation to the displacement data curve. Here, a data set having a predetermined time width (6.4 ms width in this embodiment) is defined as one window Win. One window W
Each displacement data input by shifting in by 80 times at 1 ms intervals corresponds to learning data. Further, a data set of displacement data for 30 ms continuing from this learning data is defined as a window Wout of desired output data. Two types of data sets from the input side and the output side are given to the neural network simultaneously and in parallel, and the learning is advanced.

Further, in this embodiment, 10 displacement curves (d1 to d10) were used as learning data, and the neural network was learned. Each learning is repeated a predetermined number of times. The convergence error limit at this time is set within 1% of the output range. The value of the convergence error limit may be freely set as long as the continuity at the boundary between the input value and the output value is ensured. Further, the width of the input window can be arbitrarily set within the range of the width where the serial data can be effectively used other than the 6.4 ms width. At this time, it is more advantageous if the width of the window is expanded to a range in which the characteristic curve shape of the displacement curve formed by the input data in one window can be easily grasped. Further, the number of learning data, the number of displacement data, and the number of times of learning are values that can be set appropriately in consideration of individual learning effects, and the values described above are not limited at all.

(6) Examples of learning data and unlearned data The following table shows examples of the learning data and unlearned data used in this embodiment.

[0034]

[Table 2]

In the table, in the cases marked with a circle, both the input data and the desired output data were adopted for learning the neural network. In other cases, only the input data was input to the neural network and used to confirm the output value.

(Learning Plan of Airbag Deployment Identification Circuit) When learning is completed by using the neural network of the occupant displacement prediction circuit described above, the threshold value (5 inches forward displacement of the passenger's head as prediction output data ( The time point (time) at which the displacement threshold value is obtained means T5 ″ after 30 ms. That is, the current time point (time) at which it is recognized that the threshold value has been reached corresponds to TTF.

However, even if the required TTF is the "OFF condition" set so as not to ignite the airbag as in the case A-10 in Table 2, T5 "is calculated as a value corresponding to the displacement of 5 inches. = 72.4 ms is obtained.
And TTF is also calculated as follows. TTF = 72.4-30 = 42.4 ms Such a long TTF often occurs at low speed where it is not necessary to deploy the airbag. Therefore, it is possible to set a time limit in order to exclude such data. However, as in case B-8 in Table 2, TT
There is a case where a request TTF of "ON condition" that the ignition of the airbag is required even when F = 45 ms is set, and it is not possible to uniformly adopt the time limit.

Therefore, in this embodiment, whether or not the inflator needs to be actuated to inflate the air bag is discriminated and ON / OF is determined.
An airbag deployment identifying circuit that outputs an F signal is provided as a separate circuit independent of the neural network for predicting the displacement of the head of the occupant. The airbag deployment identifying circuit shares the input layer 1 with the occupant displacement predicting circuit described above, and functions by the same collision waveform data input at the same time (see FIG. 1).

In this airbag deployment identifying circuit, a plurality of acceleration data corresponding to the limit waveform of airbag deployment availability is used in advance as learning data. Learning is performed based on the airbag deployment limit waveform, and whether or not the collision that is the basis of the waveform has reached a level at which the airbag should be activated. In addition to the TTF prediction obtained by the displacement prediction circuit, whether the airbag can be deployed is 1,
It is output to the output layer 6 as an identification signal such as 0. Based on this identification signal, it is possible to determine ON / OFF of the inflator ignition based on the collision waveform shape. Therefore, the identification signal and the output TTF of the displacement prediction circuit are input to the logic circuit 22 to obtain the logical product of both signals, and in either case, only when the ignition is turned on, the inflator 2 of the airbag 2 is obtained.
An ignition command is output to 3. As a result, it is possible to prevent the airbag from malfunctioning in the case of a slight collision.

(Prediction Operation in Actual Vehicle) An embodiment in which the collision state prediction system using the neural network according to this embodiment is applied to an actual vehicle will be briefly described. The learned neural network network program can be stored in the ROM as an application by compiling it and converting it into a machine language. Then, it can be incorporated as a circuit specific to a predetermined IC chip or the like, and this chip is mounted on the circuit board of the airbag operation control device of the actual vehicle.

In a vehicle equipped with a collision state prediction circuit for operating such an airbag only at an early and effective timing, when the driver starts the engine, the acceleration sensor S connected to the collision state prediction circuit is activated. Detectable state. The analog acceleration data detected by the acceleration sensor S is A / D converted into digital data by a signal conversion unit 20 equipped with a sample hold circuit and an A / D converter, and is further passed through a parallel input interface 21 to obtain the data shown in FIG. As shown, the data is sequentially converted into a parallel data set forming a window Win having a predetermined window width.
Then, each data set is continuously input as parallel data to the input layer 1 of the neural network incorporated on the circuit board one after another at predetermined time intervals. In a normal driving state, acceleration generated by a change in speed during normal traveling is treated as noise and is removed by noise removing means such as an analog filter. Therefore, during normal driving, zero-value data is continuously input to the input layer of the neural network.

When a collision occurs at a certain point in time, the impact acceleration data is input to the acceleration sensor S, and the acceleration data is input to the neural network through the signal conversion unit 20 and the parallel input interface 21 in a predetermined window Wink.
Iterative calculation of weight calculation is performed on the series of data in the neural network. As a result, the predicted output data corresponding to the displacement data up to 30 ms after the input data is obtained almost immediately. When the input acceleration data changes from moment to moment and the output value corresponding to the input value reaches the threshold dT corresponding to a 5-inch displacement, the repetitive calculation ends. The last time of the data set of the window Wink at this time is output as TTF. Further, a logical product of the obtained TTF and the identification signal of the airbag deployment identification circuit is obtained by the logic circuit 22.

As described above, in this embodiment, T5 "can be accurately predicted at a time point 30 ms before T5", and the TTF at that time can be adopted to judge whether or not the airbag can be deployed. Further, this judgment is made by the neural network self-organized by the learned data. Therefore, it is possible to subsequently improve learning accuracy by performing learning with a simple algorithm.

[Second Embodiment] Next, a self-organizing function and a competitive learning function are incorporated in an intermediate layer in a neural network in order to accurately respond to a large amount of unlearned data with a small amount of learned data. Examples will be described. In this embodiment, FIG.
A two-dimensional self-organizing / competitive learning layer 10 is provided in front of the hidden layers 3 and 5, as shown in FIG. This two-dimensional self-organizing / competitive learning layer 10 is a two-dimensional-Kohonen layer (hereinafter, 2D-
In the present embodiment, 8 × 5 = 40 (P).
E). In this 2D-K layer 10, a predetermined geometric distance of 40 PE is calculated for each input data set. As a result, the 64 pattern attributes that have been input are standardized, and the attribute determination for unlearned data becomes easy.

Specifically, among the input 40 PE data, three outputs are selected by a competitive learning method such as mapping. Further, the unlearned data is preprocessed according to the division of the input data by self-organization at the time of learning, and the output value is passed to the next hidden layer. 2D-
By additionally inserting the normalization layer 11 immediately before the K layer 10, the learning efficiency in the 2D-K layer 10 can be improved. By providing the 2D-K layer 10 in this way, it is clarified that the convergence efficiency at the time of learning can be significantly improved, and the accuracy rate for unlearned data can be extremely improved.

The neural network having this 2D-K layer can be completed by expanding the above-mentioned BP method. Therefore, the learning plan can be determined according to the BP method. In this neural network, feedback is not performed from the output side due to internal self-organized learning in the repeated learning state of thousands of times immediately after the start of learning. Then, after the self-organization is completed, a predetermined BP method is learned. As a result, learning can be completed with a rapid number of convergences.

In the above-mentioned two embodiments, a compact executable application can be constructed by creating a neural network after learning as a program in C language or the like and compiling it. It can also be built into the chip as a control microcomputer package. At this time, if the corresponding vehicle type is stored in the ROM, it is possible to construct a prediction system with a compact neural network corresponding to the vehicle type to be mounted, and it is possible to reduce the cost.

It is also possible to apply an expert system to an algorithm part for which logic construction is easy, and construct a prediction system by a hybrid logic circuit of a neural network and an expert system.

[0049]

As is apparent from the above description, according to the present invention, a parallel processing algorithm capable of performing a calculation quickly without applying a complicated and huge amount of data is applied to the input data. Since the state to be obtained is quantitatively predicted, it is possible to quickly predict the state that will occur at the time of a collision, and it is possible to give an instruction to promptly and reliably ensure the safety of passengers.

[Brief description of drawings]

FIG. 1 is a schematic configuration diagram including a neural network showing an example of the neural network of the first embodiment of the collision state prediction system by the neural network according to the present invention.

FIG. 2 is a conceptual diagram schematically showing one element of a neural network applied to the present invention.

FIG. 3 is a characteristic curve diagram showing an example of a transfer function used for learning calculation in the present invention.

FIG. 4 is a relationship schematic diagram showing a relationship between parallel data and a displacement curve.

FIG. 5 is a relationship schematic diagram showing a relationship between an input window of parallel data and a displacement curve.

FIG. 6 is a relational schematic diagram showing a prediction operation in an actual vehicle based on a relation between a parallel data input window and a displacement curve.

FIG. 7 is a neural network configuration diagram showing an example of the neural network of the second embodiment of the collision state prediction system by the neural network according to the present invention.

FIG. 8 is a characteristic curve diagram showing each relationship of a collision waveform curve acting on an occupant collected by an acceleration sensor, a velocity curve of movement of the occupant's body at that time, and a displacement curve of the occupant's body.

FIG. 9 is an explanatory view schematically showing a state in which movement of the occupant, which is a premise of T5 ″, is restrained by the deployed airbag when the occupant moves forward by 5 inches.

[Explanation of symbols]

 1 Input Layer 2a, 2b Integral Layer 3, 5 Hidden Layer 4, 6 Output Layer 10 Two-dimensional Self-Organization / Competitive Learning Layer (2D-K Layer) 11 Normalization Layer 20 Preprocessor 21 Input Interface 22 Logic Circuit 23 Inflator α Acceleration curve v Velocity change curve d Displacement curve B Airbag C Automobile M Passenger Ob Collision target S Front acceleration sensor

Claims (5)

(57) [Claims] 1. A detection means which is mounted on a vehicle and is capable of collecting an impact force acting on an occupant at the time of a collision of the vehicle as a continuous collision waveform, and a plurality of learning collision waveforms having a first parallel learning data group in advance. Is input as the learning collision waveform to the displacement data of a predetermined part of the occupant's body and learning is performed on the first neural network having the learned intermediate layer, and when the collision means is used, The obtained collision waveform data is input as a data set having a predetermined time width, and the arrival time at which a predetermined part of the occupant's body reaches a displacement threshold is extrapolated and predicted based on the learning result in the first neural network. The occupant displacement predicting means for predicting the latest threshold time for deploying the occupant's body so as to be protected by the airbag, and the limit waveform for deploying the airbag are the second learning data as a learning collision waveform. The collision waveform data at the time of the collision is input to the learned second neural network which is input to the group and on which the learning operation is performed, simultaneously with the input to the first neural network, and the second neural network is input. Based on the learning result of the circuit network, an airbag deployment identifying means for identifying the necessity of airbag deployment and outputting an identification signal, and inputting the latest threshold time and the identification signal, the latest threshold time The necessity of deploying the airbag with respect to the collision state of the vehicle, and a logic circuit that outputs an airbag deployment operation command signal based on the determination, and an airbag deployment operation command signal from the logic circuit. A collision state prediction system using a neural network, comprising: an airbag deployment operation means for inflating and deploying the airbag. 2. The neural network according to claim 1, wherein the first neural network and the second neural network are constructed by constructing a two-dimensional self-organizing competitive learning layer in the intermediate layer. Collision prediction system. 3. The collision state prediction system by a neural network according to claim 1, wherein the learning operation is performed by using a backpropagation method. 4. The first parallel learning data group includes input data input to an input layer of the neural network by performing parallel conversion of continuous displacement data in a predetermined range, and a predetermined connection to the continuous displacement data. Collision by the neural network according to claim 1 or 2, wherein the displacement data of the range is parallel-converted and desired output data is applied to the output layer of the neural network. State prediction system. 5. The first neural network and the second neural network are coded in a learning completed state and incorporated in a substrate of a control circuit for operating the airbag. A collision state prediction system using the neural network according to claim 4.