EP2054274A1

EP2054274A1 - Method and device for the actuation of personal protection means

Info

Publication number: EP2054274A1
Application number: EP07787284A
Authority: EP
Inventors: Josef Kolatschek; Joerg Breuninger
Original assignee: Robert Bosch GmbH
Current assignee: Robert Bosch GmbH
Priority date: 2006-08-16
Filing date: 2007-07-10
Publication date: 2009-05-06
Also published as: CN101506001B; US20090306858A1; US8374752B2; DE102006038151B4; CN101506001A; JP2010500227A; WO2008019915A1; DE102006038151A1

Abstract

A device and a method for the actuation of personal protection means is provided, wherein a feature vector, having at least two features of at least one signal of an accident sensor system is formed by an evaluation circuit. The evaluation circuit classifies the feature vector into the respective dimension by means of at least one classification limit. The actuation circuit generates an actuation signal, wherein an actuation circuit actuates the personal protection means as a function of the actuation signal.

Description

description

title

Method and device for controlling personal protective equipment

State of the art

The invention relates to a method and a device for controlling personal protection means according to the preamble of the independent claims.

From DE 103 60 893 Al it is already known to control personal protection means in response to a comparison of a forward displacement with a threshold value. The threshold is set in response to a deceleration and deceleration. The decay and deceleration span a two-dimensional feature space which is divided by the threshold into two areas. These two areas characterize the classes that are significant for the activation of the personal protection devices, the threshold value representing the class boundary.

Disclosure of the invention

The inventive method for controlling personal protection devices or the inventive device for controlling personal protection means have the advantage that the application of the class boundaries is not limited to one, two or the three-dimensional feature space. In particular, dependencies of the features in rooms larger than the third dimension can be used. The class membership is determined by a linear combination of a nonlinear function with feature values. Thus, the method on a control unit is well calculated or reproducible. With the device according to the invention or the method according to the invention complicated classification tasks are very easy to solve. The invented The method according to the invention preferably uses a so-called support vector machine (SVM). This is well founded by the statistical learning theory. The determination of the class boundaries is given by an analytically solvable optimization problem, so that this process can be performed automatically without additional expert knowledge from a computing machine, in particular an evaluation circuit, which can be designed as a microcontroller. There are only a few parameters that have to be set by the user, ie an application engineer. Therefore, it is easily adaptable to various classification problems associated with accident detection. The classification quality of the method according to the invention is very high. Due to the setting in class boundary finding, the method according to the invention allows additional free space to be used. The inventive method has a high generalization ability. Ie. there is no danger that the decision-making process will be optimized too much for the data record (training data record) used during the application and therefore offers a poor classification performance for previously unknown data not contained in the training data set.

The feature vector is at least two-dimensional according to the independent patent claims. For classification, the feature vector is compared with the class boundary. If this is within a corresponding class, then the feature vector is assigned to this class.

The accident sensor system can contain several accident sensors, including different types. The interface to the accident sensors can be implemented in terms of hardware or software. In particular, a software executed interface on the evaluation, in particular a microcontroller, be present. Instead of a microcontroller, other processors or ASICs can serve as an evaluation circuit. The drive circuit can also be present as an integrated circuit in a control device for controlling personal protection devices. Overall, the device can be arranged as a control device for controlling personal protection means or integrated in a control device for controlling safety means. The last control unit can namely also drive a vehicle dynamics control. The measures and refinements recited in the dependent claims, advantageous improvements of the specified in the independent claims device for controlling personal protective equipment or specified in the independent claims method for controlling personal protection means are possible.

The class boundary or limits can already be determined in advance. For this purpose, preferably a data-oriented modeling method of a support vector machine (SVM) can be used. This method is, for example, from Bernhard Schölkopf and Alex Smole: Learning with Kernels, MIT

Press, Cambridge, MA, 2002. It will be briefly explained below.

It is particularly advantageous that the class boundary is loaded from a memory. Alternatively, it is possible that the class boundary is determined by means of at least one training vector and by means of a kernel function. This special

Training vector is in the support vector machine a so-called support vector, which is shown below, to specific solutions in a constraint for determining the minimum of a function. For the determination of the solution, it is necessary to have the features such that they can be separated with a simple class boundary in the form of a straight line, or with higher-dimensional input data of a hyperplane (ie linear in both cases). As will be explained in detail later, the kernel function implicitly allows the features to be brought into such a linearly separable representation without having to perform this step explicitly and thus with a high computational effort. This method of the support vector machine provides an efficient and well reproducible way to design a driving algorithm. In particular, this allows complicated classification tasks to be solved. Above all, the support vector machine method enables expert knowledge, which is necessary in the prior art solutions, to be minimized or even eliminated altogether. This also makes an algorithm easier to understand and interpret. The resource requirement for displaying the class boundary (characteristic curve) is reduced and the application effort is reduced. - A -

It is also advantageous that the classification is carried out in binary form. This is easy to implement and, by using a tree structure, allows for a grading with a successive refinement. In each case, binary classifiers are provided at the branches. Thus, it is then possible to realize complicated classification problems by modular assembly of such binary classifiers. By omitting unneeded binary classifiers within the tree, the inventive method or device according to the invention can be simplified to a meaningful level. In extreme cases, it is then reduced to a simple binary classifier.

It is advantageously possible to use two or more similar or different classification trees in parallel. Thus, one tree can for example determine the crash severity and the other tree regardless of the crash type. This can then be reconnected later to find the right drive.

In addition, it is advantageous to assign a specific retention means to each tree. For example, it is possible to assign the control of a belt tensioner to a first tree and the activation of a second tree

Airbags. If both trees use simple single binary classifiers, which only know the classes Fire and NoFire respectively, a different control of the restraining means can nevertheless be realized in this way. The activation of the retaining means thus takes place on the basis of the accident situation determined by the classifier. For example, with a simple classifier Fire / NoFire, the drive signal, upon detection of the Fire class, causes the immediate activation of an associated personal protection device.

In a more sophisticated embodiment, the drive signal uses z. B. the information about the crash type and the crash severity together. For this purpose, in a special table for each combination of these two size a specific set of personal protection deposited, which is then activated. In addition to the method exemplified above to solve one of the listed classification problems with more than two classes, other methods are also applicable. The above reference Schölkopf et al. shows still other methods of how such a multi-class problem can be solved in a favorable manner. These methods are mentioned in the literature under

Names "one versus the rest", "pairwise classification", "error correcting output coding" and "multiclass objective functions" known.

For special problems that occur in the context of accident detection and classification, it may happen that only training data of a particular class are available. The task then is to determine whether or not a feature vector measured by measurement belongs to an event of that class. This corresponds to a so-called one-class classifier. For this purpose, too, the method according to the invention and in particular the support vector machine algorithm can be used in an advantageous manner. Therein, the quality or sensitivity can then typically be set via a parameter V. For example, in the classification of misuse events, such a situation exists. Since misuse events, ie non-triggering events, are much easier and cheaper to carry out than crash tests in which the vehicle is destroyed, a large number of such test data usually suffice. With the help of a one-class classifier, a misuse recognition system can be set up. Ie. all feature vectors that are outside this class are trigger cases.

Advantageously, a regression can be used to generate a continuous value for the classification. This is particularly advantageous in the case of a classification which describes a steady increase of a certain property, for example when the class describes crash severity or crash speeds. The advantage here is that in such a case the output of the system consists not only of the discrete numerical values but of a real numerical value of a continuous range of values. The corresponding methods are also described in Schölkopf et al. as described above. The numerical value (s) can then be Therefore, tables are assigned to a specific activation pattern of personal protective equipment.

Advantageously, the at least two features are determined from a time block of the at least one signal. These are about a specific

Time, namely the block length, the signals of individual sensors, which are already pre-processed, such as a filtering or integration, considered in context. The feature vector consists of the signal characterizing variables within this block. This may be, for example, the mean of the sensor data, the variance or higher moments, the first integral, the second integral, the coefficients of a wavelet decomposition, a Fourier decomposition, the index values of a codebook, if a vector quantization is applied to the input data within the block applies. Likewise, the coefficients of a polynomial regression can be determined. The one or more selected feature quantities can be determined in one step at the time of the end of the block or can also be determined continuously or recursively with the arrival of the data, ie the signal. In extreme cases, it is also possible to match the block length T to the sampling time of the sensor so that in each data block only exactly one data value is contained, which is then correspondingly converted into a feature vector. The feature vector may also contain appropriately processed data from various sensors with optionally different sensing principles. If different blocks are used, these blocks may also overlap or be separated in time.

Advantageously, the use of a feature vector allows features of the feature vector to be formed from signals from different sensors. This allows a comprehensive description of the accident event.

Advantageously, there is a computer program which runs on the control unit, in particular of the evaluation circuit, such as a microcontroller, for example. This computer program may be written in an object-oriented language or other common computer languages. In particular, this computer program can be used as a computer program product on a data carrier. ger, which is machine-readable, such as a hard disk, an electronic memory such as an EEPROM or on a magneto-optical disk or an optical disk such as a DVD or CD.

Embodiments of the invention are illustrated in the drawings and are explained in more detail in the following description.

Show it:

FIG. 1 shows a block diagram of the device according to the invention,

FIG. 2 shows a software structure on the microcontroller,

FIG. 3 shows a flow chart of the method according to the invention,

FIG. 4 shows a data flow diagram of the method according to the invention,

FIG. 5 shows a further data flow diagram of the method according to the invention,

FIG. 6 shows a block structure and

Figure 7 is a tree structure.

The basic idea of the Support Vector Machine Algorithm is explained below, which can preferably be used for the method according to the invention.

The SVM algorithm is, in the simplest case, able to perform a binary classification, ie to assign an unknown data vector to one of the two classes on the basis of a training data set of data vectors. Since the two classes are not necessarily separable by a simple class boundary in the form of a straight line, or in the case of higher-dimensional input data, a hyperplane (ie linear in both cases), they would have to be mapped into a higher-dimensional feature space by a transformation. where possible. In the original space, in turn, this separating hyperplane would correspond to a nonlinear separation surface.

With the aid of a kernel function, the SVM algorithm now allows to calculate this separation surface without first performing the mapping into the feature space. This significantly reduces the demands on the computing power, makes the implementation of the classification technically possible in many cases (above all those with a high-dimensional feature space). Since the non-linear separation surface in the input data space can in each case be traced back to a linear one in the feature space, the generalizability of the classification method can be set well.

As detailed descriptions of the method can be found in the literature (eg in Bernhard Schölkopf and Alex Smola, Learning with Kernels, MIT Press, Cambridge, MA, 2002), only a brief overview is given here.

Simple case: There are two classes which are linearly separable, i. by a class boundary that exists in the form of a straight line (in the two-dimensional feature space) or a plane or hyperplane in the higher-dimensional feature space. Each of the 1 feature vectors Xj of a training data set can be combined with the class information yj to form a data pair z \. The two classes are assigned the values +1 and -1, where z. For example, + l can mean Fire crash and -1 can be NoFire crash.

So:

Data pairs z \ = {x _t , y _t ), with 1 <i <I

in which

y _j e {-l, + l} specifies the class information and the X _{j represent} the multi-dimensional feature vectors. Then a linear multidimensional separation plane (hyperplane) between both classes has the form:

f (x) = W ^τ x + b It then applies:

w ^τ xj + b> 1, if y, = 1 (class 1) w ^τ xj + b> - 1, if yj = -1 (class 2)

Summarized:

yj ^(τ w x j + b)> - 1

Now we are looking for exactly that hyperplane for which the margin m between the classes 1 and 2 becomes maximal ("maximal margin"). It is precisely this hyperplane that is optimal for separating both classes and has the best generalization properties.

The normalized edge can be expressed as

1 m = ^■ w

The maximum edge is now the one for which the function

E (w) = I w

becomes minimal. Thus, the determination of the class boundary can be attributed to the following task:

It is E (w) = _ _w | 2 to minimize under the constraint of the correct

2 '

Classification of training data, d. H. :

yj (w ^τ xj + b)> 1, with l ≤ i ≤ L This is a standard problem (minimizing a quadratic function with linear inequalities as a constraint) and can be solved with appropriate methods.

It can be formulated as a Lagrangian function:

L (w, b, a) = - w | 2 ^" ΣaXyX w ^τ xj + b) -l), where the aj are the Lagrange multipliers.

According to the Karush-Kuhn-Tucker (KTT) theorem, the following additional conditions lead to the inequality constraints:

For the optimal values of w, b and aj, denoted by w *, b * and a *, the following applies:

a * (yj (w ^τ xj + b *) - l) = 0 with l ≤ i ≤ l

d. H. the NB (constraint) is either met or the associated lagrangian multiplier is zero (KTT complementarity condition). This is an extremely advantageous feature. It clearly shows that only a fraction of the feature vectors make any appreciable contribution to the determination of the class boundaries and must be taken into account, namely exactly those for which the Lagrange multipliers are not equal to zero. These vectors are called support vectors. In them, therefore, the information about the optimal hyperplane is available in compressed form.

By reshaping, finding the solution for the optimal hyperplane can be formulated as a "dual optimization problem", which depends only on the scalar products of the feature vectors Xj:

1 ' ¹ I

^{L (a) =} -τ Σ ^a . ^a jy.yj ^x . ^Tχ j ⁺ Σ ^fl . ^{= "} - a ^T Ka ⁺ l ^T a

¹ 'J = I 1 = 1

This can be solved with the standard methods of quadratic programming. For a practical application, the above approach needs to be extended. With training feature vectors generated from real crash records, it is always possible to include therein points that are not separable or fall within the range of the edge (eg, due to measurement errors). In order to get a solution anyway, a slip variable ξ C, slack variable "must be introduced into the NB, which introduces a tolerance to such erroneous training data.

yj (w ^τ xj + b)> 1 - j ξ> 0th

The function to be minimized can then be defined as

E (w) = - w | + C ^ ξ _; where C can be understood as a noise parameter i = l, which limits the influence of a training data value on the class boundary.

The Lagrange function of the dual optimization problem then arises as described above.

Typical case:

The case of a linear separability is rather rare in classification problems for the activation of restraints. Much more frequently, the feature vectors form an arrangement which can only be separated into the associated classes by a non-linear class boundary. This problem could be circumvented by introducing a mapping Φ (x): R "-> E into the one higher-dimensional feature space E, in which the classes are again linearly separable, and the corresponding Lagrange function would be:

L (a) = - \ Y _j a _ι a _] y _ι y _] {Φ {x _ι ), Φ {x _] )) + Y _j a _ι

^ ι, j = l ι = l

The previous scalar product (XJ, x;) is replaced by the scalar product (Φ (x _ι ), Φ (x _J )) in the feature space. Here comes a decisive advantage of an SVM to bear. In SVM, the mapping Φ (x): R "-> E is not performed at all, using instead a certain property of scalar products resulting from the

Set of mercer results. Thus, under certain conditions, the scalar product (Φ ^^ Φ ^)) can be replaced by a so-called kernel function k ( _xj , xp), which yields the same result, thus leaving out the projection into the higher-dimensional space Instead, the solution is calculated directly in the lower dimensional input space, which is called a "core trick." The core function (kernel), for example, is:

Gaussian kernel: k (x _j , xp = exp

Polynomial kernel: k (xj, xp = ((xf xp + ö) ^d

Sigmoider core; k (xj, xp = + Θ)

1

Inverse multi-quadratic kernel: k (x _j , xp = x, -x. + C

In addition, in the literature mentioned a variety of other possible nuclei are mentioned, all of which may be appropriate. Just as in the simple linear case, it is advantageous for the case of non-separable classes to introduce the slip variable ξ C, slack variable "into the NB The equations listed there remain the same mutatis mutandis It can be found by applying a standard method for solving a quadratic optimization problem under linear constraints, eg, LOQO or equivalent methods).

All in all, it can be said that the SVM, like many other methods, does not

Estimation of a density function underlying the input data, but it minimizes the so-called worst-case risk of classification. It belongs to the class of "distribution-free procedures".

The following describes how the method according to the invention or the device according to the invention can be embodied by way of example. In principle, the method according to the invention or the device according to the invention is trained in an off-line phase, ie before use in the vehicle, ie. H. the core, class boundaries, or support vectors are determined based on training data. This information is then stored in the control unit in a suitable form, that is stored in a memory and then forms the classifier for the online operation of the device according to the invention or of the method according to the invention. If the support vectors and the core function are stored in the device according to the invention, the device according to the invention can determine the class limits online via the listed equations or calculate the class affiliation directly. Of course it is also possible to store the class boundaries directly.

Basically, the data processing takes place as follows: First, a recording of the signals of the accident sensors is made and there is a feature extraction. Then the classification is carried out with the inventive

Method and ultimately the control of the personal protection means.

FIG. 1 illustrates in a block diagram the device according to the invention. The device according to the invention is embodied here by way of example as a control device for the activation of personal protection devices. In this case, the control unit SG a

Control unit, which is configured only for controlling personal protection means PS, but it is alternatively possible that it is a control device for controlling safety means in general and can also make interventions in a vehicle dynamics control or a braking system. The control unit SG has as a central element a microcontroller μC.

This microcontroller μC is an evaluation circuit according to the independent device claim. Alternatively, it is possible to use other processor types or an ASIC. It may even be possible to use a discretely constructed circuit. The microcontroller .mu.C is connected to a memory S via a data input / output. This memory can be a permanently descriptive be a volatile memory, as it is a so-called RAM in the usual way. However, combinations of memories which can also record data permanently under the name S are also possible. In particular, such memories from which according to the invention the class boundary or the core function and the support vectors can be loaded in order to determine this class boundary. Furthermore, two interfaces IF1 and IF2 are connected to the microcontroller .mu.C, which are embodied here as discrete components. Ie. they are integrated circuits and convert signals from sensors located outside of the SG control unit into a data format that the microcontroller μC can process efficiently. Furthermore, an acceleration sensor system BS1, which is located within the control unit SG, is connected to the microcontroller .mu.C via a data input. This acceleration sensor can detect accelerations at least in the vehicle longitudinal direction. Usually, however, it is possible that this acceleration sensor system BS1 also detects accelerations transversely or obliquely to the vehicle longitudinal direction. An acceleration sensor in vehicle vertical direction is possible.

The microcontroller .mu.C now has a software interface via which the sensor system BS1 is connected to the microcontroller .mu.C. The sensor system BS1 can transmit its data analog or digital to the microcontroller .mu.C. The acceleration sensor is usually constructed, d. H. a micromechanical element provides for the detection of accelerations. Alternatively, it is possible that further sensor types, such as a structure-borne noise sensor or a rotation rate sensor are arranged in the control unit SG.

To the interface IFl a pressure sensor P and an outsourced acceleration sensor BS2 is connected. The pressure sensor P is preferably arranged in the side parts of the vehicle to detect a side impact. In this case, the pressure sensor P detects an air pressure, the by a

Side impact in the side part is adiabatically compressed. This allows a very fast detection of such a side impact. The acceleration sensor BS2 may be installed in the vehicle front, for example, to detect a pedestrian impact or a frontal impact. In this case, the acceleration sensor BS2 is installed behind the bumper, for example or on the radiator grille. It is additionally or instead possible that the acceleration sensor system BS2 is installed in the sides of the vehicle. Thus, the acceleration sensor BS2 then serves for the detection or plausibility of a side impact. In this case, the acceleration sensor BS2 can also be sensitive in various directions in order to serve for the plausibility check or detection of specific impact types.

The data transmission to the interface IF1 of the pressure sensor P and the acceleration sensor BS2 is usually done digitally. It is possible to use a sensor bus, but in the present case point-to-point

Connections provided with a power data transmission.

Connected to the interface IF2 is an environment sensor system U which records data from the surroundings of the vehicle. In this case, other collision objects are detected, detected and characterized, for example via a trajectory or the

Collision speed. Radar, ultrasound, infrared, lidar or video sensors are possible as environment sensors. Other external sensors, such as indoor sensors, are possible.

The microcontroller .mu.C now controls a drive circuit FLIC in response to these sensor signals and its activation algorithm, which serves to activate the personal protection means PS. The drive circuit FLIC has output stages, which are turned on when a drive signal from the microcontroller μC comes. Also, a logic that the signal of the microcontroller μC with a signal plausibility or a parallel evaluation, which is not shown here for simplicity, may be present. The personal protective equipment is, for example, airbags, belt tensioners, roll bars, external airbags, a liftable front hood and other possible personal protective equipment for occupant or pedestrian protection. These can be controlled pyrotechnically or reversibly, for example by an electric motor. Other components which are generally necessary to the operation of the controller SG but which do not contribute to the understanding of the invention have been omitted here for the sake of simplicity. The microcontroller .mu.C forms a feature vector from the signals of the sensors BS1, P, BS2 and U and determines, based on, for example, stored linear class boundaries, which class this feature vector belongs to. As stated above, it is alternatively possible that the class boundaries in operation can be determined based on support vectors and a kernel function.

On the basis of the classification, the microcontroller .mu.C decides whether a drive signal is generated and what content it has. This drive signal is then transmitted to the drive circuit FLIC. The transmission within the control unit SG is usually carried out via the so-called SPI bus.

Figure 2 illustrates important software modules used by the microcontroller μC. First, the above-mentioned interface IF3 is shown, which is used for the application of the acceleration sensor BS1. The interface IF3 has the same function as the hardware interfaces IF1 and IF2 to provide the sensor signals. With the software module 20, a feature vector is then formed from the signals of the sensors. This is classified in the manner according to the invention with the software module 21 and based on the classification is then optionally generated with the software module 22, a control signal indicating which personal protection means are to be controlled. Other software modules are possible, but not shown here for the sake of simplicity.

FIG. 3 explains in a flow chart the sequence of the method according to the invention. In method step 300, a feature vector is formed from the signals of the sensors. This is classified in method step 301, based on the class boundaries. These are either loaded or determined by the support vectors and the kernel function. On the basis of the classification, a control signal is generated in method step 302, which indicates which personal protection devices are to be controlled.

FIG. 4 illustrates in a data flow diagram the function of the device according to the invention or the sequence of the method according to the invention which takes place on the device according to the invention. Block 40 identifies the individual processing steps. In the processing step 402, the sensors 41, 42 and 43 generate their signals, which are then available as measurement data 48. In method step 403, a feature extract from these signals is generated. tion, since it may be, for example, the signals themselves or filtered, integrated, derived, averaged, etc. processed signals. As a result, as indicated by the block 44, the feature vector 49 is present. In step 404, the classification is performed by block 46. This block 46 classifies the feature vector in the manner described above so that the class information 400 is present at the output, which then enters block 47, which is considered to be processing step 405 and the drive signal 401 generated.

FIG. 5 illustrates the formation of a feature vector 508. The sensors are located

500, 501 and 502 ago. Their signals are subjected to pre-processing in blocks 503, 504 and 505, for example filtering or integration or some other mathematical operation. In blocks 506, 507 and 508, the temporal block formation and the feature extraction then take place. In the feature vector 508, the features which have been extracted are then arranged in a vector, wherein a multiplicity of features can be present per sensor, here four for the sensor 1 and five for the sensor N, for example. The number of sensors is not fixed, but there must be at least one sensor. The installation location of the sensors can also be selected in a variety of ways. The method according to the invention can operate, for example, with centrally installed sensors, but also with peripheral sensors in the vehicle side, in the rear of the vehicle or in the front of the vehicle. Also, a combination, as shown above, this installation locations is possible. Examples of suitable sensors are acceleration sensors, pressure sensors, body sound sensors, temperature sensors or sensors with other physical properties

Measuring principles.

The conversion of the measured individual data in a feature vector is preferably done in time blocks. Feature vectors of more than one time block can be calculated. For this purpose, as shown in Figure 6, over a certain time, here the block length T, which is denoted by Bl, B2 and B3, the data of individual sensors, which may already be subjected to a certain preprocessing considered in a context. As indicated above, the feature vector consists of the signal within these blocks of characterizing quantities. This can be, for example, the mean value of the sensor the variance or higher moments, the first integral, the second integral, the coefficients of a wavelet decomposition, a Fourier decomposition, the index values of a codebook, if one applies a vector quantization to the input data within the block. Likewise, the coefficients of a polynomial regression can be determined. The one or more chosen

Feature sizes can be determined in one step at the time of the end of the block or can also be determined continuously or recursively with the arrival of the data. In extreme cases, it is also possible to match the block length T of the scanning side of the sensor so that only one data value is contained in each data block, which is then correspondingly converted into a feature vector. The feature vector may also contain correspondingly processed data from various sensors with optionally different sensing principles.

The classification then has the task of classifying the event which has generated the feature vector into specific classes. Such a classification can z. B. from the two classes Fire and NoFire consist (Example 1). In this case, it is a binary classifier. But it is also the accident event more precisely characterizing class divisions conceivable:

Example 2:

Ci = no triggering event C2 = soft barrier crash C3 = hard barrier crash

Example 3:

Ci = no triggering event C2 = symmetric crash event

C3 = left crash C4 = right crash

Example 4: C ^ = crash severity 1 C2 = crash severity 2 C3 = crash severity 3 C4 = crash severity 4 C5 = crash severity 5

CQ = crash severity 6 C 7 = crash severity 7

Example 5:

Ci = crash speed between Okm / h and 10km / h

C2 = crash speed between 10km / h and 20km / h

C3 = crash speed between 20km / h and 30km / h

C4 = crash speed between 30km / h and 40km / h C5 = crash speed between 40km / h and 50km / h

CQ = crash speed between 50km / h and 60km / h

With binary classification, as shown in FIG. 7, such a refined classification can be successively performed. In the classification level 70 it is determined with the classifier 74 whether the crash severity is less than the value 4. If this is the case, then it goes to the classifier 75, which determines whether the crash severity <2, where we are in the classifier level 71. If so, then the crash severity is determined to be 1 with the classification result as indicated in block 700. If this is not the case, then another classifier level 72 is inserted, so that the classifier 79 determines whether the

Crash severity <3 is. If this is the case, the classification result 701 is determined with the crash severity = 2, this is not the case, then the classification result 702 is determined with the crash severity = 3. If, however, it has been determined in the classifier level that the crash severity is not <4, then the classifier level 71 is jumped to classifier 76, which then checks whether the crash severity is <6. If this is the case, then it jumps to the classifier level 72 and thereby to the classifier 78. This checks whether the crash severity is <5. If this is the case, then the classification result 73 is jumped to block 703, which determines that the crash severity = 4. Was determined by the classifier 78 that the crash severity is not <5, then jump to the classification result 704 and determine that the crash severity = 5. If it was determined by the classifier 76 in the classification level 71 that the crash severity is not <6, then the classifier 77 is jumped to the classifier level 72, where it is checked whether the crash severity is <7. If so, then the classification result 705 is determined that the crash severity = 6. If this is not the case, then it is determined that the classification result is 706, namely the crash severity = 7. These values are exemplified here. You can use completely different values. As FIG. 7 also shows, it is not necessary to have a classifier for each branch in every classifier level, it is also possible to jump directly to a classification result.

Claims

claims

1. Method for controlling personal protective equipment (PS) with the following method steps:

Formation of a feature vector with at least two features from at least one signal of an accident sensor system (BS1, BS2, P, O)

Classification of the feature vector in the corresponding dimension by means of at least one class boundary

Control of personal protection equipment (PS) depending on the classification.

2. The method according to claim 1, characterized in that the at least one class boundary is loaded from a memory (S).

3. The method according to claim 1, characterized in that the at least one class boundary is determined by means of at least one training vector and by means of a kernel function.

4. The method according to any one of the preceding claims, characterized in that the classification is carried out in binary.

5. The method according to claim 4, characterized in that for the binary classification at least a first tree is used, at the branches of each a respective binary classification is performed.

6. The method according to claim 5, characterized in that a respective tree is assigned a personal protection means (PS).

7. The method according to claim 5, characterized in that a second

Tree is assigned a crash severity and a crash type is assigned to a third tree.

8. The method according to any one of claims 1 to 3, characterized in that a regression for generating a continuous value is used for the classification.

9. The method according to any one of the preceding claims, characterized in that the at least two features are obtained from a time block.

10. The method according to any one of the preceding claims, characterized in that the at least two features are formed from signals from different accident sensors.

11. Device for controlling personal protective equipment (PS) with:

at least one interface (IF1, IF2, IF3), which provides at least one signal of an accident sensor system (BS1, BS2, P, O) to an evaluation circuit (.mu.C) which forms a feature vector with at least two features from the at least one signal and the characteristic value Vector classified in its corresponding dimension by means of at least one class boundary, wherein the evaluation circuit (.mu.C) depending on the classification, a drive signal generates a drive circuit (FLIC) which controls the personal protection means (PS) in response to the drive signal.

A computer program executing all the steps of a method according to any one of claims 1 to 10 when running on a controller (SG).

13. Computer program product with program code, which is stored on a machine-readable carrier for carrying out the method according to one of claims 1 to 10, when the program is executed on a control unit (SG).