A SMART AIRBAG SYSTEM use with anticipatory sensing systems to identify threateningobjects, such as an approachingvehicle about David S. Breed to impact the side of the vehicle. Neural networks have Automotive Technologies International, Inc. also been applied to sense automobile crashes for the United States purposeof determiningwhether or not to deploy an airbag Paper Number: 98-%-O- 13 or otherpassive restraint, or to tighten the seatbelts,cutoff the fuel system, or unlock the doors after the crash. ABSTRACT Heretofore, neural networks have not been applied to forecastthe severity of automobilecrashes for the purpose Pattern recognition techniques, such as neural of controlling the flow of gas into or out of an airbag in networks, have beenappiied to identify objects within the order to tailor the airbag inflation characteristicsto the passengercompartment of the vehicle, such as a rear crash severity. Neural networks have also not beenused facing child seat or an out-of-position occupant, and to to tailor the airbag inflation characteristicsto the size, suppressthe airbagwhen an occupantis more likely to be position or relative velocity of the occupant or other injured by the air-bagthan by the accident. Neural factors such as seatbelt usage, seat and seat back networks have also been applied to sense automobile positions,headrest position, vehicle velocity, etc. crashes. The use of neural networks is extendedhere to “Pattern recognition” as usedherein meansany system tailoring the airbag inflation to the severity of the crash, which processesa signal that is generatedby an object. or the size, position andrelative velocity of the occupantand is modified by interacting with an object, in order to other factors such as seatbeltusage, seat and seat back determinewhich one of a set of classesthe object belongs positions, vehicle velocity, and any other relevant to. In this case, the object can be a vehicle with an information. accelerometerwhich generatesa signal based on the It is well known that a neural network based crash deceleration of the vehicle. Such a system might sensor can forecast, basedon the first part of the crash determineonly that the object is or is not a memberof one pulse, that the crash wilI be of a severity which requires specified class (e.g., airbag required crashes),or it might that an airbag be deployed. This is extended here to attempt to assign the object to one of a larger set of enhancethe capabilities of this sensor to forecast the specifiedclasses, or find that it is not a memberof any of velocity changeof the crash over the entire crash period. the classes in the set. One such class might consist of Then a patternrecognition occupant position and velocity vehicles undergoinga crash of a certain severity into a determination sensor is added. Finaiy, an occupant pole. The signals processed are generaIly electrical weight sensor is inctuded to permit a measure of the signals coming from transducerswhich are sensitive to occupant’s momentum or kinetic energy. The either acceleration, or acoustic or electromagnetic combination of these systems in various forms will be radiation and, if electromagnetic, they can be either usedto optimize inflation and/or deflation of the airbagto visible light, intiared, ultraviolet or radar. The particular createa “‘smartairbag ” system. pattern recognition techniques used here are neural Crash sensorscan predict that a crash is of a severity networks. which requires the deployment of an airbag for the To “identifjr” as usedherein meansto determinethat majority of real world crashes. A more difficult problem the object belongsto a particular set or class. The class is to predict the crash velocity versus time function and may be one containing all frontal impact air-bag-desired then to adjust the airbag inflation/deflation over time so crashesinto a pole at 20 mph, one containing all events that just the proper amount of gas is in the airbag at all where the airbag is not required, or one containing all times even without considering the influence of the eventsrequiring a triggering of both stagesof a dual stage occupant. To also simultaneouslyconsider the occupant gas generatorwith a 15 millisecond deiay between the size, weight, position and velocity renders this problem triggering of the first and secondstages. unsolvableby conventionalmethods. SINGLE POINT CRASH SENSORS BACKGROUND All electronic crash sensorscurrently used in sensing Patternrecognition techniques, such as artificial neural frontal impacts include accelerometersthat detect and networks, are finding increasedapplication in solving a measurethe vehicle accelerationsduring the crash. The variety of problems such as optical characterrecognition, accelerometerproduces an analog signal proportional to voice recognition, and military target identification. In the acceleration experiencedby the accelerometer,and the automotive industry, neural networks have now been hencethe vehicle on which it is mounted. An analog to applied to engine control and to identify various objects digital converter(ADC) transformsthis analog signal into within the passengercompartment of the vehicle, such as a digital time series. Crash sensordesigners study this a rear facing child seat. They have also beenproposed for digital accelerationdata and derive therefrom computer
1080 algorithms that determine whether the accelerationdata position, size and weight of the occupant. More Tom a particular crash event warrants deploymentof the particularly, no such system exists which uses pattern airbag. This is usually a trial and error processwherein recognition techniquesto match the airbagdeployment or the crash sensor designer observes data from crashes gas dischargefrom the airbag to the severity of the crash where the airbag is desiredand when it is not needed,and or the size, weight, position, velocity and seatbeltuse of other eventswhere the airbag is not needed. Finally, the an occupant. crash sensordesigner settles on the “rules” for controlling Since there is insufficient information in the deploymentof the airbag which are programmedinto an acceleration data, as measured in the passenger algorithm which appearsto satisfy the requirementsof the compartment,to senseall crashesand since some of the crashlibrary. The resulting algorithm is not universaland failure modesof published single point sensoralgorithms most such crash sensor designers will answer in the can be easily demonstratedusing the techniquesof crash negative when asked whether their algorithm will work and velocity scaling describedin the referencedtechnical for all vehicles. Such an algorithm also merely papers, and, moreover, since the process by which determines that the airbag should or should not be engineersdevelop crash sensor algorithms is based on triggered. Heretofore, no attempt has been made to trial and error, pattern recognition techniques such as ascertainor forecastthe eventualseverity of the crash or, neural networks, should be able to create an algorithm more specifically, the velocity changeversus time of the basedon training the systemon a large number of crash passenger compartment during the crash from the and non-crashevents which, althoughnot perfect, will be accelerationdata obtained from the accelerometer. superior to all others. Such a crash sensor has been Severalpapers have beenpublished pointing out some demonstratedwhich is based on the ability of neural of the problemsand limitations of electronic crashsensors networks to forecast, basedon the first part of the crash that are mounted out of the crush zone, usually in a pulse, that the crashwill be of a severity requiring airbag protected location in the passengercompartment of the deployment. vehicle (I-6). The crush zone is defined, for the purposes herein, as that portion of the vehicle that has crushedat SMART AIRBAG CATEGORIES the time that the crash sensormust trigger deploymentof the restraint system. These sensorsare frequently called An improvementto this neural network based crash single point crashsensors. sensor carries this process fiuther by using the neural These papers demonstrate,among other things, that network to forecastthe velocity changeof the crash over there is no known theory which allows an engineer to time so that the inflation and/ordeflation of the airbag can develop an algorithm for sensingcrashes and selectively be optimized. Then by the addition of a neural network deploying the airbag except when the sensoris located in occupant position and velocity determination system as the crush zone of the vehicle. Thesepapers show that, in disclosed in (lo,1 I) the occupancy category (forward general, there is insufficient information within the facing human, rear facing child seat, box etc.), position acceleration signal measured in the passenger and velocity can be obtained. Finally, the addition of the compartment to sense all crashes. Another conclusion weight of the occupant provides a measure of the suggestedby thesetechnical papersis that if an algorithm occupantskinetic energyas a further input to the system. can be found which works for one vehicle, it will also The combination of these sub-systemsin various forms work for all vehicles since it is possible to create any can be called “smart airbags” or “smart restraints”. In a crash pulse measuredin one vehicle, in any other vehicle. preferred implementation, the crash severity is not Note in particular SAE paper920 124 (3). explicitly forecasted. Rather, the value of a control In spite of the problems associatedwith finding the parameterused to corm-01the flows of inflator gas into or optimum crash sensor algorithm, many vehicles on the out of the airbagis insteadforecasted. road today have electronic single point crash sensors. Smart airbags can take several fbrms which can be Someof the problemsassociated with single point sensors roughly categorizedinto four evolutionary stages,which have the result that an out-of-position occupant who is will hereinafterbe referred to as Phase 1 (2,3,4) Smart sufficiently close to the airbag at the time of deployment Airbags, as follows: will be injured or killed by the deployment itself I> Occupantsensors use various technologiesto turn off Fortunately, systems are now being developed which the airbag where there is a rear facing child seat monitor the location of occupantswithin the vehicle and present or if either the driver or passengeris out-of- can suppressdeployment of the airbag if the occupant is position to where he/she is more likely to be injured more likely to be injured by the deploymentthan by the by the airbagthan from the accident. accident. At Present,these systems do not provide the 2) Occupant sensors are used along with variable information necessaryfor the control of airbag systems inflation or deflation rate airbags to adjust the that have the capability of varying the flow of gas into or inflation/deflation rate to match the occupant,first as out of the airbag, and thus to tailor the airbag to the to his/her position and then to his/her morphology.
1081 The neural network occupant sensors discussed in data sets or vectors as every combination of occupancy (10,ll) will also handle this with the addition of an characteristics and acceleration segment is considered. occupant weighing system. One particular weight Fortunately, the occupancy data can be acquired measuring system, for example, makes use of strain independentlyand is currently being done for solving the gagesmounted onto the seat supportingstructure. At occupant position sensing problem of Phase I smart the end of this phase, little more can be done with airbags. The crash data is available in abundanceand occupantmeasurement or characterizationsystems. more can be analytically created using the crash and 3) The next improvementis to use a neural network as velocity scaling techniques described in the referenced the basis of a crashsensor not only to determineif the papers. The training using combinationsof the two data airbagshould be deployed,but also to predict the crash sets, which must also take into account occupantmotion severity from the pattern of the initial portion of the that is not adequatelyrepresented in the occupancydata crashpulse. Additionally, the crashpulse can continue can then be doneby computer. to be monitoredeven after the decisionhas beenmade to deploy the airbag to seeif the initial assumptionof CRASH SEVERJTY FOR!XIASTIlW the crash type, based on the pattern up to the deployment decision, was correct. If the pattern When a crash commences, the vehicle starts changesindicating a different crashtype, the flow rate decelerating and an accelerometer located in the to the airbagcan be alteredinstantaneously. passengercompartment begins sensing this deceleration 4) Finally, anticipatory sensing using neural networks and producesan electronic signal that varies over time in can be used to identify the crash before it takes place proportion to the magnitude of the deceleration. This and selectthe deploymentcharacteristics of the airbag signal containsinformation as to the type of the crashthat to match the anticipatedcrash with the occupantsize, can be used to identify the crash. A crash into a pole position, velocity etc.. gives a different signal than a crashinto a rigid barrier, for Any of these phases can also be combined with example,even during the early portion of the crashbefore various methods of controlling the pretensioning, the airbag triggering decision has been made. A neural retraction or energy dissipation characteristics of the network pattern recognition system can be trained to seatbelt. Although the main focus here is the control of recognize and identify the crash type from this early the flows of gas into and out of the airbag, the control of signal, and other available information such as vehicle the seatbeltcan also be accomplishedand the condition of speed,and further to forecast aheadthe velocity change the seatbelt can be valuable input information into the versustime of the crash. Once this forecastis made, the neuralnetwork system. severity and timing of the crash can be predicted. Thus, The smart airbag problem is complex and difficult to for a rigid barrier impact, for example,an estimateof the solve by ordinary mathematicalmethods. Looking first at eventualvelocity changeof the crashcan be madeand the the influence of the crash pulse, the variation of crash amount of gas needed in the airbag to cushion an pulses in the real world is vast and quite different from occupant as well as the time available to inject that the typical crashes run by the automobile industry as amount of gas into the airbag can be determinedand used reported in the referenced technical papers. It is one to control the airbaginflation. problem to predict that a crash is of a severity level to Alternately, consider a crash into a highway energy requirethe deploymentof an airbag. It is quite a different absorbingcrash cushion. In this case,the neural network problem to predict exactly what the velocity versus time basedsensor determines that this is a very slow crash and function will be and then to adjust the airbag causesthe airbagto inflate more slowly thereby reducing inflation/deflation control system to make sure that just the incidence of collateral injuries such as broken arms the properamount of gas is in the airbagat all times, even and eyelacerations. without considering the influence of the occupant. To In both of these cases, the entire decision making also simultaneously consider the influence of occupant process takes place before the airbag deployment is size, weight, position and velocity renders this problem, initiated. In another situation where a soft crash is for all practical purposes, unsolvable by conventional precededby a hard crash, such as might happenif a pole methods. were in front of a barrier, the neutral network system On the other hand, if a neural network is used and would first identify the sofi pole crash and begin slowly trained on a large variety of crash accelerationsegments, inflating the airbag. However, once the barrier impact and a setting for the inflation/deflation control system is began, the system recognizes that the crash type has specified for each segment, then the problem can be changedand recalculatesthe amount and timing of the solved. Furthermore,inputs from the occupantposition introduction of gas into the airbag and sendsappropriate and occupantweight sensorscan also be included. The commandsto the inflation control systemof the airbag to result will be a training set for the neural network increasethe introduction of gas into the airbag. involving many millions, and perhapstens of millions, of
1082 VARIABLE INFLATORS trained on data from real crashesand non-crashevents, it can handle data Corn many different information sources Thereare many ways of controlling the inflation of the and sort out what patterns correspondto airbag-required airbag and several are now under development by the events in a way which is nearly impossible for an inflator companies. One way is to divide the airbag into engineerto do. For this reason,a crash sensorbased on different charges and to initiate these charges neural networks, for example,will always perform better independentlyas a function of time to control the airbag than one devised by engineers. The theory of neural inflation. An alternative is to always generate the networks including many examples can be found in maximum amountof gas but to control the amount going severalbooks on the subjectincluding (7-9). into the airbag, dumping the rest into the atmosphere. A The process can be programmedto begin when an third way is to put all of the gas into the airbagbut control event occurs which indicatesan abnormalsituation such the outflow of the gas from the airbag through a variable as the acceleration in the longitu~mal direction, for vent valve. For the purposes herein, all controllable example, exceedingthe accelerationof gravity, or it can apparatusfor varying the gas flow into or out of the take place continuouslydepending on the demandson the airbag over time will be considered as a gas control computersystem. The digital accelerationvalues from the modulewhether the decision is made at the time of initial ADC may be preprocessed,as for example by filtering, airbag deployment,at one or more discretetimes later or and then entered successiveiy into the neural network continuouslyduring the crashevent. algorithm which comparesthe pattern of values on nodes 1 through N with patternsfor which it has been trained. INTEGRATION Each of the input nodesis connectedto eachof the second layer nodes h-l,...,hn, called the hidden layer, either The use of neural networks in crash sensors has electrically as in the caseof a neural computer,or through anothersignificant advantagein that it can sharethe same mathematical functions containing multiplying hardwareand software with other systemsin the vehicle. coefficients called weights. Neural networks have proven to be effective in solving The weights are determinedduring the training other problems related to airbag passive restraints. In phase while creating the neural network as described in particular, the identification of a rear-facing child seat detaii in the text references. At eachhidden layer node, a locatedon the front passengerseat, so that the deployment summation occurs of the values Tom each of the input of the airbag can be suppressed,has been demonstrated. layer nodes, which have been operatedon by functions Also, the use of neural networks for the classification of containingthe weights, to createa node value. Similarly, vehicles or objectsabout to impact the side of the subject the hidden layer nodes are comected to the output layer vehicle for use in anticipatory side impact crash sensing nodes, which in this example is only a single node shows great promise. Both of these neural network representingthe control parameterto be sent to the gas systems,as well as othersunder development, can use the control module. If this value exceedsa certain threshold samecomputer systemas the crash sensorand prediction the gascontrol module initiates depIoymentof the airbag. system. Moreover, both of these systems will need to During the training phase,an output node value interactwith, and shouldbe part of, the diagnosticmodule is assignedfor every setting of the gas control module used for frontal impacts. It would be desirable for cost correspondingto the desired gas flow for that particular and reliability considerations, therefore, for all such crash as it has occurredat a particular point in time. As systems to use the same computer system. This is the crash progressesand more accelerationvalues appear particularly desirable since computersdesigned specially on the input nodes, the vahe of the output node may for solving neural network problems, such as neural- change. In this way, as long as the crashis approximately computers,are now available. representedin the training se$the gas flow can be varied at eachone OFtwo milliseconds dependingon the system THE NEURAL NETWORK SYSTEM designto optimally match the quantity of gas in the airbag to the crash as it is occurring. Similarly, if an occupant The neural network crash sensordescribed is capable sensorand a weight sensorare present,that information of using information from three accelerometers,each can additionally be fed into a set of input nodesso that the measuringacceleration I?om an orthogonaldirection. As gas module can optimize the quantity of gas in the airbag will be describedin more detail below, other information taking into account both the crash decelerationand also can also be consideredby the neural network algorithm the position, velocity, size and weight of the occupantto such as the position of the occupants,noise, data horn optimally deploy the airbag to minimize airbag induced anticipatory acoustic, radar, inbed or other injuries andmaximize the protectionto the occupant. The electromagneticsensors, seat position sensots, seatbelt details of the neuralnetwork processand how it is trained sensors,speed sensors, or any other infmation presentin are describedin referencedtexts and will not be presented the vehicle which is relevant. Since the algorithm is in detail here.
1083 A time step such as two milliseconds is selectedas the OCCUPANT MOMTORWG SYSTEM period in which the ADC preprocessesthe output from the accelerometersand feeds data to input node 1. Thus, using this time step, at time equal to 2 milliseconds horn the start of the process,node 1 contains a value obtained from the ADC and the remaining input nodes have a random value or a value of 0. At time equal 4 milliseconds,the vaIue which was on node 1 is transferred to node2 and a new value Tom the ADC is fed into node 1. In a similar manner,data continuesto be fed from the ADC to node 1 and the data on node 1 is transferredto node 2 whose previous value was transferred to node 3 etc.. Naturally, the actual transfer of data to different memory locations need not take place but only a redetinition of the location which the neural network should find the data for node 1. For one case,a total of Fire 1. Occupant monitoring system one hundred input nodes were used representing two hundredmilliseconds of accelerationdata. At each step, Figure 1 illustratesan occupantmonitoring systemthat the neural network is evaluatedand if the value at the is capableof identifying the occupancyof a vehicle and output node exceeds some value such as .5 then the measuringthe location and velocity of human occupants. airbags are deployed by the remainder of the electronic This system is now being developedfor implementation circuit. In this manner,the systemdoes not needto know on a production vehicle. In one implementation, four when the crash begins, that is, there is no need for a ultrasonic transducers are used to provide accurate separatesensor to determinethe start of the crash or of a identification and position monitoring of the passengerof particular algorithm operatingon the accelerationdata to the vehicle. Naturally, a similar system can be make that determination, implementedon the driver side. The systemis capableof In the example above, one hundred input nodes were determiningthe pre-crashlocation of the critical parts of used, twelve hidden layer nodes and one output layer the occupant,such as his/her head and chest, and then to node. In this example, accelerations liom only the track their motion toward the airbag with readingsas fast longitudinal direction were considered. If other data such as once every 10 milliseconds. This is sufficient to as accelerationsfrom the vertical or lateral directions determinethe position and velocity of the occupantduring were also used, then the number of input layer nodes a crash event. The implementation described can would increase. If the neural network is to be used for therefore determine at what point the occupant will get sensingrear impacts, or side impacts,2 or 3 output nodes sufficiently out-of-position so that deployment of the might be used, one for each gas control module. The airbag should be suppressedas in solving the standard theory for determiningthe complexity of a neuralnetwork occupantsensing problem. Alternately, the information is for a particular application has been the subject of many used to determinehow fast to deploy the airbag. If the technical papersand will not be presentedin detail here. weight of the occupantis also known, the amount of gas Determining the requisite complexity for the example which should be injected into the airbag and perhapsthe presentedhere can be accomplishedby those skilled in out flow resistance can be controlled to optimize the the art of neural network design and is discussedbriefly airbag systemnot only basedon the crash pulse but also below. In another implementation, the integral of the the occupant properties. This provides the design for accelerationdata is used and it has been found that the Phase3 Smart Airbags. numberof input nodescan be significantly reducedin this Although the system illustrated uses ultrasonic manner. transducers, other systems use a variety of other The particular neuralnetwork describedand illustrated technologies including electromagnetic(optical, passive above contains a single series of hidden layer nodes. In or active infrared, radar), capacitive, seatbeltswitch, seat somenetwork designs,more than one hidden layer is used and seatbacklocation transducers,weight sensorsand in although only rarely will more than two such layers fact any sensing system which can provide relevant appear. There are of coursemany other variations of the information. The neural network is the ultimate “sensor neural network architecture illustrated above which fusion” technology and can use any type of sensorsand appearin the literature. will provide the system designer with a quantitative measureof the importance of any of the sensors. The optimum combination of four sensors, for example, might be one active infrared sensor,two ultrasonic sensorsand a single strain gageweight sensor. The initiaI investigation
1084 might have included, four ultrasonic sensors,two active determinewhat that object is and to display an image of infrared sensors fwr weight sensors, a scat position the object. Neural network pattern recognition systems sensor,a seatbackposition sensor,and a seatbeltbuckle have that capability and Wure collision avoidance sensor. A cost benefit analysiscan easily be performedto systemsmay needthis capabifity. Naturally, as above,the determine the effect of adding any particular additional sameneural network computer systemwhich is proposed sensorto the system. herein for sensingcrashes can also the used for collision avoidanceneural network as well as anticipatory sensing. ANTfCIPATORY SENSING OPERATION OF THE SYSTEM
1. Obtain data from staged crashesand other non-crash events pius occupant position weight, size, velocity etc. from sled tests
J 2. Analytically derive additional crash and event data from staged crashesand analytically determine occupant motion.
J 3. Design candidate neural network.
Figure 2. Side impact anticipatory sensorsystem.
Figure 2 illustrates a side impact anticipatory sensor system using transducersthat are situated in different locations on one side of the vehicIe, using the same 5. Test the neural network using different data. computer systemas discussedabove. These sensorscan provide the data to permit the identification of an object that is about to impact the vehicle at that side as well as its velocity. An estimatecan then be made of the object’s weight and thereforethe severity of the pending accident. This provides the information for the initial inflation of the side airbag before the accident begins. If additional 7. Output neural network algorithm information is provided from the occupant sensors,the deployment of the side airbag can be tailored to the occupantand the crash in a similar manner as described Figure 3. Smart airbag system development block above. Figure 2 also illustrates additional inputs that, in diagram some applications, provide useful information in determining whether an airbag should be deployed. The neural network algorithm which forms an integral These include inputs from a t?ont crash sensormounted part of the crash sensor described herein can be on the vehicle radiator, an engine speed sensor, and a implementedeither as an algorithm using a conventional wheel speed sensor, as used in the antilock braking microprocessoror through a neural computer which is systemsensor. now available. ln the former case, the training is This anticipatory sensorcan act in concert with or in accomplishedusing a neural pattern recognition program place of the accelerometer-basedneural network crash and the result is a computer algorithm fiquentIy written sensor described above. In the preferred case, both in the C computer language. ln the latter case,the same sensorsare used with the anticipatory sensor forecasting neural computercan be used for the training as used on the crash severity before the collision occurs and the the vehicle. Neural network software for use on a accelerometerbased sensor confirming that forecast. conventional microcomputer is available horn several Collision avoidance systems currently under commercialsources. development use radar or laser radar to locate objects A block diagram of the neural network computer such as other vehicles that are in a potential path of the methodof obtaining a smart airbagalgorithm is illustrated subject vehicle. ln some systems,a symbol is projected in Figure 3. ln the first step,one or more vehicle models onto the windshield in a heads-updispIay signifying that are crashedunder controlled conditions where the vehicle some object is within a possible collision spacewith the and crash dummies are fully instrumented so that the subject vehicle. No attempt at present is made to severity of the crash,and thus the need for an airbag, can
1085 be determined. An occupantsensor is also presentand in and thus the initial position plus the accelerationof the use so that occupantmotion data can be obtained. The vehicle allows a reasonably accurate determination of accelerationduring the crash is measuredat all potential his/her position over time. The problem is more locations for mounting the crash sensor. Normally, any complicated for the belted occupant and the rules position which is rigidIy attachedto the main structural governing occupant motion must be learned from membersof the vehicle is an adequatemounting location modeling and verified by sled and crash tests. for the sensor. Fortunately, belted occupants are unlikely to move The following crash event types, at various velocities, significantly during the critical part of the crash and thus are representativeof those which should be consideredin the initial position at least for the chest is a good establishing crash sensor designs and calibrations for approximation. frontal impacts: The vehicle manufacturerwill be loath to conduct all Frontal Barrier impact of the crasheslisted above at several different velocities Right Angle Barrier Impact for a particular vehicle since crashtests are expensive. If, Left Angle Barrier Impact on the other hand, a particular crash type that occurs in Frontal Offset Barrier Impact the real world is omitted from the library, there is a Frontal Far Offset {Outside of Rails) Barrier Impact chancethat the system will not perform optimally when High Pole on Center Impact the event occurs later resulting in death or injury. One High Poleoff Center Impact way to partially solve this dilemma is to use crash data Low Pole (below bumper)Impact from other vehicles as discussedabove. Another method Frontal Car-to-CarImpact is to create data using the data obtained from the staged Partial Frontal Car-to-CarImpact crash tests and operating on the data using various Angle car-to-carImpact mathematicaltechniques which permits the creation of Front to Rearcar-to-car Impact data which is representativeof crashes not run. One Front to Side Car-to-CarImpact, Both Cars Moving methodof accomplishingthis is to use velocity and crash Bumper UnderrideImpact scaling as describedin detail in the referencedpapers and Animal Impact - SimulatedDeer particularly in reference (1) page 8 and reference (2) UndercarriageImpact (hangup on railroad track type pages 37-49 . This is the second step in the process of object) illustrated in Figure 3. Also included in this step is the Impact Into Highway Energy Absorbing Device analytical determinationof the occupantmotion discussed (Yellow Barrels,etc.) above. Impact Into Guardrail The third step is to assumea candidateneural network Curb Impacts architecture. A choice which is moderately complex is The following non-crashevent types are representative suggested. If the network is too simple, there will be of those consideredin establishing crash sensordesigns casesfor which the systemcannot be trained and, if these and calibrations: are important crashes,the network will have to be revised HammerAbuse (shop abuse) by adding more nodes. If the initial choice is too Rough Road(rough driving conditions) complex, this will usually show up after the training with Normally, a vehicle manufacturer will only be one or more of the weights having a near zero value. In concernedwith a particuIarvehicle model and instruct the any event, the network can be tested later by removing crash sensordesigner to designa sensorfor that particular one node at a time to see if the accuracyof the network vehicle model. This is in general not necessarywhen degrades. Alternately, genetic algorithms are used to using the techniquesdescribed herein and vehicle crash searchfor the optimum network architecture. data from a variety of different vehicle models can and The training data must now be organizedin a fashion should be includedin the training data. similar to the way it will be seen on a vehicle during a Since the system is being designed for a particular crash. Although data from a previously staged crash is vehicle model, static occupantdata needsto be obtained available for the full time period of the crash,the vehicle for that particular model. Although crash data from one mountedsystem will only seethe dataone value at a time. vehicle can be used for the training purposesfor other Thus, the training data must be fed to the neural network vehicles, occupantdata cannotin generalbe interchanged computer,or computerprogram, in that manner. This can from one vehicle model to another. Dynamic position be accomplished by taking each crash data file and data for an occupant will be in general be analytically creating 100 casesfrom it, assumingthat the time period derived based on his/her initial position and rules as to chosenfor a crash is 200 milliseconds and that each data how the body translates and rotates which will be point is the preprocessed acceleration over two determined from sled and crash tests. This is not as milliseconds. This data must also be combined with the complicated as might first appear since an unbelted occupant data derived as discussed above. The first occupantwill usuallyjust translateforward as a free mass training case contains the first crash data point and the
1086 remaining 99 points are zero, or random small values for gas module to inject a particular amount of gas into the the crashdata nodes, and the segmentedoccupant position airbag. For anothercrash such as an 8 mph barrier crash data for the occupantnodes. where airbag deployment is not desired, the desired The secondcrash data casecontains the fust two output value for all of the data lines which are used to data points with the remaining 98 points set to zero or representthis crash (100 lines) would have associated random low values etc. For the tenth data file, datapoint with them a desired output node value of 0 which one will contain the average acceleration at twenty correspondsto a commandto the gas control module not milliseconds into the crash, data point two the average to inject gas into the airbag. The network program then accelerationat eighteenmilliseconds into the crash, and assigns different weights to the nodes until all of the data point ten will contain the data from the first two airbag-deployment-not-desiredcases have an output node miiiiseconds of the crash. This processis continued until value nearly equal to 0 and similarly all of the airbag- the one hundreddata cases are createdfor the crash. Each deployment-desiredcases have an output value close to case is representedas a line of data in the training file. that which is required for the gas control module to inject This sameprocess must be done for each of the crashes the proper amount of gas into the airbag. The program and non-crashevents for which there is data. A typical finds those weights that minimize the error between the training set will finally contain on the order of 50,000 desiredoutput valuesand the calculatedoutput values. crashdata cases and 500,000occupant static datacases. The term weight is a generalterm in the art used to In the pure neural network crash sensorcase, it was describethe mathematicaloperation which is performed possible to substantially trim the data set to exclude all on each datum at each node at one layer before it is those casesfor which there is no defmite requirementto inputted into a node at a higher layer. The data at input deploy the restraint, and the same is true here. For a layer node 1, for example, will be operated on by a particular 30 mph frontal barrier crash, for example, function that contains at least one factor which is analysisof the crash has determinedthat the sensormust determinedby the training process. ln generalthis factor, trigger the deploymentof the airbag by 20 milliseconds. or weight, is different for each combination of an input It is thereforenot necessaryto use data from that crash at node and hidden layer node. Thus, in the exampleabove less than 20 milliseconds since we are indifferent as to where there were 100 input nodes, 12 hidden layer nodes whether the sensorshould trigger or not. Although data and 1 output node,there will in generalbe 1,212 weights greater than 20 milliseconds is of little value from the which are determined by the neural network program point of view of a neuralnetwork crashsensor which only during the training period. An exampleof a function used needsto determinewhether to deploy the airbag since that to operateon the data from onenode before it is input to a would representa late deployment, such is not the case higher level node is the sigmoid function: here since, for some gas control modules, the In the usual back propagationtrained network, let inflation/deflation rate can be controiled after the decision 0, be the output of nodej in layer i, to deploy. Also, the 20 millisecond triggering then the input to nodek in layer i+l is requirementis no longer applicable since it dependson wh~~l&~c~j Y&(i) Qj the initial seating position of the occupant. For cases is the weight applied to the connection where the airbagshould not trigger, on the other hand,the betw:en nodej in layer i andnode k in layer i+ 1. entire data set of 200 data files must be used. Finally, the Then the output of node k in layer i+l is found by training set must be balancedso that there are about as transforming its input, for example, with the many no-trigger casesas trigger casesso that the output sigmoid function: will not be biasedtoward one or the other decision. This OiLI&= l/(l+e’“+‘~“) then is the fourth step in the processas depictedin Figure and this is usedin the input to the next, i+2, layer. 3. If the neural network is sufficiently complex, that is if In the fifth step, the neurat network program is run it has many hidden layer nodes,and if the training set is with the training set. The program uses a variety of small, the network may “memorize” the training set with techniques,such as “‘back propagation”,to assignweights the result that it can fail to respondproperly on a slightly to the connections from the input layer nodes to the different case from those presented. This is one of the hidden layer nodesand horn the hidden layer nodesto the problems associatedwith neural networks which is now output Iayer nodes to try to minimize the error at the being solved by more advanced pattern recognition output nodes betweenthe value calculatedand the value systems including genetic algorithms which permits the desired. For example,for a particular crash such as a 30 determination of the minimum complexity network to mph fiontai barrier impact, an analysis of the crash and solve a particular problem. Memorizing generally occurs the particular occupanthas yielded the fact that the sensor only when the number of vectors in the training set is not Must trigger in 20 milliseconds and the data file sufficiently large or varied comparedto the number of representingthe first 20 milliseconds of the crash would weights. The goal is to have a network which generalizes havea desiredoutput node value which would instruct the from the datapresented and thereforewhich wi1I respond
1087 properly to a new case which is similar to but only optimize the airbag systemfor the occupantand the crash slightly different from one of the casespresented. The severity. network can also effectively memorize the input data if Airbags have traditionally beendesigned based on the many casesare nearly the same. It is sometimesdifficult assumptionthat 30 milliseconds of deployment time is to determine this by looking at the network so it is available beforethe occupant,as representedby a dummy important that the network not be trained on all available corresponding to the average male, has moved five databut that somesignificant representativesample of the inches. An occupant can be seriously injured or even data be held out of the training set to be used to test the killed by the deploymentof the airbag if he or she is too network. It is also important to have a training set that is close to the airbag when it deploys and in fact many very large and varied (one hundredto one thousandtimes people, particularly children and small adults, have now the number of weights or more is desirable). This is the beenso killed. It is known that this is particularly serious function of step five, to test the network using data that it when the occupantis against the airbag when it deploys has not seenbefore, i.e., which did not constitute part of which correspondsto about 12 inches of motion for the the training data. averagemale occupant,and it is also known that he will Step six involves redesigning the network and then be uninjured by the deploying airbagwhen he has moved repeating steps three through five until the results are less than 5 inches when the airbag is completely satisfactory. This step is automatically accomplishedby depIoyed. Thesedimensions are basedon the dummy that some of the neural network software products available represents the average male, the so-called 50% male on the market. dummy, sitting in the mid-seatingposition. The threshold The final step is to output the computer code for the for significant injury is thus somewherein betweenthese algorithm and to program a microprocessor,or a neural two points and thus for the purposesof this table, two computer, with this code. One important feature of this benchmarkshave been selectedas being approximations system is that the neural network system chosen is very to the thresholdof significant injury. These benchmarks simple and yet, becauseof the way that the data is fed to are, basedon the motion of an unrestrainedoccupant, (i) the network, ail relevant calculations are made with a if the occupanthas already moved 5 inches at the time single network. There is no need, for example,to use an that deployment in initiated, and (ii) if the occupanthas additional network to translate a prediction of a vehicle moved 12 inches by the time that the airbag is fully velocity change,and thus the crashseverity, into a setting deployed. Both benchmarks really mean that the for the gascontroller. In fact, to do this would be difficult occupantwill be signiticantly interacting with the airbag since the entire time history would needto be considered. as it is deploying. Other benchmarkscould of coursebe The output from the network is the setting of the gas used; however, it is believed that these two benchmarks controller in the preferredsystem design. are reasonablelacking a significant number of test results to demonstrate otherwise, at least for me 50% male OPERATION OF THE NEURAL NETWORK dummy. CRASH SENSOR - AN EXAMPLE The tables shown in Figures 4 and 5, therefore, provide data as to the displacement of the occupant In Figure 4, the results of a neural network pattern relative to the ah-bagat the time that deployment is recognition algorithm for use as a single point crash initiated and 30 milliseconds later. If the first number is sensor are presented for a matrix of crashes created greater than 5 inches or the secondnumber greater than according to the velocity and crash scaling techniques I2 inches, it is assumedthat there is a risk of significant presentedin the referencedpapers. The table containsthe injury and thus the sensorhas failed to trigger the airbag results for different impact velocities (vertical column) in time. For these cases,the ceil in the table has been and different crash durations (horizontal row). The outlined. As can be.seen in Figure 4, which represents the results presentedfor each combination of impact velocity neuralnetwork crashsensor, none of the cells are outlined and crash duration consist of the displacement of an so the performanceof the sensoris consideredgood. unrestrainedoccupant at the time that air-bagdeployment The table shown in Figure 5 representsa model of a is initiated and 30 milliseconds later. This is presented single point crash sensor used on several production here as an exampleof the resultsobtained from the use of vehicle modelsin usetoday. In fact, it was designedto be a neural network crash sensorthat forms the basis of the optimized for the crashesshown in the table. As shown in smart airbag system. In Figure 4, the successof the Fig. 5, the sensor fails to provide timely airbag sensor in predicting that the velocity change of the deployment in a significant percentageof the crashes accident will exceeda threshold value is demonstrated. representedin the table. Since that sensorwas developed, Here this capability is extended to where the particular several manufacturers have developed crash sensor severity of the accidentis indirectly determinedand then algorithms by trial and error which probably perform used to set the flow of gas into or out of the airbag to better than that of Figure 5. It is not possible to ascertain the success of these improved sensors since the
1088 algorithms are consideredproprietary. Some algorithms can now be analyzedusing the abovemethods. have recently been published in the patent literature and
SCALED BARRIER SCALING FACTOR
VELOCITY
1 1.2 14 16 18 2
8 MPH NT NT NT NT NT NT
10 MPH NT 0.7R 9 0.913 1 1 o/3 0 NT NT
12 MPH 0 011 1 08/35 09/35 1 o/3 4 1 4/3.9 2 014 7
14 MPH 0 Oil 2 0 914.1 1Ot38 1 w.0 I.314 0 1.714.5
16 MPH 00/l 4 09144 1 o/4.0 1 l/4.0 1.414.3 1714.6
18MPH 0011 6 0 814 2 0 713 6 1 214.5 1614 8 1 8f4.9
20 MPH DO/l 8 0 714 3 0 714 0 11143 1.314.4 1 O/3.8
22 MPH 00/l 9 05i39 0 714 0 0 Q/4 1 12146 11142
24 MPH OORl OlR3 0 814 4 08/42 1 3150 ? 414 8
26 MPH OOR3 OlR5 0 5/4 0 0 Q/4 5 1 o/4 4 12/46
28 MPH OOR5 OORI 01124 0 714 2 0 8/4 1 0 5132
30 MPH 0 o/2 7 OOR3 0 l/2 6 0 l/2 3 0 814 4 1215 0
32 MPH OOR8 OOR4 0 l/2 8 0 1125 0 914 7 1 l/49
34 MPH 0 o/3 0 OOR3 OORO 00118 0 614 2 12/53
Figure 4. Neural network single point sensor performance.
SCALED BARRIER SCALING FACTOR
VELOCITY
8 MPH
10 MPH
12 MPH
14 MPH
16 MPH 22/76 27i79 3 4/a 5 42193 NT NT
18 MPH 2 2/8.0 2 8/a 7 3 wwi 42197 5.0110.5 178R75
M MPH 2on.9 31193 3.719 7 43112 5 0110.9 5 g/11 7
22 MPH 1 o/5 3 2 718 9 391104---. 43109 52/115 5 9/12 2
24 MPH 514 2 16165 39/108 481116 541120 6 1112 8
26 MPH 414 1x7 20/6a I-~ 4 WI1 5 5 6113 64/135
26 MPH 414 1 6l40 16/66 zg---zy-~~~;~ j
30 MPH 414 2 340 8142
32 MPH 3/4 2 514 1 717 2 2.117 0 2 6/74 3 4/a 4
34 MPH 3/4 0 514 2 714 3 914 5 26175 4 o/9 6
Figure 5. Optimized standard single point sensorperformance.
1089 GAS FLOW CONTROLLER output value of the output node of the neural network. Naturally, the abovediscussion is for illustration purposes One issue that remains to be discussed is the only and there are many ways that the interface between derivation of the relationship betweenthe gas controller the neural network system and the gas controller can be setting and the desired volume or quantity of gas in the designed. airbag. Generally, for a low velocity, long duration The gas flow controller can also make use of threshold crash, for a small light weight out-of-position additional inputs including in particular the pressure occupant, the airbag should be inflated slowly with a within the airbag. All such information inputs can be reiatively small amount of gas and the out flow of gas handled within the neural network or, in the case of the from the airbag controlled so a minimum value, constant airbagpressure input, within the control mechanismitself. pressureis maintainedunit1 the occupantjust contactsthe in this casethe output from the neural network would be vehicle interior at the end of the crash. Similarly, for a the desiredairbag pressure. high velocity crash with large heavy occupant,positioned The descriptions above have concentrated on the far from .the airbag before deployment is initiated, but control of the gas flows into and out of an airbag, with a significant forward relative velocity due to pre- Naturally, other parts of the occupantrestraint systemcan crashbraking, the airbag should be deployedrapidly with also be controlled in a similar manneras the gas flows. In a high internal pressureand an out flow control which particular, various systemsare now in use and others are maintains a high pressurein the airbag as the occupant being developedfor controlling the force applied to the exhauststhe airbag to the point where he almost contacts occupantby the seatbelt. Such systemsuse retractors or the interior vehicle surfacesat the end of the crash. These pretensioners,others use methods of limiting maximum situations are quite diff‘erent and require significantly the force exert by the seatbelt, while still others apply different flow rates into and out of the airbag. As crash damping or energy absorbing devices to provide a variability is introducedsuch as where a vehicle impactsa velocity sensitive force to the occupant. Also, the crash pole in front of a barrier, the gas flow decisions will be accelerometerand occupantsensors have been the main changedduring the crash. inputs to the neural network system as describedabove. In theory the neural network crash sensor has the Although not describedin detail, the neural network can entire history of the crash at each point in time and make optimum use of other sourceson information such therefore knows what instructions it gave to the gas as seatbeltuse, seat position, seat back position, vehicle controller during previous portions of the crash. It velocity etc. as additional inputs into the neural network therefore knows what new instructions to give the system for particular applications depending on the controller to accountfor new information. The problem is availability of such information. to determine the controller function when the occupant parametersand the crash forecastedseverity are known. CONCLUSION This requires the use of an occupant crash simulation program such as MadymoTMfrom TN0 in Del& The The systemdescribed herein usesa neural network, or Netherlands, along with a model of the gas control neural-network-derivedalgorithm, to analyzethe digitized module. A series of simulations are run with various accelerometerdata created during a crash and, in some settings of the controllable parameterssuch as the gas cases, occupant size, position, seatbeit use, weight and generationrate, gas inflow and gas outllow restriction velocity data, and, in other cases, data from an until acceptableresults are obtainedand the results stored anticipatory crash sensor, to determine not only if and for that particular crash and occupantsituation. In each when a passive restraint such as an airbag should be case, the goal may be to maintain a constant pressure deployedbut also to control the flow of gas into or out of within the airbag during the crash once the initial the airbag. deploymenthas occurred. Thoseresults for eachpoint in Generally,the presentdevice provides a smart airbag time are converted to a number and that number is the system that optimizes the deployment of an occupant desired output of the neural network used during the protection apparatusin a motor vehicle, such as an airbag, training. A more automatedapproach is to couple the to protect an occupant of the vehicle in a crash. The simulation model with the neural network training system includes an accelerometermounted to the vehicle program so that the desired results for the training are for sensingaccelerations of the vehicle and producing an generatedautomatically. Thus, as a particular case is analog signal representative thereoc an electronic being prepared as a training vector, the MadymoTM converter for receiving the analog signal from the sensor program is run which automatically determines the and for converting the analog signal into a digital signal, settings for the particular gas control moduIe, through a and a processorwhich receives the digital signal. The trial and error process,and thesesettings are convertedto processor includes a neural network and produces a a number and normalizedwhich then becomethe desired deployment signal when the pattern recognition system
1090 determines that the digital signal contains a pattern characteristic of a vehicle crash requiring occupant protection and further producesa signal which controls the flow of inflator gas into or out of the airbag. In some cases,the systemalso includes an occupantposition and velocity sensorwhich outputsa signal that is also usedby the processorin producing the signal which controls the flow of gas into or out of the airbag.
REFERENCES
1. Breed, D.S. and Castelli, V. “ProbIems in Design and Engineering of Air Bag Systems”, Society of Automotive EngineersPaper SAE 880724, I988 2. Breed, D.S., Castelli, V. ‘Trends in Sensing Frontal Impact”, Society of Automotive EngineersPaper SAE 890750, 1989. 3. Breed, D.S., Sanders, W.T. and Castelli, V. “A Critique of Single Point Crash Sensing”, Society of Automotive EngineersPaper SAE 920124, 1992. 4. Breed, D.S., Sanders, W.T. and Castelli, V. “A CompleteFrontal Crash SensorSystem - I”, Society of Automotive EngineersPaper SAE 930650,1993. 5. Breed, D.S. and Sanders, W.T. “Using Vehicle Deformation to Sense Crashes”, Presented at the International Body and Engineering Conference, Detroit MI, 1993. 6. Breed, D.S., Sanders, W.T. and Castelli, V., “A Complete Frontal Crash Sensor System - II”, ProceedingsEnhanced Safety of Vehicles Conference, Munich, 1994, Published by the US Department of Transportation, National Highway Traffic Safety Administration, Washington,D.C. 7) Techniques And Application Of Neural Networks, edited by Taylor, M. and Lisboa, P., Ellis Horwood, West Sussex,England, 1993. 8) Naturallv Intelligent Systems, by Caudill, M. and Butler, C., MIT Press, Cambridge Massachusetts, 1990. 9) Digital Neural Networks, by Kung, S. Y., PTR PrenticeHall, EnglewoodCliffs, New Jersey,1993. lO.Breed, et al., U.S. Patent5,563,462 Vehicle Occupant Position And Velocity Sensor. 1l.Breed, et al., U.S. Patent (allowed) Vehicie Interior Identification and Monitoring System.
1091