EP2276650A1

EP2276650A1 - Method and control unit for activating at least one security means

Info

Publication number: EP2276650A1
Application number: EP09731577A
Authority: EP
Inventors: Marcus Hiemer; Mike Schwarz
Original assignee: Robert Bosch GmbH
Current assignee: Robert Bosch GmbH
Priority date: 2008-04-16
Filing date: 2009-02-16
Publication date: 2011-01-26
Also published as: DE102008001215A1; CN102007019A; US20110153164A1; WO2009127453A1

Abstract

A method is proposed for activating at least one security means (FDR, PS1), which comprises a first step of acquiring at least two features (M1, M2, M3) from at least one signal of an accident sensor system, in order to calculate a feature vector from the acquired features (M1, M2, M3). Subsequently, in a second method step, a classification of the calculated feature vector is performed with the aid of a classifier (MSVM) on the basis of the statistical learning theory, in order to classify the feature vector in one of at least three possible feature classes (K1, K2, K3). An activation of the security means (PS1, FDR1) according to an activation guidelines (T1, FLIC1, T2, T2, FLIC2) for the feature class (K1, K2, K3) in which the feature vector was classified is provided as a third method step.

Description

description

title

Method and control device for driving at least one security means

State of the art

The present invention relates to a method for controlling at least one security means according to claim 1, a control device for controlling at least one security means according to claim 10, a computer program according to claim 11 and a computer program product according to claim 12.

Existing triggering algorithms for personal protective equipment, such as airbags, do not or only too imprecisely evaluate a history of the motion of a vehicle in order to perform an optimal release of safety devices. In existing systems, a deployment decision is made based essentially on the acceleration signals that occur in the event of an accident and are measured via acceleration sensors. Predictive systems such as PreCrash try to pre-condition the triggering algorithms, for example, with RA-DAR or LASER sensors, but these signals are not or only insufficiently combined with the information of the other sensors installed in the vehicle, so that a combined analysis of the data All of the principle already available sensor signals for reasons of complexity of such evaluation is currently not done.

In the future, in order to improve the safety of the occupants of a vehicle, a merger of the evaluation of the active and passive safety components will take place, which until now have worked separately. As part of this merger, the number of requirements for a security system will increase significantly since the arrival ^• number of driving situations to be considered will increase expoπentiell. This increasing number of driving situations to be considered should be countered by combining and evaluating as many signals as possible available in the vehicle. In contrast to conventional triggering systems, it would be beneficial to take into account the history of the movement of the vehicle under consideration and to efficiently use or combine information from the existing sensors. However, this currently requires a very complex circuit structure for the evaluation of the available signals.

DE 10 2006 038151 A1 discloses a device and a method for controlling personal protection devices, in which the activation of the personal protection means is carried out using a support vector machine. The evaluation of an accident sensor signal is carried out using different classification trees, in which a binary classification is performed. Compared to the technique of "neural networks", the use of the support vector machine has the advantage that for each classification problem an optimal solution can be found, which can also be determined relatively easily.

In the patent application D E102007027649 a method and control device for controlling personal protection means is disclosed, in which a decision algorithm for evaluating features of an accident sensor signal and a support vector machine are used to control personal protection means, the multi-dimensional classification of further, additional Processing features for the decision algorithm. In this case, a classification of the additional features into two different classes is used.

In both documents, however, a further optimization of the release decision of security means by the trigger circuit is possible to efficiently combine and evaluate several signals of accident sensors (or derived variables). In this way, the sensor signals already available in vehicles can be better evaluated from a history of the vehicle movement for the control of the safety means.

Description of the invention

It is therefore the object of the present invention to provide a possibility for improved evaluation of the available sensor signals.

This object is achieved by a method according to claim 1, a control device according to claim 10, a .Computer program according to claim 11 and a computer program product according to claim 12. Advantageous embodiments of the invention are the subject of the dependent claims. The present invention provides a method for controlling at least one security agent, the method comprising the following steps:

Detecting at least two features from at least one signal of an accident sensor system in order to form a feature vector from the detected features;

Classifying the formed feature vector using a classifier based on the statistical learning theory to classify the feature vector into one of at least three possible feature classes; and

- Activation of the security agent according to a tax code for the feature class in which the feature vector has been classified.

The invention is based on the recognition that a classification of signal characteristics of an accident sensor signal in more than two classes ^' by a classifier on the basis of statistical learning theory a significant improvement of the linking possibilities and the rapid evaluation of such signal characteristics is possible. Essentially, this optimization is based on the fact that the various classes can be used to classify directly into more than two classes and that the signal characteristics can be well processed or separated, which simplifies signal processing in a drive unit following in the signal path. Since the classifiers on the basis of statistical learning theory on the one hand numerically efficient and fast work and on the other hand can process a large number of signal characteristics, can be optimally evaluated by using such classifiers also a large amount of already available in the vehicle accident sensor signals , This speeds up the desired fusion of the active and passive safety systems with their corresponding sensors.

In particular, when using more than two feature classes, an accident classification of a collision of vehicles can be carried out better and, above all, more accurately than in the prior art. In order to. For example, a more precise reaction of the vehicle safety system to the collision of two vehicles or even before such a collision of two vehicles can be realized. This makes it possible to initiate exactly the appropriate countermeasure against such an accident scenario by an exact addition of the required for a respective type of accident security means. According to an advantageous embodiment of the invention, classifying using the classifier based statistical theorem theory comprises using a multi-class support vector machine. The use of such a multi-class support vector machine provides an excellent choice for a fast, numerical, or circuit efficient, and most importantly, precise, classifier based on statistical learning theory.

In a further embodiment of the invention, the activation of the security device in accordance with a driving instruction for a first feature class may include activating a personal protection device. Furthermore, the activation of the security device in accordance with a driving instruction for a second feature class may also include the activation of a driving dynamics assistance control. This advantageously ensures that the features extracted from a single accident signal are used for different safety functions in a vehicle, so that merging of the active and passive safety components in the vehicle is simplified by the approach proposed here. At the same time a fast and precise classification is achieved in particular by the use of the multi-class support vector machine, a reduction of the computational or circuit complexity for the operation of the corresponding trip unit of the individual Sicherheitsmitte! allowed in the vehicle.

It is also favorable if the actuation of the safety device also takes place by using at least one feature of the feature vector or of a further feature from a signal of the accident sensor system. In this way, the one feature from the feature vector itself or the feature of the signal from the crash sensor system can be used in a physical core algorithm, which forms a fallback during the control of the corresponding safety means. In this way, it is made possible that tripping occurs reliably even in the event of a malfunction of said classifier, wherein an improvement and / or specification of the triggering of the corresponding safety means can then be implemented by the classifier described above. This implies an exclusive gain in security in the implementation of this embodiment of the invention.

Furthermore, it is advantageously possible that a classification function value is determined during classification and that the activation of the security means is performed using the

Classification function value takes place. This represents a further refinement of the classification result, since not only a class as such, but also a differentiation of the triggering within a class is possible. Such a differentiation on the basis of the classification function value then makes possible an even more precise control of the corresponding safety means, for example by a graduated activation of different airbag stages.

According to another embodiment of the present invention, the driving of the security means may be carried out according to an activation rule based on a decision threshold value. As a result, a very simple and numerically or circuitry technically simple to implement control rule is implemented, so that only components with low complexity to implement the invention according to this embodiment are to be used.

Furthermore, in the step of the Ansteuems the tax law can be changed according to a feature class dependent on the change rule. By such a change in the tax code depending on the particular

Feature class can be prompted in a simple and above all very fast way the control of the safety device. As a result, by linking and simple modification of the security means or their components which are usually already present in a vehicle, a significant increase in occupant safety can be achieved

Use this product to achieve the correctness.

Specifically, in the step of the activation, depending on the feature class, an increase or decrease of the decision threshold value can take place or the decision threshold value can be replaced by a second decision threshold value. This easy-to-implement change of the driving instruction can very well improve the safety of the occupants of a vehicle by classifying a feature vector into several (especially more than three) classes. In this case, changing or replacing the decision threshold value also requires only slight changes to the structure of the corresponding security means or its associated triggering circuit.

In another embodiment of the invention, classifying may be based on class boundaries between the feature classes loaded from a memory. In this case, a classifier, for example, is pre-trained in the manufacturer's laboratory and already optimally based on accident scenarios or simulations set, after which the trained parameters are then stored in a memory. As a result, a classifier is obtained which operates quickly and precisely during operation, since a complex adaptation of the settings of the classifier during operation is no longer required.

In order to realize the advantages according to the invention, in a further embodiment of the invention, a control device for controlling at least one safety means may also be provided which comprises the following features:

at least one interface, which is designed to form a feature vector of at least two features from at least one signal of an accident sensor system; an evaluation circuit designed to classify the feature vector formed into one of at least three possible feature classes using a classifier based on statistical learning theory; and a drive unit configured to drive the safety means in accordance with a drive instruction for the feature class into which the feature vector has been classified.

Also by this embodiment of the invention in the form of a device, the object underlying the invention can be solved quickly and efficiently. In particular, by combining the use of a classifier on the basis of statistical learning theory with the possibility of classifying the feature vector into one of at least three feature classes, a more precise, faster, and thus better than the prior art, evaluation of the available sensor signals is achieved possible.

In a further embodiment of the invention, a computer program is provided which executes all the steps of the method according to one of the embodiments described above, when it runs on a control unit. This computer program may originally be written in a high-level language and is then translated into machine-readable code.

Also of advantage is a computer program product with program code which is stored on a machine-readable carrier such as a semiconductor memory, a hard disk memory or an optical memory and for carrying out the method according to nem the Ausführuπgsformen described above is used when the program is executed on a control unit.

Brief description of the drawings

The invention will now be described by way of example with reference to the accompanying drawings. Show it:

1 is a block diagram of a first Ausführuπgsbeispieis of the present invention as a unit installed in a vehicle.

Fig. 2 is a block diagram of a second embodiment of the present invention

Invention; 3 is a block diagram of a third embodiment of the present invention

Invention; 4 is a block diagram of a fourth embodiment of the present invention

Invention; and FIG. 5 is a flow chart of a fifth embodiment of the present invention

Invention.

The same or similar elements may be provided in the following figures by the same or similar reference numerals. Furthermore, the figures of the drawings, the description and the claims contain numerous features in combination. It is clear to a person skilled in the art that these features are also considered individually or that they can be combined to form further combinations not explicitly described here.

FIG. 1 shows a block diagram of a first exemplary embodiment of the present invention. With the aid of this block diagram, the control unit SG according to the invention is explained in more detail with connected components. In a vehicle FZ, the control unit SG is arranged, which is connected to various components. By way of example, only the components necessary for understanding the invention are shown outside and inside the control unit.

To the control unit SG different accident sensors are connected, such as a structure-borne sound sensor KS, a Beschleuniguπgssensorik BS1, a pressure sensor DS and an environment sensor US. Other sensors, such as a driving dynamics sensor and / or rotational rate sensors, etc. may be connected in addition to or instead of the sensors described above. Various installation positions in the vehicle FZ are known to the skilled person. The structure-borne sound sensor KS and the acceleration sensor system BS1 are connected to a first interface IF1 of the control unit SG, wherein the interface 1F1 supplies the signals of the evaluation circuit .mu.C, which is designed according to the first embodiment as a microcontroller .mu.C. Alternatively, the evaluation circuit .mu.C can be another element having a data processing functionality such as a general microprocessor, a digital signal processor DSP, an application specific integrated circuit ASIC (ASIC) or a programmable logic device FPGA (FPGA = field programmable gate array ~ on-site programmable logic device). A second interface IF2, to which, for example, an air pressure sensor DS and the environmental sensor system US are connected, provides these signals to the evaluation circuit .mu.C. The air pressure sensor DS can also be installed in the side parts of the vehicle and should then serve as a side impact sensor. The environment 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 .mu.C contains further sensor signals from an acceleration sensor BS2 within the control unit SG via an internal control unit interface. Additional sensors can be located within the control unit SG and deliver signals to the MTkrαcontroller μC via appropriate interfaces. These include driving dynamics sensors and / or structure-borne noise sensors.

The interfaces receive the signals from the accident sensors, extract certain characteristics from this accident sensor signal such as an acceleration, an Integra! a speed, etc., and combine a certain number of these features into a feature vector. For example, the signal can be an acceleration signal and one of the interfaces can determine the speed by simple integration and then form the acceleration and the velocity into a two-dimensional feature vector, which is made available to the evaluation circuit, especially the classifier.

In the microcontroller μC a classifier based on the statistical learning theory is arranged, which will be explained in more detail below. This classifier is supplied with the feature vector, whereby the classifier also processes a multi-dimensional feature vector, depending on how many features are to be included in the classification can. The classifier divides the feature vector into one of at least three feature classes K1, K2 or K3. These feature classes characterize, for example, different types of accidents or accidents, so that for each type of accident or for each severity of severity the control of correspondingly suitable safety means can take place. For example, the activation of a first personal protection device PS1 in the form of an airbag can be effected by the first drive circuit FL1C1 if the classifier in the microcontroller .mu.C has classified the feature vector into the feature class K1. Similarly, a second personal protection means PS2 (for example, a belt tensioner) can be activated by a second drive circuit FLIC2 if the classifier in the microcontroller .mu.C has arranged the feature vector in the second feature class K2. In the event that the classifier in the microcontroller .mu.C has arranged the feature vector in the third feature class K3, a vehicle dynamics control FDR (such as, for example, an ESP control) can be activated via a third drive circuit FL1C3. This transfer can be particularly secure if it is done via the SPI bus (serial peripheral interface bus). This activation of the corresponding drive circuits can be done very easily by, for example, by the classification in a (feature JKIasse K1, K2 or K3 only a (binary) on / off Akti vation of the corresponding drive circuit takes place, which is quickly and inexpensively evaluated ,

In the present case, the control unit SG has a housing that can be made of metal and / or plastic. The microcontroller μC itself has an internal memory, but can also access external memory, which is also located in the control unit SG. The memory can also store class boundaries, which were determined, for example, in a pre-training of the classifier in the laboratory, as will be described in more detail below. Using these class boundaries, the classifier in the microcontroller .mu.C can implement the feature vector very quickly and easily in the different feature classes K1, K2 or K3.

It is possible that more or less than the illustrated sensors are used. The communication of the interfaces IF1 and IF2 to the microcontroller .mu.C can be done, for example, via the controller-internal bus SPI. The SPI bus can also be used for the communication between the microcontroller .mu.C and the drive circuits FLIC1, FL1C2 and FLIC3. In the present case, the drive circuits FLIC1, FLIC2 and FLIC3 consist of one or more integrated circuits which, for example, have power switches and, in the case of actuation, energization of the ignition or ignition circuits Activation elements of the personal protection means PS1 or PS2 or of the driving dynamics controller FDR. These personal protection means PS1 or PS2 or the driving dynamics controller FDR can also have different characteristics, which consist of one or more integrated circuits and / or discrete components.

Classifiers based on statistical learning theory, which subdivide a feature vector into one of at least three feature classes, are particularly suitable for use in the present invention. In this way, it is possible to evaluate the multiplicity of signals and signal combinations to be considered by means of an automated method. The automatic evaluation makes it possible that the signal combinations due to the limited manageable data volume should no longer result exclusively from actually driven driving tests (such as the standardized crash hail tests EuroNCAP, ...), but increasingly results of driving dynamics and FEM simulations (FEM = finite element model) can be processed. If no automatic evaluation were used, the large number of sigma combinations to be considered would not be manageable. Thus, by mapping a real-world safety development process, the automatic evaluation of any simulated accident situations can advantageously take place, allowing an even better training of the classifier.

It should also be noted that the classification of the feature vector into one of more than two feature classes results in a more detailed evaluation of the features of accident signals than is possible in the prior art. In particular by the automated evaluation of the history of a vehicle movement in a very early unfail phase, different branches of the further algorithm processing can be activated by the feature class-based control based on the finely graduated classification of the characteristics of an accident signal into many feature classes. Due to the advantageous early classification of the characteristics of an accident signal and the correspondingly fast control of the most suitable safety means, on the one hand, the reaction time of the safety means is reduced and, on the other hand, only those safety devices are actuated that are actually affected by the current driving situation. This saves resources.

Due to the possibility of combining features of a large number of available accident signals, it is also possible to continue to map or react to driving situations caused by active and passive safety components were. In addition, the classifier can be easily tuned to customer-requested hazard and accident situations by using simulation data. The use of machine-lime-based processes also significantly reduces the application time, so that trajectories with comprehensive signal combinations for training the cassifier in a laboratory can be carried out in a realistic time. Compared to classifiers from the prior art, therefore, the proposed classifier can be trained much better, which is advantageous in its use in a more precise selection of the correct feature class for a given feature vector.

In the present invention, as a classifier based on the statistical learning theory, specifically, a multi-class support vector machine (MSVM) is used because such a multi-class support vector machine as well as a support vector machine always provides an optimal solution and a low tendency for specialization (ie memorizing the trained data).

A detailed operation of a support vector machine (SVM) as the basis of the multi-class support vector machine can be found, for example, in the patent application DE102007027649. Further information on SVMs can also be found in the relevant literature (eg Cristianini, NeIIo and Shawe-Taylor, John: "An introductory to support vector machines and other kernel-based learning methods" or Hastie: "The Elements of Statistical Learning"). ). To avoid redundancies, a detailed description of the functionality of SVMs should be omitted here.

Unlike conventional support vector machines that are between two classes

is different (for example, when accident discrimination between a feature class "Fire" for activating a safety means and a feature class "no-fire" to non-activation of the corresponding safety means or between the Merkmalsklasseπ "ODB7 _D NichtODB ^H (ODB = offset deformable barrier), the Mehrklassen- is support Vector Machine (MSVM) is able to distinguish several classes, especially more than 3 classes.The MSVM is also a machine learning based method of the class of statistical learning theory in which the classifier is trained by pairwise specification of feature vectors and associated class. The training of such an MSVM will be discussed in more detail below. The concrete use of a classifier on the basis of the statistical learning theory is briefly sketched in FIG. 2 as a block diagram illustration of a second exemplary embodiment of the present invention. Here, the classifier, which is arranged, for example, in the microcontroller .mu.C, M1 and M2 of accident signals (for example, with respect to a wheel speed, a yaw acceleration, an integral of the longitudinal acceleration or a degree of coverage of an accident sensor (PreCrash sensor)) and receive trained such that the individual feature classes K1 to KN map different vehicle states (such as "skidding", "frontal crash", "light side crash soft crash", ...) Accordingly, different control circuits or algorithms parts or control rules for controlling For example, when classifying the feature vector from the features M1 and M2 into the feature class K1, a first sub-algorithm T1 can be activated as a triggering algorithm in the microcontroller .mu.C, which then uses a drive control FLIC for a front airbag to protect the person tel PS1 in the form of squibs, rev. Restraints (rev. Restraint = reversible retention means) or the like activated. Also conceivable would be a separate trigger circuit, in which the functionalities of the first sub-algorithm T1 and. the driving circuit FLIC are implemented, which is activated by the classification of the characteristics M1 and M2 into the first class K1 and which in turn activates the personal protection means PS1.

Similarly, in a classification of the feature vector in the feature class K2, a second sub-algorithm T2 in the microcontroller μC be activated, which in turn activates the drive circuit FLIC for performing a soft-crash functionality, which in turn then activates a vehicle dynamics controller FDR1 in the form of a brake specification. Again, possibly a separate component for implementing the second sub-algorithm T2 and the functionality of the FLIC can be used again.

Accordingly, when classifying the feature vector into a third feature class, a third part algorithm, which is not explicitly shown in FIG. 2, can be activated, which then uses the drive circuit FLIC, here in the form of a control unit, to improve the vehicle dynamics, a second vehicle dynamics controller FDR2 activates wheel-selective braking or steering.

If the feature vector is divided by the classifier into a fourth feature class, a fourth sub-algorithm, which is also not shown in FIG. 2, can be activated. This fourth sub-algorithm can trigger via the drive circuit FLIC the second person protection PS2, such as a side airbag, initiate, so that the second personal protection PS2 accordingly squib or rev. Restraints on.

If the classification of the feature vector is carried out, for example, in a further feature class KN, an EPP algorithm (EPP = pedestrian protection algorithm) can be activated as a corresponding nth partial algorithm TN, by means of which a further personeπ protection means PS3 in the form of the activation circuit FLIC of detonators or rev. Restraints is controlled.

It is also possible that, contrary to the representation from FIG. 2, separate driver circuits FLIC1, FLIC2,... Are provided for the individual passenger protection means PS1, PS2 or PS3 or the individual vehicle dynamics controllers FDR1 or FDR2, as already briefly described above was cut and is shown in more detail with reference to Figure 4.

The representation from FIG. 2 can also be continued for a classification of the feature vector into any number of (but more than three) feature classes, in which case suitable safety means are activated by activating a correspondingly fitting partial algorithm and the drive circuit FLIC. The design of the classifier for the classification of the feature vector in at least three feature classes thus makes it possible to precisely activate those parts of a safety system of a vehicle that are needed precisely in the driving situation that has occurred, from the characteristics of one or more accident signals. An elaborate execution of all available algorithm parts of the security system or a permanent activation of all control circuits can thus be dispensed with.

FIG. 3 shows a third exemplary embodiment of the present invention as a block diagram illustration, with a single partial algorithm T of the partial algorithms T1 to TN shown in FIG. 2 being shown specifically to illustrate the mode of operation of the invention. However, the invention can also be used only using a single sub-algorithm T, so that not several sub-algorithms are needed. In the exemplary embodiment illustrated in FIG. 3, the classifier used is a multi-class support vector machine MSVM in the microcontroller .mu.C, which is acted upon by the features M1, M2 and M3. These features can be generated, for example, from an accident signal, as described above with reference to FIG. speed signals, yaw acceleration, vehicle acceleration, etc., or their integrals. The classifier MSVM can divide the features M1, M2 and M3 into a first, second or third feature class K1, K2 or K3 and feeds them to the sub-algorithm T, which activates a drive circuit FLIC1. In combination with the drive circuit FLIC1 and the sub-algorithm T, a triggering instruction is thus implemented numerically and / or by circuitry, by means of which, in response to the accident signal features M4 and M5, the personal protection means PS1, for example an airbag, is activated. The sub-algorithm T can be designed such that it implements a physics-based core threshold decision whose decision threshold value is influenced by the feature classes K1, K2 or K3. In this case, the triggering or activation of the personal protection means PS1 takes place in response to the accident signal features M4 and M5, which, however, may be identical to or derived from one or more of the input features M1 to M3.

The influencing of the decision threshold value can consist of a discount or a surcharge in accordance with a modification rule for the respective selected feature class K1 to K3. In this way, it is ensured that even with a possibly incorrect classification, the personal protection means PS1 is always activated by the partial algorithm T with the physically-based core threshold separation implemented therein (although not optimal but nevertheless).

As discussed above, after multi-class support vector machines implement a lerπ-based method, post-training classification can be based on a mathematical relationship such as the following equation

f (x) = Σy _/ - <V / c (xi, x) + ö.

/ = I

Here, the variables y _(r ctj and b results of the training or k (xι, x) are the used trained kernel function of the multiclass support vector machine The result of this classification function corresponds to the class determined in the classifier, where, for example, a real, ie non-binary, classification function value of 3.1 corresponds to the feature class K3, which comprises all the classification function values between 3.0 and 3.9 The sub-algorithm T in Figure 3 can then be replaced by a (binary) on / off _^ Activation of the signal path for the feature class K3 for activation, the transmission of the precisely determined classification function value of 3.1 to the partial algorithm T takes place, as a result, for example, of a quantitatively more exact increase or reduction of the decision threshold! value in the driving instruction of the sub-algorithm. The same also applies to the transmission of classification function values for the feature classes K1 and K2, it being possible, if appropriate, to amplify the respective classification function values by the amplifiers V1 to V3 in order to largely compensate or mask disturbances on the signal lines.

Figure 4 shows a fourth embodiment of the present invention in a block diagram representation. According to the exemplary embodiment illustrated in FIG. 4, a classifier in the form of a multi-class support vector machine MSVM is again provided in a microcontroller μC, to which features M1 to M3 are supplied by one or more accident signals. The features M1 to M3 are combined in the classifier MSVM (or via an upstream integrated interface) to a feature vector M and this classified into one of the feature classes K1 to K3. Each of these feature classes K1 to K3 serves to control a partial algorithm T1 to T3, which in each case again receives the characteristics M4 and M5 of an accident signal. According to the statements on the exemplary embodiment shown above in FIG. 3, the features M4 and / or M5 may again be identical to or derived from one or more of the input features M1 to M3 of the kiassifikator MSVM.

In the sub-algorithms T1 to T3, a triggering instruction in the form of a physical core threshold decision can be implemented in each case, wherein the classification of the feature vector into one of the feature classes K1 to K3 makes it possible to switch between different core thresholds in the different subalgorithms T1 to T3. For example, a first decision threshold value can be implemented in the first sub-algorithm T1, wherein the first sub-algorithm T1 is activated by the classification of the feature vector into the feature class K1. Upon activation of the first partial algorithm T1, a personal protection device PS1, such as an airbag, can be activated via the first control circuit FLIC1.

Furthermore, in the second sub-algorithm T2, a second decision threshold different from the first decision threshold can be implemented, the second sub-algorithm T2 being classified by the classification of the feature vector into the feature class K2 is activated. The second sub-algorithm also again uses the features M4 and M5 to trigger a security agent and also implements again a physical-based core threshold decision. If the feature vector is classified into the third feature class M3, the third sub-algorithm T3 can be activated, which implements a third physical-based core threshold decision with a further decision threshold using the features M4 and M5. In this case, the decision threshold value in the second or third sub-algorithm T2 and T3 can again be changed by evaluating a transmitted classification function value for the second and third feature classes K2 and K3.

The second sub-algorithm T2 and the third sub-algorithm T3 can activate a vehicle dynamics control FDR1, for example a control of an ESP function, via a common second drive circuit FL1C2. In this case, for example, a comparison of the two activation signals of the second and third sub-algorithms T2 and T3 for the first personal protection means PS1 according to a predefined specification makes it possible to check the activation of this protection means PS1. If, for example, the second decision threshold is lower than the third decision threshold, there must be an error if the second sub-algorithm T2 signals that a value from the considered features of the accident signal is below the second decision threshold but the third sub-algorithm T3 signals that the value from the considered features of the accident signal is above the third decision threshold.

The first and second drive circuits FLIC1 and FLIC2 may also be implemented together in a drive circuit, as shown in FIG. 2, respectively. The individual sub-algorithms can also be executed together in the microcontroller μC or on separate signal processing modules. Furthermore, when classifying the feature vector into the first feature class K1, the first decision threshold value can for example be loaded from a look-up table or a memory into the first sub-algorithm. Analogously, in the classification of the feature vector into the second feature class K2, the second decision threshold value can be loaded from a look-up table or a memory into the second subalgorithm and the third decision threshold value from the look-up when the feature vector is classified into the third feature class K3 Table or memory into the third sub-algorithm. Depending on the severity of the accident, which is characterized by the classification of the feature vector into a corresponding feature class then the vehicle dynamics control FDR1, for example an automatic brake controller, can be controlled to different degrees (for example in different stages) via the activation of the second or third sub-algorithm T2 and T3, as shown in FIG.

Thus, it is apparent that various embodiments of the invention are disclosed by the above description, which can be combined in any known manner for a person skilled in the art in order to realize the advantages of the invention as best as possible. For example, a combination of the embodiments of FIGS. 2 and 4 may be implemented. In the figure 1, such an embodiment is represented by the two dashed lines, in accordance with the embodiment of Figure 4, the microcontroller .mu.C several control signals, which are obtained from different subalgorithms, to a single drive circuit provides (dashed line between the microcontroller .mu.C and the first drive circuit FLIC1) or that according to Figure 2, a drive circuit activates a plurality of protection means (dashed line between the second drive circuit FLIC2 and the first personal protection means PS1). An application of a single partial algorithm with a plurality of classification signals according to the exemplary embodiment from FIG. 3 is not explicitly shown in FIG. 1; However, it is obvious to a person skilled in the art that also this further combination of the obvious embodiments is easy to implement.

Especially in the embodiments of the present invention illustrated in FIGS. 3 and 4, not only is a binary classification decision output, but a distinction is made between a plurality of feature classes. This is advantageous since it is desirable in the classification of an accident to be able to differentiate between different accident types and it is often not precise enough, at the output of a classifier or a drive circuit only a binary "Fire7" No Fire decision for Depending on the classification result, the components of a vehicle safety system required in the current driving situation can be specifically added, for example, based on the type of accident classified in more detail a core threshold in a drive unit to change the same so that the triggering requirements are met for the classified accident type .This substantially corresponds to the embodiment shown in Figure 3. However, it is conceivable au ch, that different look-up tables for the corresponding core threshold are loaded on the basis of the classification result, that is to say, in fact, switching back and forth between the core thresholds adapted to the respective accident types. This is essentially implemented in the Ausführungsbetspiel shown in Figure 4.

Furthermore, it is conceivable that special functions are added on the basis of the classification result. For example, if a pedestrian accident (i.e., an accident with no overlap against a non-deformable barrier) is classified, the so-called low-risk function may be started to cause the suppression of a second airbag step if necessary. For this purpose, it may therefore be necessary, for example, to distinguish between accident classes (ie feature classes) K1 = "ODB", K2 = Λ2T "and K3 =" FullFrontal ", the class" AZT "representing a" non-triggering crash test against a rigid barrier ".

In the manner proposed above, it is possible to ensure the best response of a vehicle safety system to a current driving situation extremely quickly and numerically or very simply from a plurality of accident signal feature combinations. As can be seen from the above-mentioned equation, this is not expensive to calculate numerically and thus does not represent a major challenge for modern Datenverarbeituπgskoneone, which is a great advantage for the implementation.

However, an important aspect is the training of the classifier used in the proposed invention. In contrast to the conventional support vector machine (SVM), which only differentiates between two classes (and for example in the accident discrimination between "Fire" and "No Fire"), the multi-class Suppαrt vector machine (MSVM) is in the Able to distinguish several, especially more than two, feature classes. The multi-class support vector Maschiπe is also a lerπbastertes method, being trained by pairwise predetermining Eingangsmerkmalsvektoreπ with the features to be trained accident signals and output signals in the form of the respective feature class to be assigned the classifier. This calculates in training the support vectors, which contain the most important data points of each class. The support vectors can be understood as the support vectors of a dividing line or separation surface separating each class. The remarkable thing about the multi-class support vector machine as well as the support vector machine is that by calculating the support vectors exactly that Dividing line is determined, which has the maximum distance to the different classes. This is particularly advantageous since, in the case of sensor signal fluctuations, this means the most robust separation of the classes. Another advantage is the fact that this optimal separation line is always found, which is not the case with other machine-based methods such as neural networks. The training takes place in a laboratory, wherein the found support vectors, for example, in a memory (such as an EEPROM of an airbag control unit in the form of a microprocessor) are stored. In doing so, the abovementioned variables of the listed equation can be determined in training, so that during the runtime of the algorithm before or in the event of the accident, the classification can carry out the classification of the feature vector with the aid of the (trained) simple equation reproduced above.

As a special feature of the training of a multi-class Supporf vector machine is to be noted that the training of such is ultimately always reduced to the two-class case, so that a distinction is made essentially between two different training variants. A first training variant ("One versus One") is based on successively training two classes against each other, in the case of 3 classes, first class 1 against class 2, then class 2 against class 3 and thereafter the class 3 is trained against class 1. The results of the classification are then combined, and a second training variant ("one versus rest") is based on successively training one class against all the remaining classes. Thus, in the case of three classes, class 1 is trained against classes 2 and 3, then class 2 against classes 1 and 3, and thereafter class 3 against classes 1 and 2. The classification results achieved are then also combined. Depending on the problem, the first training variant can be used once and the second training variant can be used the other time. In this way, the application time of additional functions for triggering a security agent can be significantly reduced by the automated calculation of the separating surface.

Figure 5 shows a fifth embodiment of the present invention. In this case, the invention is designed as a method 50 for controlling at least one safety device according to the procedure described above in the operation of such a classifier on the basis of the statistical learning theory. The method 50 has a first step 52 of detecting at least two features M1, M2 from at least one signal of an accident sensor system, in order to determine one of the detected features To form feature vector. In a second step 54, the feature vector formed is classified using a classifier on the basis of statistical theory in order to classify the feature vector into one of at least three possible feature classes K1, K2, K3. In a third method step 56, the security means FDR, PS1, PS2 are activated in accordance with a control specification for that feature class K1, K2, K3 into which the feature vector has been classified. By this method 50, the object of the invention can be achieved, wherein the illustrated advantageous effects set.

Depending on the circumstances, the method according to the invention can be implemented in hardware or in software. The implementation can be carried out on a digital storage medium, in particular a floppy disk, a CD or a DVD with electronically readable control signals, which can interact with a programmable computer system such that the corresponding method is carried out. In general, the invention thus also consists in a computer program product with a program code stored on a machine-readable carrier for carrying out the method according to the invention, when the computer program product runs on a computer. In other words, the invention can thus be realized as a computer program with a program code for carrying out the method when the computer program runs on a computer.

Claims

claims

A method (50) for controlling at least one safety means (FDR, PS1, PS2), the method (50) comprising the following steps:

- Detecting (52) of at least two features (M1, M2, M3) from at least one signal of a Uπfallsensorik (KS, BS1, DS, US, BS2) to form from the detected features (M1, M2, M3) a feature vector ;

- classifying (54) the formed feature vector using a classifier (MSVM) based on the statistical learning theory to classify the feature vector into one of at least three possible feature classes (K1, K2, K3); and

- Controlling (56) of the safety means (FDR, PS1, PS2) according to a control instruction (Tϊ, T2, TN, FLIC1, FLIC2, FLIC3) for that feature class (K1, K2, K3), in which the feature vector has been classified.

The method (50) of claim 1, characterized in that classifying (54) using the classifier (MSVM) based on the statistical learning theory comprises using a multi-class support vector machine.

Method (50) according to one of claims 1 or 2, characterized in that the driving (56) of the safety means (PS1, FDR1, FDR2, PS2, PS3, FDR1, PS1) is performed in accordance with a driving instruction (T1, TN, FLiC; T1, FLIC1) for a first feature class (K1, KN; K1) the activation of a personal protection device (PS1, PS2, PS3; PS1) and the activation of the safety device (PS1, FDR1, FDR2, PS2,

PS3) according to a driving instruction (T2, FLIC; T2, T3, FLIC2) for a second feature class (K2; K1) comprises activating a driving dynamics assistance control (FDR1).

4. Method (50) according to one of claims 1 to 3, characterized in that the driving (56) of the security means (FDR1, PS1) is further characterized by using at least one feature of the feature vector or another feature (M4, M5) Signal of the accident sensors (KS, BS1, DS, US, BS2) takes place.

5. Method (50) according to one of claims 1 to 4, characterized in that a classification function value (f (x)) is determined during classification (54) and the Driving (56) of the safety means (FDR, PS1, PS2) using the classification function value (f (x)) takes place.

6. The method (50) according to any one of claims 1 to 5, characterized in that the driving (56) of the security means (FDR1, PS1) is carried out according to a driving instruction (T1, T2, T3), which is based on a decision threshold.

7. Method (50) according to one of claims 1 to 6, characterized in that the activation instruction (T1, T2, T3) is changed in accordance with a modification rule dependent on the feature class (K1, K2, K3) in the step of the activation (56) ,

The method (50) according to claim 6, characterized in that in the step of driving (56), depending on the feature class (K1, K2, K3), an increase or decrease in the decision threshold or the decision threshold is replaced by a second decision threshold.

9. Method (50) according to one of claims 1 to 7, characterized in that the classification (64) takes place on the basis of class boundaries between the feature classes (K1, K2, K3) which are loaded from a memory.

10. Control device (SG) for controlling at least one Sicherheitsmittei (FDR, PS1, PS2), comprising the following features:

- At least one interface (IF1, IF2), which is adapted to at least two features (M1, M2, M3) from at least one signal of an accident sensor system (KS, BS1, DS ₁ US, BS2) form a feature vector;

an evaluation circuit (.mu.C), which is designed to classify the formed feature vector into one of at least three possible feature classes (K1, K2, K3) with the aid of a classifier (MSVM) on the basis of statistical learning theory; and

- A drive unit (T1, T2, TN, FLIC1, FLIC2, FLIC3), which is designed to control the safety means in accordance with a driving instruction for that feature class (K1, K2, K3), in which the feature vector has been classified.

A computer program executing all the steps of the method (50) of any one of claims 1 to 9 when running on a controller (SG).

12. Computer program product with program code, which is stored on a machine-readable carrier, for carrying out the method (50) according to one of claims 1 to 9, when the program is executed on a control unit (SG).