EP2155520A1

EP2155520A1 - Method and control device for activating passenger protection means, and computer program and computer program product

Info

Publication number: EP2155520A1
Application number: EP08759692A
Authority: EP
Inventors: Alfons Doerr; Marcus Hiemer
Original assignee: Robert Bosch GmbH
Current assignee: Robert Bosch GmbH
Priority date: 2007-06-15
Filing date: 2008-05-16
Publication date: 2010-02-24
Also published as: DE102007027649B4; RU2010101000A; CN101678803A; WO2008151901A1; DE102007027649A1; CN101678803B; US20100305818A1

Abstract

The invention relates to a control device and a method for activating passenger protection means, wherein a feature vector is formed by at least two features of at least one signal of an accident sensor. Passenger protection means are activated by means of a core algorithm as a function of the feature vector, or of a first partial feature vector. The feature vector, or a second partial feature vector, is classified by means of a support vector machine (SVM), and the core algorithm is influenced by said classification.

Description

description

title

Method and control device for controlling personal protection devices and computer program and computer program product

State of the art

The invention relates to a method or a control device for controlling personal protection devices or a computer program or a computer program product according to the type of independent

Claims.

From DE 103 60 893 Al it is known to control personal protection means in dependence on a comparison of a forward displacement with a threshold value. The threshold value depends on a speed reduction and a

Delay set. 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 driving the personal protection devices, the threshold being the smallest limit.

In order to automate an application process and shorten the application time, automatic learning-based methods are proposed. One possible implementation is neural networks, as are known from WO 2005/037609 A1, WO 2005/037610 A1, WO 2005/037611 A1, WO 2005/035319 A1, EP 1133418 and DE 198 54 380 A1. In a training process that takes place offline in the laboratory, the otherwise manually applied dividing line is set automatically by machine learning methods. The algorithm based on neural networks ultimately provides a triggering decision based on a learned characteristic. The use of such neural networks is intransparent. There is no fallback in case of misclassification. Neural networks also require a large amount of training data that often does not exist. The problem of so-called overfittings, which is too strong a specialization of neural networks, is disadvantageous.

Disclosure of the invention

The inventive method or control device for controlling

Personal protective equipment, in contrast, has the advantage that a core algorithm known from the prior art is combined with a classification method, so that the strengths of both methods complement one another. As a classifier in the present case a support vector machine (SVM) is used. The SVM is trained in the laboratory. It provides a multidimensional separation surface, z. Between a trip area and a non-trip area; but possibly also between different crash classes such as ACT, ODB40kmh, ODB64kmh, 56kFull frontal, angle crash, etc. The real-world comparison of the crash data with the support vectors corresponding to the dividing line yields one

Classification of the crash signal. This classification creates an influence on the core algorithm so that its triggering performance is optimized.

This brings a number of advantages:

1. By combining the core algorithm with the

Classification procedures, the interfaces are identical to the outside, d. H. the data acquisition of the sensors and environmental parameters such as the buckle and the control of the

Personal protection can be carried out according to the proven principle. Even an existing security concept does not need to be changed. 2. By combining with the core algorithm, there is a physically safe fallback level, in case the classification was unsuccessful.

3. The separation surface found by the SVM divides the different ones

Crash classes optimal. The dividing line is therefore maximally robust with regard to the use of inexpensive hardware. For example, a simpler, less well-resolving sensor can be used.

4. The optimal dividing line or dividing surface, ie separating functions, is always found. It is said that the learning goal is always achieved. This is not the case with neural networks, for example. The optimization algorithm for defining the separation plane or interface can be stuck in a local minimum in neural networks. The quality of the separation function can therefore be very poor under certain circumstances. The support vector optimization features do not have this problem.

5. The classification is universally applicable. This is described in more detail in the dependent claims.

6. By using more than two dimensions, more crash information can be concurrently linked. For this reason, the classification quality is improved.

7. Learning-based classifiers such as the SVM can be assessed with objective quality measures. Thus, one can take advantage of the quantitative quality of the classifier, so that they can be transferred to the quality of an application and taken in numerical values.

8. Due to the automatic nature of the application, application time can be saved as the separation function is calculated automatically.

9. Due to the automatic nature of the application, a large number of numerical experiments can be carried out, which are even easier for the applicator no longer manageable. As a result of this increase in, for example, F EM simulation data or vehicle dynamics simulation data, the application can simply be extended to Real World scenarios via the previously used crash hall scenarios.

10. The separation function of the Support Vector machine replaces several additional functions. Selecting the right additional function is time-consuming in the standard application process. This time is saved by the proposed method.

11. Flexibilizing the algorithm decision-making for control saves on classifying computation time, which can be used for other computations, such as the merger of various additional functions.

12. The inventive method allows a reduction of the term, which will also be reflected in simple and thus cheaper hardware.

13. It is possible to react more flexibly to events during the crash, because some fire decisions take place later.

In the present case, the gist of the invention is the classification of the feature or partial feature vector by a support vector engine. The kernel algorithm is then influenced by this classification. The support vector machine is based on a statistical learning method, which is described in detail below.

Activation of personal protection means such as airbags, belt tensioners, roll bars or active personal protection means such as

Brakes or a vehicle dynamics control understood.

In the present case, a feature vector includes at least two features which are formed from a signal of an accident sensor system. If the signal is for example an acceleration signal, the acceleration signal itself or the first or second integral are used as features. From this, the vector is formed, which enters into the kernel algorithm on the one hand and into the support vector machine on the other hand. It is possible that only a part of the feature vector is included in the support vector machine. This is then designated by a partial feature vector. This also applies in the reverse case, ie, a feature vector is included in the support vector machine, while only a partial feature vector is included in the kernel algorithm.

In the present case, the accident sensor system may be an acceleration sensor system in and / or outside the control unit, which also applies to a structure-borne sound sensor system.

Furthermore, the accident sensor system can have an air pressure sensor system in the side parts of the vehicle and also an environment sensor system. Other accident sensors known to those skilled in the art may also be included here. The signal may include one or more readings from different sensors.

In the present case, the core algorithm is an algorithm which evaluates the feature vector in such a way that a control decision can be made. This can preferably be done by a threshold value decision.

A classification here means that the feature vector is assigned to a particular class. This class then determines how the kernel algorithm is affected. For example, classes can be divided according to the severity of the accident, ie how much the accident affects the occupants. A classification according to the crash type or a combination of crash type and crash severity can also be made.

The influence is described in more detail by the dependent claims. In particular, the decision to control is influenced, d. H. The classification leads to a triggering decision being taken in a first case which would not have occurred without the influence of the classification.

In the present case, a control device is understood to mean such a device which decides the activation of personal protection devices as a function of sensor signals. Therefore, the controller has means for evaluating the signals of Crash sensor system. To deliver the control signal, then a corresponding device in the control unit is necessary.

The at least one interface is realized in the present case by means of hardware and / or software. As software, it is designed, for example, on a microcontroller in the control unit as a software module.

The evaluation circuit is usually a microcontroller, but it can also be another type of processor such as a microprocessor or a signal processor. An integrated circuit, which includes the evaluation functions and, for example, is designed as an ASIC, can be used as an evaluation circuit. The evaluation circuit can also consist of discrete components or a combination of the aforementioned components. It is also possible to construct the evaluation circuit from a plurality of processors. For the individual tasks, the

Evaluation circuit then corresponding software modules, if it is a processor type such as a microcontroller, or there are corresponding hardware modules available. These can also be arranged on a single chip.

By the measures and refinements recited in the dependent claims advantageous improvements of the specified in the independent claims method for controlling personal protection means are possible.

It is advantageous that the core algorithm forms a decision for the drive by comparing the feature vector with a first threshold value in an at least two-dimensional feature space. Thus, the kernel algorithm is designed such that it contains the feature vector with the at least two features in an at least two-dimensional

Characteristic space transferred and compared there with a threshold, the threshold may also be a function. Thus, a time-invariant kernel algorithm is realized, wherein as the features, for example, the delay and the first integral of the delay, so the speed can be used. But other variables such as the forward displacement, So twice the integral of the delay, can be used here.

Furthermore, it is advantageous that the kernel algorithm is influenced by the classification in that the first threshold value depends on the

Classification is changed. By changing this threshold, the classification directly intervenes in the decision-making as to whether or not the personal protective equipment should be activated. This change may be made by a penalty or discount depending on the classification or by replacing the first threshold with a second threshold. In this case, the second threshold value is stored, for example, or it is calculated.

It is also possible that, depending on the classification, in addition to the influence on the core algorithm, a plausibility check of the

Driving decision is performed. It is decided based on the classification, whether or not there is a triggering case for the personal protective equipment. This result is then combined with the decision of the kernel algorithm to arrive at a secure overall decision. Additional functions can also contribute to the combination. These additional functions include, for example, the processing of further sensor signals or a crash type detection.

A plausibility check means that a first decision is confirmed or revoked by a second decision. This is a safer overall

Final decision possible.

It is furthermore advantageous that a misuse is detected as a function of the classification and the kernel algorithm takes this into account in the activation. A misuse is an impact that does not trigger

Personal security should lead. Thus, a triggering decision that meets the core algorithm can still be prevented. This too can be determined depending on the respective classification. The classification can also be used as a supplement to an existing misuse classification. Again, the classification can be an add-on provide for shifting a misuse threshold or, for example, act as a misuse plausibility function.

It is also advantageous that, depending on the classification, a very serious crash is detected. A very heavy crash usually has to activate all the necessary front passenger protection devices, these are the belt tensioners and the first and second airbag stages. If the core algorithm classifies a control and the SVM classifies a very serious crash, then the SVM classification can initiate the activation of all front personal protection devices by a control circuit.

Furthermore, it is advantageous for a computer program to be present which carries out all the steps of the method according to any one of claims 1 to 7 when it runs on a control unit. This computer program may originally be written in a programming language and is then translated into machine-readable code.

Advantageously, a computer program product in the program code, which is stored on a machine-readable carrier such as a semiconductor memory, a hard disk memory or an optical memory and the

Implementation of the method according to one of claims 1 to 7 is used when the program is executed on a control unit.

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 control device according to the invention with connected components,

FIG. 2 shows different software modules on the microcontroller,

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

FIG. 4 shows a first signal flow diagram,

FIG. 5 shows a second signal flow diagram, FIG. 6 shows a third signal flow diagram, 7 shows a dividing line between two classes in the SVM, FIG. 9 shows a dividing line in the starting room and FIG. 10 shows the dividing line in the image space, and FIG. 11 shows a diagram for explaining the training procedure by deliberately simultaneous inputting of input and output data.

Central to the present invention is the use of the Support Vector Machine (SVM) as a classifier of the feature vector. This will be explained in more detail below.

In the following, the classification principle of the SVM for two classes will be described, for example, for the separation of Fire and NoFire crashes. In principle, it can easily be transferred to the classification of several classes.

A detailed description of SVM can be found in the pertinent literature (e.g., Cristianini, Neil and Shawe-Taylor, John: An Introduction to Support Vector Machines and Other Kernel-Based Learning Methods or Hastie, Trevor: The Elements of Statistical Learning).

Multiclass Support Vector Classification is described, for example, in Schölkopf, Bernhard et al .: Extracting Support Data for a Given Task, Proceedings of the First International Conference on Knowledge Discovery and Data Mining, AAAI Press, Menlo Park, CA, 1995, pages 252-257.

At this point, only the principle should be described qualitatively.

Linear separation

The Support Vector machine is a linear classifier. The linear dividing line has the following form: f (x) = w- x + b (1) The goal is to lay a dividing line between the two classes to be classified, which is optimal in terms of the distance of the training data (Figure 8). In Fig. 8, this is the solid line 84. Although the two finer separation lines 80, 81 separate also, but not optimal in terms of robustness. Only the dividing line 84 provides maximum robustness and enables the use of simpler and therefore cheaper hardware described in point 3 of the "advantages of the invention".

Finding the optimal straight lines for class separation is known in mathematics as a "quadratic problem with linear constraints." A quadratic problem with linear constraints can be efficiently solved by quadratic programming algorithms (point 3 of the "Advantages of the Invention"), (See, for example, "R. Vanderbei, LOQO: An Interior Point Code for Quadratic Programming.") A great advantage of this is the fact that this optimal solution is always found by the algorithms, so there is no danger in a local minimum to stick to the optimization (point 4 from the "Benefits of

As a result of the optimization, one has finally the characteristic curve shown in FIG.

In mathematics equation (1) can be represented by the so-called "dual form":

/ (*) = Σ>, α, ^■ *, ^■ * + *> (2)

Both representations are identical. The yi are the class affiliation of the training datum i (usually +1 or -1), xi represents the so-called support vectors, x are the e.g. Characteristics to be classified in a crash. In FIG. 8, the support vectors can be seen as those features shown on dashed line 82,

83 lie. In a sense, they represent those support vectors which "lie closest to the other class." Considering equation (2), the factors αi, the so-called Lagrange multipliers, have not yet been discussed the support vectors are different from zero, in other words, the equation (2) must only be applied to the

Support vectors are evaluated. To illustrate even more clearly: new features which, for example, are added during the crash, no longer have to be evaluated in relation to the entire solid dividing line 84 illustrated in FIG. but only with respect to the support vectors on the dashed lines 82, 83. The number of support vectors can be kept small by the method and thus the computational effort in the ECU be limited.

In summary, it can be said: the support vector algorithm that is used in the

Training is always an optimal, i. maximum robust dividing line of the two classes. After the training, in the test or in the crash, the generated features are not evaluated with respect to the entire dividing line but only in relation to the (significantly less) support vectors.

Nonlinear separation

In reality, the classes will normally not be linearly separable, but will only be separable by a non-linear dividing line. For this reason, the so-called "Kernel-trick" is used

Transformation by means of a kernel is obtained from the output space (x1, x2) in FIG. 9, which is described by two of the three features 1... 3 in FIG. 7 in the so-called image space (z1, z2, z3) in FIG. 10. With 90 in FIG. 9, the non-linear parting line in the exit space and 10 the corresponding linear parting line in FIG. 10 are designated.

In the image space, the features are again linearly separable (see FIGS. 9 and 10) and equation 2 can be used again: the algorithm for finding the optimal linear separation line in the image space, which always optimally converges. The kernel trick now has the following advantage: the transformation into the image space does not take place explicitly, that is, one does not really count in the image space. One uses only the mathematical "kernel function" to achieve a linear separability in the image space, but on the other hand, every calculation still takes place in the output space, and equation (2) then becomes the case for the nonlinear case

f (x) = γ _j yμ _i - k (x ₁ , x) + b (3)

The kernel function k (x "x) must satisfy some mathematical requirements, such as: B. Cristianini, Neil and Shawe-Taylor, John: "An introduction to support vector Machines and other kernel-based learning methods. "The following standard kernels are normally used as kernel functions:

Radial basis kernel

Polynomkernel

Sigmoidalkernel

It should be expressly understood that the described invention is not dependent on the kernel function. As can easily be seen from equation (3), the usually nonlinear kernel function k (xi, x) must also be calculated exclusively on the support vectors. For the example of a radial base kernel this means: the distance of the features x only has to be calculated to the support vectors xi. The e-function could be in the control unit, for example by a

Taylor approximation or realized by a look-up table. The parameter σ in equation (4) makes it possible to influence the robustness of the classifier and thus the number of support vectors.

In summary, it can be said that by using the kernel

Tricks and non-linear characteristics can be optimally separated without having to explicitly perform the transformation into the image space. The kernel function and the formula (3) for the separation only have to be evaluated with regard to the support vectors.

Slack variable

By using so-called Slack variables, the robustness of the classification can be further increased. Slack variables may allow misclassifications to be tolerated. For this purpose, incorrectly classified features are weighted by a factor C:

Since it may be advantageous to penalize the misclassification of one class more than that of another class (eg, for example, it may be more tolerant that a must-fire be classified as no-fire than vice versa), equation (5) should be extended :

^G = ^c w ^' Σξ * + ^C Hϊ ^' Σξ * ^{with c} in » ^c ι + n ^■ ( ^ß )

Vie {+1} Vie {-1}

Equation (6) causes class -1 fault classifications (eg "NoFire") to be weighted much more heavily than those of class +1 ("MustFire"). Allowing misclassifications can also affect the number of support vectors, and thus indirectly the computational time. When using Slack

Variables, the applicator a priori still bring knowledge about his data. If he is aware that the data is very scattered, misclassifications can be tolerable.

training

As with all learning-based methods, a training phase also takes place in the support vector machine before the actual ECU application (see FIG. 11). This takes place offline. It is used to determine the support vectors, which are then stored in the control unit. During training, the classifier 111 receives one

110 and output data 112 supplied in pairs. As input data, the three features of Fig. 7 could be used. Output data could be, for example, the desired trigger times. Care must be taken to use a well-balanced crash set during training and to take due account of the usual robustness criteria, such as amplitude and offset variations.

The support vectors determined during training must then be placed in the control unit.

validation Frequently, in particular in an early airbag project phase, insufficient crash data are available. Cross-validation methods can increase the amount of training and increase the safety of the classification. Cross Validation splits the existing crashset into subsets. Some subsets then serve as training data, others are used to assess the classification quality. The most well-known of these methods is likely to be the Leave One Out Cross Validation, which always uses exactly one record for the test and used all other records previously for training. By permuting this one test dataset through the total set of datasets, one obtains a very large number of tests for the classification and, using a statistical evaluation, can determine the quality measures for the classifier described in item 7 of the "Advantages of the Invention" On the basis of the quality measures, an optimization of the classification parameters, for example σ in equation (4), continues to be carried out.

FIG. 1 illustrates in a block diagram the control unit SG according to the invention with connected components. In a vehicle FZ, the control unit SG is arranged to which various components are connected. By way of example, in the present case, only the components necessary for understanding the invention are shown both outside and inside the control unit.

To the control unit SG different accident sensors are connected as a structure-borne sound sensor KS, an acceleration sensor BSL, a

Pressure sensor DS and an environment sensor US. Other sensors such as a vehicle dynamics sensor and / or yaw rate sensors, etc. may be additionally or instead connected. Various installation positions in the vehicle FZ are known to the skilled person. The structure-borne noise sensor system and the acceleration sensor system BS1 are connected to a first interface I F1, the interface IF1 providing these signals to the evaluation circuit, namely the microcontroller μC. A second interface I F2, to which the air pressure sensor DS and the environmental sensor system US are connected, provides these signals to the microcontroller .mu.C. The DS air pressure sensor is installed in the side panels of the vehicle and is intended to serve as a side impact sensor. The environmental sensor system US may include various environmental sensors such as radar, LIDAR, video or ultrasound to analyze the environment of the vehicle FZ with respect to collision objects. The microcontroller μC receives from an acceleration sensor

BS2 within the control unit SG further sensor signals. Additional sensors can be located inside the controller and send signals to the microcontroller μC. These include vehicle dynamics sensors and structure-borne noise sensors.

The control unit SG in this case has a housing made of metal and / or

Plastic can be made. The microcontroller .mu.C has internal memory of its own, but can also access external memories which are also located in the control unit SG. By means of a core algorithm located in the memory, the microcontroller .mu.C evaluates a feature vector from features of these crash signals and decides whether the personal protection means PS which are connected via the

Control circuit FLIC are controlled to be controlled. For this purpose, the core algorithm is influenced by a support vector machine with a classification of the feature vector. This influence ensures that the decision is more accurate and appropriate.

It is possible that more or less than the illustrated sensors may be used. The communication of the interfaces I Fl and I F2 to the microcontroller .mu.C can be done, for example, via the control unit-internal bus SPI (serial peripheral interface bus). The SPI bus can also be used for communication between the microcontroller μC and the drive circuit

FLIC can be used. In the present case, the drive circuit FLIC consists of several integrated circuits which have power switches and, in the case of activation, enable the ignition or control elements of the personal protection means PS to be energized. This drive circuit may also have different forms, which consist of one or more integrated circuits and / or discrete components.

FIG. 2 now shows software modules which are necessary for the function of the invention and are present on the evaluation circuit in the microcontroller .mu.C. The microcontroller μC usually has its own memory. However, it can also be a memory connected to the microcontroller .mu.C via lines. An interface I F3 is used to connect the acceleration sensor BS2 and provides the signals of this acceleration sensor BS2 ready. These signals are recorded on the one hand by the feature module M, which consists of the signals of accident sensors

Features and formed from the features of the feature vector, for example, the signal is the acceleration signal and the module M determines therefrom by simple integration, the speed and then from the acceleration and the speed forms a two-dimensional feature vector.

This feature vector, which can also have more dimensions, depending on how many features are to be included, then goes firstly to the module SVM, which contains the support vector machine, and secondly to the kernel algorithm K. It is possible that the feature module M only one

Part of the module SVM provides because only a part of the characteristics for the classification is necessary. The same applies to the core algorithm. The module SVM now classifies the feature vector with the support vector machine. This classification result is also provided in the kernel algorithm K. It is possible that this

Classification result can also be provided to other modules not shown here. For example, the classification result can be used as a plausibility for a trigger decision, which is obtained from another part of the algorithm. It is also conceivable that the classification result is used for the control of the further algorithm processing. Conceivable, for example, the targeted switching on and off of functionality. The kernel algorithm now influences the evaluation of the feature module with the classification result, whether the activation of the personal protection means PS should take place or not. If it is decided that the personal protection means should be controlled, then the module A is activated for the activation in order to generate a drive signal with the hardware of the microcontroller .mu.C and to transmit it to the drive circuit FLIC. This transmission can be particularly secure when it happens over the SPI bus. FIG. 3 illustrates in a first flowchart the sequence of the method according to the invention. In step 300, the signal of the accident, environment and / or driving dynamics sensor is provided. And through the interfaces I Fl, I F2 or I F3. From this, method step 301 is then used to shape the feature vector in the manner described above. This

The feature vector is completely fed to the kernel algorithm 303 and completely or partially the support vector machine 302. The support vector machine classifies the feature vector or partial feature vector and transmits this classification result to the kernel algorithm 303. The kernel algorithm 303 decides the activation of the personal protection means PS in dependence on the feature vector and the classification result. Control then takes place in method step 304.

FIG. 4 shows a further signal flow diagram. In block 400, the feature vector is provided and provided to the kernel algorithm 401 which spans a two-dimensional decision space here from the acceleration A and velocity DV, where A is plotted on the abscissa and DV on the ordinate. The threshold value 408 separates the trigger case 403 from the non-trigger case 402. The feature vector is entered in this decision space, and it is checked whether the

Feature vector is above the threshold 408 or below. Depending on this, the output is then made that an activation is to take place, namely at the block 406. In parallel, the feature vector 400 has been made available to the support vector machine SVM in block 404, with the support vector machine performing the classification. This classification influences, for example, the threshold value 408. However, it is also possible to carry out a plausibility check in block 405 from the classification. H. It is checked whether the classification indicates that a trigger case exists. The result of the plausibility check and the kernel algorithm 401 is linked in block 406. If this link indicates a triggering case, this is done in the block

407 the control.

FIG. 5 shows a further signal flow diagram. Only a part is shown here. The support vector engine 500 provides its classification to a search algorithm 501 which is at a lookup table after a Threshold depending on the classification searches and loads it and then the core algorithm 502 provides.

FIG. 6 shows a further detail of the signal flow diagram. The support vector machine in turn classifies the feature vector. This leads in the block

601 at a surcharge or discount for the threshold value, which is supplied to the core algorithm 602, so that here the threshold 603, the supplement 604 is supplied.

Fig. 7 shows a signal flow diagram of the method according to the invention.

Features Ml-3 generated from the accident sensor signal (s) are applied to the support vector machine 70 for classifying the feature vector formed from features Ml-3. These features Ml-3 or a subset of the features Ml-3 and possibly other features, even from different sensors are the core algorithm

71, which falls in response to these overall characteristics of the drive decision. However, this drive decision is also affected by the classification by the support vector engine 70. The influencing is carried out, for example, by a threshold value change as a function of the classification. In this case, a respective

Class lead to a predetermined surcharge or discount or for a particular class, a respective threshold is loaded.

In addition to or instead of the classification, a separate plausibility decision can be made, the result of this

Plausibility and the decision of the core algorithm then be linked together to ultimately make the driving decision.

Claims

claims

1. A method for controlling personal protection means (PS), wherein a feature vector (M) with at least two features from at least one signal of an accident sensor system (BSl, BS2, DS, U) is formed, wherein by a kernel algorithm (K) in dependence on the Feature vector (M) or by a first partial feature vector the personal protection means (PS) are driven, characterized in that the feature vector (M) or a second partial feature vector by a support vector machine (SVM) is classified and the core algorithm (K) influenced by the classification becomes.

2. The method according to claim 1, characterized in that the kernel algorithm (K) forms a decision for the drive in that the feature vector (M) or the first partial feature vector is compared with a first threshold value in an at least two-dimensional feature space.

3. The method according to claim 2, characterized in that the kernel algorithm (K) is influenced by the classification in that the first threshold value is changed depending on the classification.

4. The method according to claim 3, characterized in that the change of the first threshold value is made by a surcharge or discount or by a replacement of the first threshold value by a second threshold value.

5. The method according to any one of the preceding claims, characterized in that carried out depending on the classification, a plausibility control, wherein the kernel algorithm (K) takes into account the plausibility of the control.

6. The method according to any one of the preceding claims, characterized in that a misuse is detected in dependence on the classification and the kernel algorithm (K) takes this into account in the control.

7. The method according to any one of the preceding claims, characterized in that, depending on the classification, a very serious crash is detected.

8. The method according to claim 7, characterized in that the support vector machine (SVM) allows misclassifications and that the misclassifications of different classes get different weight.

9. Control device (SG) for controlling personal protection devices:

- At least one interface (I Fl, I F2, I F3), which provides at least one signal of accident sensors;

- An evaluation circuit (.mu.C), which forms a feature vector (M) having at least two features of the at least one signal, wherein the

Evaluation circuit (.mu.C) has a core algorithm (K) which, depending on the feature vector (M) or of a first partial feature vector, activates the personal protection device (PS), characterized in that the evaluation circuit (.mu.C) has a support vector machine (SVM), which classifies the feature vector (M) or a second partial feature vector and influences the kernel algorithm (K) as a function of the classification.

A computer program executing all steps of a method according to any one of claims 1 to 8 when running on a controller (SG).

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 8, when the program is executed on a control unit (SG).