JPH1067294A - Method for detecting collision of vehicle and device thereof - Google Patents

Abstract

PROBLEM TO BE SOLVED: To determine the development of a passive protection device such as an air bag by analyzing data of an accelerometer which is prepared at the time of collision and is digitalized. SOLUTION: In a front part of a car room 100 of a vehicle, dual air bags 110, 111 and a diagnosis module (SDM) 120 including a three-shaft accelerometer 122, an analog-digital converter, and a one-point type collision sensor including pattern recognition algorithm included in a microprocessor 130 are provided. The three-shaft accelerometer is a device which includes three accelerometers and measures acceleration in three orthogonal directions of a longitudinal direction, a horizontal direction, and a vertical direction. The pattern recognition technique is a neutral network, and this neutral network analyzes data from all three accelerometers and judges whether a vehicle receives collision from any direction or not.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a method and an apparatus for detecting a collision of a vehicle.

[0002]

2. Description of the Related Art Pattern recognition techniques, such as artificial neural networks, are becoming more and more widely applied to solve various problems, such as optical feature recognition, speech recognition, and military target identification. It has become to. In the automotive industry, pattern recognition technology is now identifying various objects, such as rear-facing (back-facing) children's seats in the passenger compartment, and threatening objects, such as approaching vehicles that are likely to collide against the side of the vehicle. It has also been applied to confirming. See, for example, U.S. Patent Application No. 0, May 9, 1994, which is incorporated herein by reference.
See 8 / 239,978 and U.S. patent application Ser. No. 08 / 247,760, issued May 23, 1994. To date, pattern recognition technology has sensed whether a car crashed and deployed airbags or other passive protection after the crash, whether to fasten seat belts, shut off the fuel system, or to close doors. It is not used to determine whether to release the lock.

The term "pattern recognition" as used herein refers to a signal emanating from an object or modulated by interaction with an object to determine to which class the object belongs. Means the system to process. Such a system may determine whether an object is a member of a particular classification, or may apply the object to a particular classification in a classification group containing multiple classifications, or It is only determined that the category does not correspond to any of the categories. The processed signal is typically an electrical signal output from a transducer that is sensitive to acceleration, sound or electromagnetic radiation, and if electromagnetic, may be either visible light, infrared light, ultraviolet light or radar. it can.

[0004] As used herein, the term "confirm" means:
This means determining that an object belongs to a specific classification or classification group. The classification is determined according to the purpose of the system. For example, a classification that includes all collisions that require a frontal impact airbag, a classification that includes all events that do not require an airbag, a classification that includes all events that require the position of an occupant headrest to be moved, and a classification that includes a side collision. At this time, the airbag is classified as a classification including all events that require the deployment of the airbag.

[0005] All electronic crash sensors currently used for frontal crashes include an accelerometer that senses and measures the acceleration of the vehicle during a crash. The accelerometer generates an analog signal proportional to the acceleration acting on the accelerometer, that is, the acceleration of the vehicle on which the accelerometer is mounted.
An analog-to-digital converter converts this analog signal into a time-series digital value. The crash sensor designer reviews the digital acceleration data and derives a computer algorithm that determines whether the acceleration data obtained from a particular crash is such that the airbag deployment is permitted. This is usually done by trial and error, with the technician or crash sensor designer observing when an airbag is required but not required, and observing crash data for other events where an airbag is not required. And ultimately, the technician or crash sensor designer determines a crash library, an algorithm that is likely to meet the requirements of the crash data accumulated from many crashes and other events. The algorithms so obtained are not universal and most technicians or crash sensor designers will give a negative answer when asked if they work for all vehicles .

[0006] Several articles have been published pointing out the problems and limitations of electronic crash sensors mounted in protected locations outside the vehicle's collision zone, usually in the passenger compartment. Here, the collision zone is a portion where the collision sensor has already collapsed when a collision occurs in which the deployment of the occupant restraint system has to be started. These sensors are often called single point crash sensors. The following technical papers discuss the theory of collision detection together with the limitations of this one-point collision sensor. These are relevant to the present invention and are incorporated herein by reference. 1) Breed, DS and Castelli, V. “Problems in the design and engineering of airbag systems.
in Design and Engineering of Air Bag System
s) ", American Society of Automotive Engineers SAE 880724,
1992. 2) Breed, DS and Castelli, V. "Trends in Sensing Frontal Impact", American Society of Automotive Engineers SAE 890750, 1989. 3) Breed, DS, Double Tea Sanders (Sanders, WT) and Buoy
Castelli, V. “A Critique of Single Point Crash Sensin
g) ", American Society of Automotive Engineers SAE 920124,
1992. 4) Breed, DS, Double T Sanders (Sanders, WT) and Buoy
Castelli, V. "A complete Frontal Crash Sensor System I"
I) ", American Society of Automotive Engineers paper SAE 930650,
1993. 5) Breed, DS and Sanders, WT, "Using Vehicle Deformation to Collision Detection"
Sense Crashes ", International Body and Engineering Conferenc
e), Detroit, Michigan, 1993. 6) Breed (DS), Double Tea Sanders (Sanders, WT) and Buoy
Castelli, V. “A complete Frontal Crash Sensor System I
I) ", Minutes of the Automotive Safety Enhancement Conference, Munich,
1994, US Department of Transportation National Highway Traffic Safety Administration, Washington, D.C. C. Published.

[0007]

These papers enable engineers to develop algorithms that sense a collision and selectively deploy an airbag, especially when sensors are not located in the collision zone of the vehicle. This indicates that there is no such known theory. These articles show that the information contained in the acceleration signal, which is typically measured in the cabin, is not sufficient to detect all crashes. Another conclusion supported by these technical papers is that if an algorithm that works on one vehicle can be found, it will be applicable to all vehicles. This is because any vehicle can produce some kind of collision pulse. See in particular the SAE article 920124 referenced above.

[0008] Despite the problems associated with finding an optimal crash sensor algorithm, many vehicles currently in motion are equipped with electronic one-point crash sensors. Due to some issues associated with single point sensors,
If the occupant is not in a predetermined position and is too close to the airbag when the airbag is deployed, the deployment of the airbag itself may cause injury or death. Fortunately, systems are currently being developed to monitor the position of the occupant in the vehicle and to suppress deployment of the airbag if deployment of the airbag in an accident could result in injury to the occupant.

[0009] The information contained in the acceleration data measured in the cabin is insufficient to detect all crashes, and some of the failure modes that the published single-point sensor algorithm has. Can be easily proved by using the collision and velocity scaling technique described in the technical paper referred to above. Furthermore, algorithm development performed by engineers is based on trial and error. Therefore, pattern recognition techniques, such as neural networks, could be used to create algorithms by training the system for many collision and non-collision events, which would be anything else excellent. Would. In fact, this has proven to be the case and is also the subject of the present invention.

Of course, if any of the crash sensors determine that the airbag should be deployed, the system may tighten the seatbelt for a vehicle with a seatbelt retractor system, or provide a vehicle during a crash or crash. Some other functions must also be provided, such as shutting off the fuel system to prevent fuel leaks later, and unlocking the doors to help the occupant escape after a collision.

[0011] The use of pattern recognition technology in a collision sensor also has another important advantage in that the same pattern recognition hardware and software can be shared with other systems in the vehicle. Pattern recognition technology has already proven to be effective in solving other problems associated with airbag passive protection devices. In particular, it is shown that the deployment of the airbag can be suppressed by checking the back-type child seat placed on the passenger seat. It is also very promising to use pattern recognition technology for predictive sensing of side collisions and to discriminate vehicles that are likely to collide with the side of the vehicle.
These pattern recognition systems, and other systems under development, can share the same computer system as the collision sensor of the present invention. In addition, all of these systems need to interact with the main sensors and diagnostic modules used for frontal collisions. It is therefore desirable from a cost and reliability standpoint that all three systems use the same computer system. This is particularly desirable because computers are now available that are specifically designed to solve pattern recognition problems, such as neural computers.

[0012]

SUMMARY OF THE INVENTION The present invention analyzes digitized accelerometer data produced during a crash using a pattern recognition technique, such as a neural network, or an algorithm derived from the neural network. It determines whether and when to deploy a passive protection device such as an airbag. The basic objectives are as follows. 1) To provide a single point sensor, including an accelerometer, that makes the best use of the information in the acceleration data to determine whether to deploy the airbag. 2) To provide a versatile one-point collision sensor that can be used in most vehicles without making modifications specific to each vehicle. 3) In the car, several different pattern recognition functions,
For example, there is provided a single computer system capable of detecting a collision, confirming an object existing in a vehicle interior, and classifying an object outside the vehicle. 4) staged vehicle crash and non-collision events;
As well as using a set of statistically derived data to provide a training derived collision sensor algorithm. 5) To provide a collision sensor based on a pattern recognition technology. 6) Provide a collision sensor that uses other data in addition to acceleration data derived from a collision. The data is combined with the acceleration data and the use of pattern recognition techniques determines the need for deployment of passive protection devices. 7) Provide a crash sensor that can be used to sense both front, side and rear crashes. 8) To provide a collision sensor that automatically retains collision acceleration data from a predetermined time before the airbag deployment for later analysis. 9) Provide a collision sensor using a neural computer.

[0013] Other objects and advantages of the present invention will become apparent from the following disclosure.

In general, the present invention relates to a sensing system for initiating deployment of a vehicle occupant protection device, such as an airbag, to protect a vehicle occupant in the event of a collision. This system is mounted on the vehicle,
A sensor that senses the acceleration of the vehicle and generates an analog signal representing the acceleration, and receives an analog signal from the sensor,
It includes an electronic converter for converting the analog signal into a digital signal, and an arithmetic processing unit for receiving the digital signal. The processor includes a pattern recognition system,
A deployment signal is generated when it is determined that a pattern specific to a collision of a vehicle requiring passenger protection is included in the digital signal.

[0015]

DETAILED DESCRIPTION OF THE INVENTION Typically, an electronic sensing and diagnostic module (SDM) is mounted at a convenient location in a vehicle compartment, such as a transmission tunnel or a fire wall. FIG. 1 is a diagram showing a front portion of a vehicle compartment 100 of a motor vehicle, a part of which is cut away. Dual airbags 100 and 111 and three-axis accelerometer 12
2 and an SDM 120 including an analog-to-digital converter and a one-point collision sensor including a pattern recognition algorithm included in a microprocessor 130. A three-axis accelerometer is a device that contains three accelerometers,
Typically, three orthogonal accelerations are measured: longitudinal, lateral and vertical. The circuit board of the SDM 120 includes a capacitor 140 as a backup power supply,
Other electronic components 126 and various circuits are included. This SDM is connected to the airbag by wires 113 and 114. The pattern recognition technique used in this embodiment is a neural network, which analyzes data from all three accelerometers and determines whether the vehicle is undergoing a collision from either direction. Is determined. If the neural network determines that the accident requires deployment of occupant protection devices, such as front or side airbags or movable headrests, by analyzing patterns of signals emitted from, for example, accelerometers, This neural network will start the operation of the occupant protection device. Therefore, the neural network constitutes the means for starting the deployment of the airbag.
For example, for a frontal collision, the airbags 110 and 1
Signals are sent through lines 113 and 114 to initiate deployment of 11. The analog-to-digital converter is connected to a sensor, in this case a three-axis accelerometer, and converts an analog signal representing the acceleration generated by one of the accelerometers, ie the acceleration of the vehicle, into a digital signal. I do. In one embodiment, converter 124 derives a digital signal from the integral of the analog signal.

The three-axis accelerometer is mounted on the vehicle by suitable mounting means, but can be mounted in various locations, for example, to sense front, side or rear impacts. In another embodiment described below, detecting means for detecting that the occupant to be protected by the deployable occupant protection device according to the present invention is located at a position other than a predetermined position and suppressing the deployment of the airbag accordingly. May be provided in the microprocessor 130. The detecting means can also be applied to detect that the rearward facing child seat is placed on the seat of the vehicle, thereby suppressing the deployment of the airbag.

Many of the discussions herein use neural networks as an example of a pattern recognition technique or algorithm, which is one of the most developed neural networks for those techniques.
Because it is one. However, neural networks have significant limitations, which are currently being addressed by the development of new pattern recognition technologies.
The limitation is that it is difficult to describe the process used to classify the patterns, so patterns that were not included in the training set may be missed. Also, the training process of a neural network does not guarantee that it will converge to the best results. This is a local minimal problem (loca
l Minimum problem). That is,
The result of the convergence of the training algorithm is not the best solution overall. These problems are addressed in US Pat.
No. 0,136 and U.S. patent application Ser.
The problem is being solved by the development of a new pattern recognition technology as disclosed in US Pat. Both of which are incorporated herein by reference, both of which are improvements to neural networks. The invention disclosed herein provides an excellent collision sensor for use with an airbag, irrespective of the pattern recognition technology used therein.

The pattern recognition collision sensor described above and shown in FIG. 1 utilizes information from three accelerometers for measuring accelerations in orthogonal directions. As will be described in more detail below, any other information such as occupant location, noise, data from acoustic or radar predictive sensors, or any other relevant information present in the vehicle can be considered in the pattern recognition algorithm. . Because pattern recognition algorithms are trained on data from real crash and non-collision events, they can handle data from many different sources, and what patterns require an airbag An event can be selected in a way that is not possible for a technician. For these reasons, a collision sensor based on a neural network will always perform better than, for example, a technician would think. Neural network theory can be found in several books on the subject, including many examples: "Techniques And Application O"
f Neural Networks ", Taylor, M.
And Lisboa, P, edited by Ellis Horwood, West Sussex,
1993; "Naturally Intelligent Systems", M Caudill, M. and Butler, C., MIT Press, Cambridge, Mass., 1990; "Digital Neural Network".
s ", Kung, SY, PTR Prentice Hall, Inglewood Cliffs, NJ, 1993.

A diagram of one example of a neural network used in a collision sensor designed according to the teachings of the present invention is shown in FIG. The process may be initiated and performed continuously, at the request of the computer system, when an event occurs that indicates an abnormal condition, for example, when the longitudinal acceleration exceeds the gravitational acceleration. You can also. As the digital acceleration values from the analog-to-digital converter enter node 1, the neural network algorithm compares the pattern of values from nodes 1 to N to the pre-trained pattern. The input nodes are second-layer nodes h-1,..., Each called a hidden layer. . . , H-n. This connection is made electrically, as in the case of a neural computer, as described below, or through a mathematical function that includes coefficients called weights, also as described below. This weight is determined while creating the neural network during training, as detailed in the above-referenced literature. At each hidden layer node, the values from each input layer node processed by the function including the weight are totaled to generate a node value. Similarly, hidden layer nodes are connected to output layer nodes. In this embodiment, the only node that represents the decision to deploy or not deploy the airbag is the output node. During the training period, an output node value of 1 is assigned, for example, meaning deploying an airbag, and an output node value of 0 is assigned.
Is assigned the meaning of not deploying the airbag. Details of this process are also described in the above reference and will not be elaborated here.

An analog-to-digital converter (ADC) averages the output from the accelerometer and sets a time step, such as 2 milliseconds, to send data to input node 1. Thus, at the time corresponding to 2 ms from the start of the process, node 1 will hold the value obtained from the ADC and the remaining input nodes will have random or zero values. At the time corresponding to 4 ms, the value of node 1 is sent to node 2 and the new value from the ADC is sent to node 1. Similarly, data is continuously transmitted from the ADC to the node 1, the data of the node 1 is transmitted to the node 2,
The previous value of node 2 is sent to node 3, and so on. Of course, it is not necessary to actually transfer the data to a different memory location, only that location need to be redefined so that the neural network can find the data of node 1. In one preferred embodiment of the invention, a total of 100 input nodes were used to represent 200 milliseconds of acceleration data. At each step, the neural network is evaluated,
If the value of the output node exceeds a certain value, for example 0.5, the airbag is deployed by the rest of the electronics. In this way, the system does not need to know when a collision has begun, i.e., does not require a separate sensor to determine the onset of a collision, or does not have any specific information about acceleration data to make such a determination. No algorithm is required.

In the above embodiment, 100 input nodes, 12 hidden layer nodes, and one output layer node are used. In this embodiment, only the longitudinal acceleration was considered. If other data is also used, such as vertical or lateral acceleration, the number of input layer nodes would have to be increased. If the neural network is used for rear impact or side impact sensing, two or three output nodes will be used, one for each decision. The theory for determining the complexity of neural networks for a particular application has been the subject of many technical papers and will not be discussed in detail here. The complexity required in the examples presented herein can be determined by one skilled in the art of neural network design, and will be described briefly below. In another implementation, the integral of the acceleration data is used, with a corresponding reduction in the number of input nodes.

The neural network described above defines a method of sensing a collision and determining whether to deploy a deployable occupant protection device, the method comprising:
(A) obtaining an acceleration signal from an accelerometer attached to a vehicle; (b) converting the acceleration signal into time-series digital data; and (c) inputting the time-series digital data to a node of a neural network. When,
(D) performing a mathematical operation on the data from each input node and inputting the processed data to a second node series, wherein the processed values are each input before inputting to the second series node. The processing performed on the input node data is different from the processing performed on some other input node data, and (e) processing the data from all the input nodes into the second (F) performing mathematical processing on each value of the second series node and inputting the processed data to an output series node. The processing performed on each second series node data before the processed value is input to the output node is different from the processing performed on some other second series node data. (G) mixing the processed data from all second series nodes towards each output series node to produce a value for each output series node; and (h) 1
Values are within a selected range, meaning that there is a collision that requires deployment of the airbag,
Deploying the airbag.

The particular neural network described and illustrated above includes a series of hidden layer nodes. In some network designs, more than one hidden layer is used, but it is extremely rare to use more than two such layers. Of course, there are many variations on the neural network structure shown above and these can be found in the literature. Thus, for the purposes of this specification, a "neural network" means that the data to be processed is separated into discrete values, which are then processed and mixed in at least one two-stage process, and the processing performed at each stage is usually , Defined at each discrete value, and the processing operations performed thereon are determined at least throughout the training process.

The implementation of the neural network can be performed in at least two ways: by an algorithm programmed on a digital microprocessor, or by an algorithm in a neural computer. Neural computer chips are now available. As more advanced pattern recognition technologies are developed, it can be expected that chips specifically designed for those technologies will also be developed.

In the particular implementation described above, the neural network was trained using collision data from about 25 collision and non-collision events. In addition, velocity and collision scaling techniques as described in the reference technical paper were used to create a large collision library representing many events that were not classified by the vehicle manufacturer. The resulting library is believed to represent most crash events in most real-world accidents in cars. Thus, the neural network algorithm approaches the goal of a general-purpose electronic one-point sensor that can be used in most, if not all, vehicles. FIG. 8 shows the result of applying the algorithm to the collision matrix created by the above-described velocity / collision scaling technique. An explanation of the meaning of the numbers in the table can be found in the above referenced technical paper entitled "Trends in Sensing Frontal Impacts", which is summarized below. FIG. 9 of that paper shows the results of an optimized cabin one-point sensor, which is repeated here in FIG. 9 for comparison with FIG. The sensors analyzed in that technical paper are employed in at least one mass-produced vehicle currently sold in the United States. This comparison shows that dramatic improvements are possible with the use of pattern recognition technology.

Once a pattern recognition computer system is employed in a vehicle, the same system can be used for many other pattern recognition functions such as airbag system diagnostics. Testing the pattern of the airbag system during a diagnostic test at start-up of the vehicle, for example, as represented by appropriate resistance values occurring across wiring to various system components,
It's easy with pattern recognition systems. In this way, the system can perform the functions of conventional SDM, i.e., sensing and diagnostics and other functions.

[0027] US patent application Ser. No. 08/2, referenced above.
No. 39,978 discloses the use of a neural network as a preferred pattern recognition technique for identifying a rear-facing child seat located in the front seat of an automobile. The patent application also discloses many other applications of pattern recognition in connection with monitoring the interior of a car interior.

FIG. 3 shows a view of the cabin of an automobile. A sensor for detecting that an occupant is not in a fixed position and a detector for detecting a rearward facing child seat are both mounted on an A-pillar. Use the same neural computer as the neural network collision sensor. Therefore, if one neural network is provided in an automobile, the same neural network computer can be used for several pattern recognition applications.

The use of the neural network to identify the back-type child seat is when the vehicle first moves. On the other hand, the use of a neural network for collision pattern recognition is continuous, but only required when an abnormal situation occurs. It is extremely rare that both occur at the same time, so the same system can easily perform both tasks. Naturally, when a conflict occurs, one of the functions is given priority. Therefore,
A powerful motivation for using neural network collision sensors is that, in addition to their excellent performance, they can also provide significant economic benefits. To date, no use of neural networks for in-vehicle surveillance or collision detection has been found, much less a significant advantage of combining both functions with the same neural network system. It has not been.

Interestingly, once a system is employed, each of the above features can be added at a very small cost. Distance measurements to measure occupants not in place can be made with minor software changes, and if an occupant system is already installed, adding a driver system requires only adding a transducer. Often, transducers are cheaper if purchased in bulk. Conversely, if the driver system is already installed, the addition of the occupant system can be achieved only by adding the converter. Since both the driver and occupant systems can share the same electronics, there will be little additional cost for electronic components.

Referring to FIG. 3, the aforementioned US patent application Ser.
As described in US Pat. No. 239,978, four ultrasonic transducers 3 are used to identify the object occupying the passenger seat.
11, 312, 313, and 314 are used. In this particular embodiment, the additional transducer 31
1 is provided to improve the verification accuracy of this system. Ultrasonic transducers 321 and 322 are used to determine the distance from the transducer to the car driver;
Airbag module 110 attached to handle
Ultrasonic transducers 323 and 324 are used to measure the distance to the driver. The second measurement relating to the driver's position is used to confirm the measurements obtained at converters 321 and 322. The distance from the airbag depends on the distance to the driver and the steering wheel 340.
Can be calculated by knowing the distance to Of course, other types of transducers or measurement means can be used without departing from the teachings of the present invention. What is illustrated and taught herein is that there are many applications that require pattern recognition technology, and that pattern recognition can be achieved very economically by sharing a pattern recognition computer device. is there.

Another example, FIG. 4, illustrates a side impact prediction sensor system, where transducers 411-419 are shown mounted at different locations on the side of the vehicle, which are discussed above. As such, the same computer system is used and coupled to the computer system by appropriate means. FIG. 4 also shows additional inputs, which provide information useful in some applications, for example, in determining whether to deploy an airbag. These inputs include inputs from a front mounted crash sensor 420 mounted on a radiator 430 of the vehicle, an engine speed sensor 440, and a wheel speed sensor 450 used in an anti-lock brake system sensor.

The collision avoidance system currently under development uses a radar to know the position of an object such as another vehicle on the expected route of the target vehicle. In some systems, a symbol is projected onto the glass surface of the head-up display to indicate the presence of an object in an area that may collide with the vehicle of interest. At this time, no attempt has been made to determine what the object is or to display an image of the object. Neural network pattern recognition systems, as well as other pattern recognition systems, have such capabilities and future collision avoidance systems will undoubtedly require this capability. Of course, as mentioned above, the pattern recognition computer system proposed herein for sensing collisions can also be used for collision avoidance pattern recognition. FIG.
Is a front view of the vehicle showing the position of the radar-based collision avoidance sensor 510. The radar-based collision avoidance sensor 510 uses the same neural computer system as the collision sensor discussed above.
0 is connected to the neural computer system.

FIG. 6 shows a headrest 610 of an automobile seat.
The occupant position sensor devices 622 and 624 provided in the vehicle are shown. Such a sensor automatically positions the headrest 610 to protect the occupant in the event of a rear impact, automatically adjusts the seat position based on the occupant's body shape, and also adjusts the head position during a front impact. Can be used as part of a position monitoring system. In each case,
The sensor is interfaced with the neural network computer system used for collision detection. For example, in the case of rear impact protection, the neural computer would use information from the accelerometer to determine that a rear collision was in progress and issue a command to move the headrest closer to the occupant's head. Become. If a predictive sensor is used to predict a rear impact, a neural computer can be used to identify the approaching object and determine whether to allow headrest positioning. If the vertical position of the occupant's head is monitored, the neural crash sensor will take this information into account, along with other occupant position information, and should the airbag be deployed if the occupant is not in place. Used to determine

The pattern recognition algorithm that forms an integral part of the collision sensor described herein can be implemented as a conventional microprocessor-based algorithm or using a neural computer that has become more recently available. In the former case, training is performed using a neural pattern recognition program, and the result is often a computer algorithm written in the C computer language. However, FORTRAN, assembl
Other computer languages such as y, Basic, etc. can be used. In the latter case, the same neural computer used in the vehicle can be used for training.
Neural network software for use on conventional microcomputers is available from several companies, such as Neural Ware, Pittsburgh, PA. An example of an algorithm created by Neuralware software is shown below. This algorithm is after training on a collision library built using data supplied by the vehicle manufacturer for a particular vehicle type and additional data created using collision and speed scaling techniques. * Neural network for collision sensors, August 1994
March 23. 50 input nodes, * 6 hidden nodes (sigmoid transfer function), 1 output node (value of 0 or 1). * Network is Logicon Projection (Logic
on Projection). * Yin (1-50) is the raw input value. Xin (1-5
0) is a scaled input value. * Yin (50) is the sum of the last 25 accelerations, a multiple of 1/10 of gravity. * Time step is 80 microseconds, such as Yin (49) being the previous 25 totals. logical function nnmtlpn3 (Yin, firesum, Yout) real * 4 firesum, Yin (50), Yout integer i, j real * 4 biashid (6), biasout, fire_criterion, hiddenout (6), NormV, NV (4), & offset_in (50), offset_out, scale_in (50), scale_out, wgthid (51,6), & wgtout (6), Xin (51), Xsum parameter (fire_criterion = 0.0) data scale_in / (omitted) / data offset_in / ( omitted) / data scale_out, offset_out / 0.625,0.5 / data NV / 2.0,7.0,7.0711002,50.000458 / data biashid / -49.110764, -69.856407, -48.670643, & -48.36599, -52.745285, -49.013027 / data biasout / 0.99345559 / data wgthid / (omitted) / data wgtout / (omitted) / NormV = 0.0 do i = 1,50 Xin (i) = scale_in (i) * Yin (i) -offset_in (i) NormV = NormV + Xin (i ) * Xin (i) enddo NormV = NV (1) * NV (2) * NV (3) / (NV (4) + NormV) do i = 1,50 Xin (i) = NormV * Xin (i) enddo Xin (51) = NV (2) -NV (3) * NormV do i = 1,6 Xsum = biashid (i) do j = 1,51 Xsum = Xsum + wgthid (j, i) * Xin (j) enddo hiddenout (i) = 1.0 / (1.0 + exp (-Xsum)) enddo firesum = biasout do i = 1,6 firesum = firesum + wgtout (i) * hiddenout (i) enddo Yout = offset_out + scale_out * tanh (fire sum) if (firesum.GE.fire_criterion) then nnmtlpn3 = .TRUE.else nnmtlpn3 = .FALSE. endif return end

Neural computers built into integrated circuit chips are now being supplied by various chip suppliers. These chips use a massively parallel processing architecture so that all input data can be processed simultaneously. As a result, the time required to process one pattern is on the order of milliseconds for a microprocessor implemented system, but on the order of microseconds for a neural computer. According to this calculation speed, a single neural computer can perform several pattern recognition tasks simultaneously even during a collision accident. A discussion of the structure of such a neural computer is provided at page 382 of the aforementioned Kun reference book.

FIG. 7 is a block diagram of a neural computer for obtaining a sensor algorithm. In a first step, one or more vehicles are collided under controlled conditions. Vehicles and dummy figures are fully equipped with measuring instruments, which can determine the severity of the collision and thus the need for airbags. The crash acceleration is measured at all locations where a crash sensor may be installed. Usually, a part firmly attached to a main structural member of a vehicle is a good attachment place for the sensor.

The following types of crashes at various speeds are typical to consider in establishing crash sensor design and calibration: Frontal impact on obstacles Angle right Collision with an obstacle held Collision with an obstacle angled to the left Frontal collision with an off-center obstacle Frontal collision with an off-center obstacle (outside the rail) At the center of a high pillar Off-center collision on a high column Off-center collision on a low column (below bumper) Vehicle-to-vehicle head-on collision Vehicle-to-vehicle partial head-on collision Vehicle-to-vehicle angled collision Vehicle-to-vehicle collision Vehicle collision Vehicle-to-vehicle side collision in which both vehicles are moving Bumper digging collision Animal collision-Simulation with deer Collision of underbody (Riding on something like a railway track) Highway shock absorber Yellow barrel collision to collision curb to the collision guardrail to the (yellow barrels), and the like)

The following non-collision event types are representative to be considered for establishing crash sensor design and calibration. Excessive use of the accelerator (hammer abuse) (Excessive use of the accelerator in the workplace (shop abuse)) Bad road (bad driving conditions)

Usually, a vehicle manufacturer focuses on a specific vehicle type only and instructs a crash sensor designer to design a sensor for the specific vehicle type. This is not necessary if the technique described in this specification is used, and data on collisions of various different vehicle types can be included in the training data.

Because crash tests are very expensive, vehicle manufacturers are reluctant to perform all of the crashes listed above for a particular vehicle type. On the other hand, if certain types of real-life incidents are missing from the library, the system may not work optimally if the incidents actually occur later, and one or more people will die. Or be injured. One way to partially solve this dilemma is to utilize the collision data of other vehicles as discussed above. Another approach is to create data using data obtained from a staged collision test and process the data using various mathematical techniques to represent untested collisions. Creating data.
One way to achieve this is to refer to the above referenced papers, especially pages 8 of Reference 1) and pages 37-49 of Reference 2),
Utilizing velocity and collision scaling as described in detail in This is the second step in the process shown in FIG.

The third step is to provisionally set a candidate neural network structure. The choice is moderately complex, for example having 100 input nodes and 6 hidden nodes. If the network is too simple, there will be cases in which the system will not be trained, and if that is a critical collision, the network will have to be modified by adding more nodes. If the initial choice is too complex, then typically one or more of the post-training weights will exhibit a value close to zero. Either way, the network will later test whether removing the nodes one by one will not reduce the accuracy of the network. Of course, a similar sequence of steps applies to other pattern recognition techniques.

The training data must then be assembled into a form as seen on a real vehicle. While data from pre-staged crashes is available for the entire time of the crash, the on-board system only looks at one data at a time. Therefore, the training data must be sent to the pattern recognition computer, or computer program, in such a manner. This is accomplished by retrieving each collision data file and making 100 data cases therefrom. In this case, this can be achieved by assuming the time of impact to be 200 milliseconds and taking each data point as the average acceleration over 2 milliseconds. The first training case includes the first point and the remaining 99 points are zero or a random small value. The second data case contains the first two data points, the remaining 98 points are zero, or a small random value, and so on. In the tenth data file, data point 1 contains the average crash acceleration at 20 ms, data point 2 contains the average crash acceleration at 18 ms, and the tenth data point contains the first 2 ms of the crash. Will include data from This process continues until 100 data cases have been created for the collision. Each case is represented as a line of data in the training file. This same process must be performed for collision and non-collision events with data, respectively. A typical training set will eventually contain as many as 50,000 data cases.

This training set is then substantially truncated, except for all cases where there is no need to deploy the occupant protection device. For a particular frontal obstacle collision, for example, at 30 mph, analysis of the collision indicates that the sensor must begin deploying the airbag within 20 milliseconds. Therefore, in the case of the above collision, it is not necessary to use data within 20 ms. Because we are not interested in whether the sensor should start deploying. Also, data longer than 20 milliseconds would mean deployment too late, which is also of little value. On the other hand, in cases where the airbag should not be deployed, the entire set of 200 data files must be used. Ultimately, this training set must be balanced so that the number of cases that do not start deployment and the number of cases that begin deployment are approximately the same, so that the output is not determined to one side. This is the fourth step in the process depicted in FIG.

In the fifth step, the pattern recognition program is run using this training set. In the case of a neural network program, the program is called "Back Propagation (Ba
ck Propagation) ”technology.
Weights are assigned to the connection from the input layer node to the hidden layer node and the connection from the hidden layer node to the output layer node so as to reduce the error between the calculated value at the output node and the required value. For example, for a particular collision, such as a 30 mph frontal obstacle collision, the collision analysis may indicate that the sensor
The deployment must start within milliseconds, and the data file representing the first 20 milliseconds of the collision must have a value of 1 as the required output node,
The fact is clear. If the deployment of the airbag is not required in another collision, such as an 8 mph obstacle collision, the required output node value for all data lines (100 lines) used to represent this collision is zero.
Becomes The network program will then have an output node value of almost zero in all cases where no airbag is needed.
, And similarly in all cases where an airbag is required, assign different weights to the nodes until the output value is close to one. This program finds the weight that minimizes the error between the required output value and the calculated output value.

The term weight is used in the art to describe the mathematical processing performed on the data at each node of each layer before the data is input to nodes at higher layers. There is a general term. For example, the data of the input layer node 1 is processed by a function including at least one coefficient defined by the training process. Generally, this coefficient, that is, the weight, differs depending on each combination of the input node and the hidden layer node. Thus, in the above example having 100 input nodes, 12 hidden layer nodes, and one output node, the number of weights determined by the neural network program during the training period is generally 1, It becomes 212 pieces. One example of a function that is used to process data from one node before entering it into higher hierarchy nodes is a sigmoid function: In a typical backpropagation training network, Oij is replaced by i-th node j
, The input to the node k in the i + 1 layer is Ii + 1, k = Ｗj Wkj (i) Oij where Wkj (i) is the node j in the i layer and the node k in the i + 1 layer.
Is the weight applied to connect between. So i +
The output of a layer k node k is obtained by transforming its input, for example with a sigmoid function: Oi + 1, k = 1 / (1 + e-1i + 1, k) and this is the input to the next layer i + 2 used.

If the neural network is very complex, that is, it has many hidden layer nodes, and if the training set is small, the network can "store" the training set, but the result is If it is slightly different, it may not be possible to respond properly. This is one of the problems with neural networks, which is being solved by more advanced pattern recognition systems. The ultimate goal is to have a network that resembles one of the cases presented in a generalized network of the presented data, but which responds appropriately to different new cases. The network can also effectively store input data if many cases are nearly identical. It can be difficult to determine this just by looking at the network, so rather than training the network with all available data, some important representative data samples can be used to test the network. It is important to keep it separate from the training set for use. This is the function of the fifth step, which tests the network using data that has not been seen before.

The sixth step is to redesign the network,
Next, the third to fifth steps are repeated until the result is satisfactory. This step is accomplished automatically by some commercially available neural network software products.

The final step is to output computer code for the algorithm and to program the microprocessor with this code. See above for examples of those codes. One of the important features of the present invention is that the neural network system chosen is very simple, but by the way the data is fed into the network, all the relevant calculations can be performed on a single network. It is. For example, it is not necessary to use an additional network to estimate the time to trigger. Because it is already in training. As a matter of course,
If the actual position of the occupant is known as described above, a significant improvement can be obtained, and in fact, two systems are introduced at the same time, whereby the best one-point crash sensor algorithm and It is more likely to be combined with a solution for a non-occupant.

The steps described above and shown in FIG. 7 are for the case where a neural computer program is used to generate the code, which is then used to program a standard micro-processor. Similar steps apply when a neural computer is used.

FIG. 8 illustrates a neural network pattern recognition algorithm designed in accordance with the teachings of the present invention for use in a one-point collision sensor, using the speed and collision scaling technique presented in the reference. The results are shown applied to a collision matrix based on this. The table contains the results for different collision speeds (vertical axis) and different collision times (rows). The results shown for each crash velocity and crash time combination are the displacement of the unconstrained occupant at the point in time at which airbag deployment begins and after 30 milliseconds.

[0052] The airbag is typically 3 inches before the occupant, represented by the dummy equivalent to the average male, has moved 5 inches.
It is designed on the assumption that a deployment time of 0 ms is provided. If the occupant is too close to the airbag when the airbag is deployed, this can cause the occupant to be severely injured or even killed by the deployment of the airbag. If the occupant is leaning on the airbag when the airbag deploys, this is equivalent to an average male occupant traveling about 12 inches, which is known to be a very serious condition; It is also known that if the occupant moved less than 5 inches when the airbag was fully deployed, the occupant would not be injured by the deployed airbag. These figures are based on a dummy representing an average male sitting in the middle of the seat, a dummy called a 50% male dummy. Thus, the critical injury threshold is somewhere between these two points, and therefore, in this table, two benchmarks have been selected to approximate the serious injury threshold. These benchmarks are based on unconstrained occupant movement: (i) whether the occupant has already moved 5 inches when airbag deployment begins,
And (ii) whether the occupant has moved 12 inches before the airbag is fully deployed. Both benchmarks demonstrate that the occupant is substantially engaged with the airbag when the airbag is deployed. Of course, other benchmarks could be used, but it is believed that they lack sufficient test results to show them.

Thus, the tables shown in FIGS.
It presents data on the relative displacement of the occupant with respect to the airbag at the start of deployment and after 30 ms. If the first number is 5 inches or more, or the second number is 12 inches or more, there is a risk of serious injury, which means that the sensor did not catch up in triggering the airbag. Will be. In these cases, the cells in the table are shaded. FIG. 8 is for a neural network collision sensor designed according to the teachings of the present invention, but as seen here, there is no shadow in any of the squares and the sensor performance is considered to be good. .

The table shown in FIG. 9 shows a model of a one-point collision sensor used in some types of vehicles currently mass-produced. This sensor is in fact designed to be optimized for collisions as shown in the table. As shown in FIG. 9, this sensor failed in a timely manner to deploy the airbag due to a significant number of the crashes shown in the table.

Another feature of using the neural network collision sensor of the present invention is that when making a decision to deploy an airbag, in the above example, 2
00 milliseconds of collision data is stored at the input node of the network. This provides a kind of "black box" that can be used later to pinpoint the severity of the collision.

Although several preferred embodiments have been illustrated and described, other shapes, materials, and different dimensions for the components may be used to implement other forms of neural networks that perform the same function. It is possible to combine. Also, a neural network has been described as an example of the pattern recognition means, but other pattern recognition means also exist. Further measures have been developed to process each data point individually to determine if there are any identifiable patterns, and to use them to identify collisions that require deployment of occupant protection systems. And will be available soon.
Also, in the neural network system described above, the input data to the network either goes through a "feature extraction" process, such as described in Green U.S. Pat. No. 4,906,940, or integrates the data. The raw acceleration data may be preprocessed by any method, such as inputting speed data to the system. The present invention is not limited to the above embodiments, but should be determined by the appended claims.

[0057]

The effects of the present invention are as follows. 1) A single point sensor can be provided that includes an accelerometer that makes the best use of the information in the acceleration data to determine whether to deploy the airbag. 2) It is possible to provide a general-purpose one-point collision sensor that can be used in most vehicles without making a modification unique to each vehicle. 3) In the car, several different pattern recognition functions,
For example, it is possible to provide a single computer system capable of detecting a collision, confirming an object existing in a vehicle interior, and classifying an object outside the vehicle. 4) A gradual vehicle crash and non-collision event, as well as other statistically derived data sets, can be used to provide training-derived crash sensor algorithms. 5) It is possible to provide a collision sensor based on a pattern recognition technology. 6) Provide a collision sensor that uses other data in addition to acceleration data derived from a collision. The data is combined with the acceleration data and the use of pattern recognition techniques can determine the need for passive protection device deployment. 7) Can provide a crash sensor that can be used to sense both frontal, side and rear collisions. 8) It is possible to provide a collision sensor that automatically retains collision acceleration data from a predetermined time before the deployment of the airbag for later analysis. 9) A collision sensor using a neural computer can be provided.

[Brief description of the drawings]

FIG. 1 is a partially cut-away view showing a front portion of a vehicle compartment of an automobile having a dual airbag and a single-point collision sensor including an accelerometer and utilizing a pattern recognition technique;

FIG. 2 is a diagram of a neural network used in a collision sensor designed according to the teachings of the present invention.

FIG. 3 is a view in which a part showing a cabin of an automobile is notched and removed, and shows a sensor for detecting that the occupant is not in a fixed position and a detector for detecting a rear-type child seat; Both are attached to the A pillar,
Both use the same computer as the collision sensor based on pattern recognition.

FIG. 4 is a partially top perspective view of a vehicle showing a side impact predictive sensing system using the same computer as a single point crash sensor, with a front mounted crash zone sensor, engine speed sensor, and antilock. Inputs from brake system sensors are also shown.

FIG. 5 is a front view of an automobile showing the location of a radar-based collision avoidance sensor using the same pattern recognition computer system as the collision sensor.

FIG. 6 is a perspective view of a vehicle seat and headrest including an ultrasonic head position sensor composed of one transmitter and one receiver.

FIG. 7 is a block diagram illustrating a method for obtaining a sensor algorithm using a neural network.

FIG. 8 shows the results obtained with a neural network algorithm for a collision matrix created using the velocity-collision scaling technique.

FIG. 9 shows the results obtained with a standard one-point collision sensor for a collision matrix made using the velocity-collision scaling technique.

[Explanation of symbols]

 110, 111 Airbag 120 SDM (Electronic Sensing and Diagnostic Module) 122 Accelerometer 124 Analog-to-Digital Converter 130 Microprocessor 311-314, 321-324 Ultrasonic Converter 411-419 Converter 420 Collision Sensor 510 Radar Collision Avoidance sensors 622, 624 Occupant position sensor device

Claims (24)

[Claims] 1. An automobile having a deployable occupant protection device for protecting at least one occupant in the event of a collision, comprising: a sensor system for initiating deployment of the occupant protection device; Sensor means for sensing acceleration and generating an analog signal representing the acceleration;
(B) mounting means for attaching the sensor means to the vehicle, (c) converting means for receiving the analog signal and converting the analog signal into a digital signal, and (d) wrapping pattern recognition means, and (E) processing means for receiving the processing means for generating a deployment signal when the pattern recognition means determines that the digital signal contains a pattern specific to a collision of a vehicle requiring occupant protection; And a deployment start means coupled to the processing means for initiating deployment of the occupant protection device in response to the deployment signal. 2. The sensor system according to claim 1, wherein said pattern recognition means includes a neural network. 3. The sensor system according to claim 1, wherein the conversion means derives the digital signal from an integrated value of the analog signal. 4. The sensor system according to claim 1, wherein the sensor means is mounted at a location on the vehicle where frontal impact is sensed. 5. The sensor system according to claim 1, further comprising: detecting when the occupant to be protected by the deployable occupant protection device has moved from a predetermined position, and based on the detected time, A sensor system comprising detection means for suppressing deployment, said detection means including said processing means. 6. The rearward child according to claim 1, wherein the vehicle has a passenger seat and a passenger side airbag occupant protection device, and the sensor system is further located in the passenger seat. A sensor system including detection means for detecting the presence of a sheet, wherein the detection means includes the processing means. 7. The sensor system according to claim 1, wherein said sensor means includes a predictive sensor. 8. The sensor system according to claim 1, wherein said sensor means includes a sensor for a collision avoidance system. 9. The sensor system according to claim 1, wherein the sensor means is mounted at a position for detecting a rear impact. 10. In a motor vehicle having a deployable occupant protection device for protecting at least one occupant during a collision, determining that the vehicle has encountered a collision that requires deployment of the occupant protection device. A sensor for sensing an acceleration of the vehicle and generating an analog output signal representing the acceleration; (b) a mounting means for attaching the sensor means to the vehicle; Conversion means for receiving and converting the analog signal to a digital signal; and (d) a neural computer coupled to the conversion means for processing the digital signal and including pattern recognition means, requiring occupant protection. When the pattern recognition means determines that a pattern specific to a vehicle collision is included in the digital signal, a deployment signal is generated. Neural computer,
(E) output means coupled to the neural computer for initiating deployment of the occupant protection device in response to the deployment signal. 11. The sensor according to claim 10, wherein:
A sensor wherein the occupant protection device is an airbag and the collision is a forward collision. 12. The sensor according to claim 10, wherein:
A sensor wherein the occupant protection device includes a movable headrest and the collision is a rear collision. 13. The sensor according to claim 10, wherein:
And detecting means for detecting when the occupant to be protected by the deployable occupant protection device has moved from a predetermined position and for suppressing deployment of the occupant protection device based on the detection. A sensor system including: 14. The sensor according to claim 10, wherein:
The vehicle has a passenger seat, a passenger seat side airbag occupant protection device, and a detecting unit that detects the presence of a rearward facing child seat located in the passenger seat and suppresses deployment of the occupant protection device based thereon. A sensor, wherein the detecting means includes the neural computer. 15. The sensor according to claim 10, wherein:
A sensor comprising an acceleration measurement system wherein said sensor means measures acceleration in at least two directions. 16. The sensor according to claim 10, wherein:
Additional data is input to the neural computer,
A sensor in which the pattern recognition unit uses the additional data. 17. The sensor according to claim 16, wherein:
The sensor wherein the additional data includes data from a predictive sensor. 18. The sensor according to claim 16, wherein:
The sensor wherein the additional data includes data from a collision avoidance sensor. 19. The sensor according to claim 10, wherein:
A sensor for the neural computer to diagnose a responsive state of the device. 20. A method of obtaining an algorithm for use with a microprocessor-based crash sensor to determine whether a deployable occupant protection device in a vehicle should be deployed, comprising: (a) C) obtaining representative digital crash data for the vehicle using the crash sensor, wherein the crash data is obtained from a crash that requires deployment, a crash that does not require deployment, and other events. Wherein all such collision and event combinations constitute a collision library for the vehicle; (b) designing a candidate pattern recognition algorithm; and (c) the collision A pattern recognition computer program and a collision library are used to classify library events into events that require Training said pattern recognition algorithm until a classification error is minimized to obtain a trained neural network;
Testing the trained pattern recognition algorithm using representative additional collisions and events for the vehicle; and (e) redesigning the pattern recognition algorithm if the test performance is unsatisfactory; Repeating test steps (c) and (d) ((Note: the examiner may reject this step as conditional, so I recommend unconditionally replacing it with a dependent claim) And (f) outputting the obtained collision sensor algorithm from a pattern recognition program. 21. A method of sensing a vehicle collision for determining whether a deployable occupant protection device in a vehicle should be deployed, comprising: (a) detecting acceleration from an accelerometer mounted on the vehicle; Obtaining a signal; (b) converting the acceleration signal into digital time series data; and (c) inputting the digital time series data to a first series node which is an input node of a neural network.
(D) performing mathematical processing on the data from each of the first series nodes as the input nodes, and inputting the processed data to a second series of nodes, wherein the processed data is stored in a second series of nodes. The processing to be performed on the data from each of the first series nodes that are the input nodes before inputting to the series nodes is the processing to be performed on the data from the other ones of the first series nodes that are the input nodes. (E) mixing processed data from all input nodes toward each second series node to create a value for each second series node; and (f) Performing a mathematical operation on each of the values of the two series nodes and inputting the processed data to an output series node, wherein the processed values are input to an output series node A step wherein the processing performed for each of the second series node data, the processing to be performed for other second series node data are different, the
(G) mixing the processed data from all second series toward each output series node to create a value for each output series node; and (h) if the value of one output series node is an airbag Deploying the airbag when it is within a range meaning that a collision requiring deployment is in progress. 22. The method of claim 21, wherein a third series node is located between the second series node and the output series node and the processed data from the second series node. Is input to the third series node, and the processed value from the third series node is input to the output node. 23. A motor vehicle having an occupant protection device deployable to protect occupants of at least two vehicles and having a driver seat occupied by a driver and a passenger seat occupied by an object. A sensor system for initiating deployment of the occupant protection device, comprising: (a) first sensor means for sensing acceleration of the vehicle and generating an analog output signal representing the acceleration; To detect the relative position of the driver to the occupant protection device
Sensor means; (c) confirmation means for confirming an object occupying the passenger seat; (d) mounting means for attaching the first sensor means to the vehicle; and (e) coupling to the first sensor means. Conversion means for receiving the analog signal and converting the analog signal into a digital signal;
(F) arithmetic processing means coupled to the conversion means and including pattern recognition means, and (i) a pattern specific to a collision of a vehicle requiring occupant protection is included in the digital signal. The pattern recognition means judges,
(Ii) the second sensor means indicates that a driver is at least a predetermined distance away from the deployable occupant protection device; and (iii) the front passenger seat is occupied, which is a rear child seat. A processing means for generating a deployment signal when the confirmation means determines that the occupant protection device is not occupied by the occupant protection device; And a deployment start means for starting the operation. 24. The system according to claim 23, wherein said pattern recognition means comprises a neural network.