METHOD FOR PREDICTING A VEHICLE-SPECIFIC PARAMETER, METHOD FOR CREATING A MODEL, COMPUTER PROGRAM PRODUCT AND SYSTEM

Description

[0001] The invention relates to a method for predicting a vehicle-specific parameter, a method for creating a model, a computer program product, and a system. Vehicles often have a whole series of sensors and / or communicate with sensors located in the vehicle's surroundings that contain vehicle-specific parameters.

[0002] Typically, it's not just the directly recorded sensor data that's of interest to the driver, other passengers, or the vehicle manufacturer. Collecting the data in context, i.e., linking the data, also provides opportunities to draw conclusions about parameters that would otherwise not be directly detectable.

[0003] As long as there is a causal relationship between the value detected by the sensors and the vehicle-specific value to be calculated, which is not directly measured, the latter can usually be determined directly from the measured value using a mathematical relationship. For example, the vehicle's speed could be calculated from a change in GPS coordinates.

[0004] It is more difficult when the relationship does not appear to be directly causal and the variable measured by sensors has only an indirect influence on the vehicle-specific parameter to be determined. For example, the movement of wheel dampers will depend on the vehicle's load and occupants, as well as the vehicle's surface. However, this relationship cannot be represented concisely.

[0005] However, there are already approaches in the prior art to calculate such only indirectly determinable variables using machine-learned models specifically adapted to the variable to be determined. However, it has been found that the training effort for each new variable to be determined is relatively high, and the accuracy of the prediction could be further improved. A relevant prior art document is EP 3 666 596 B1, which discloses a method for identifying occupant-specific attitudes using unsupervised clustering of sensor data.

[0006] It is an object of the present invention to at least partially remedy the above-mentioned disadvantages known from the prior art. In particular, it is an object of the present invention to provide a method for predicting a vehicle-specific parameter, a method for creating a model for predicting a vehicle-specific parameter, a computer program product, and a system that enables the prediction of even several different vehicle-specific parameters quickly, cost-effectively, and precisely.

[0007] The above object is achieved by a method having the features of claim 1, a method having the features of claim 10, a computer program product having the features of claim 13, and a system having the features of claim 14. Further features and details of the invention emerge from the respective subclaims, the description, and the drawings. Features and details described in connection with the method according to the invention for predicting a vehicle-specific parameter naturally also apply in connection with the method according to the invention for creating a model, the computer program product according to the invention, and / or the system according to the invention, and vice versa, so that reciprocal reference is or can always be made to the individual aspects of the invention with regard to the disclosure.

[0008] According to a first aspect of the invention, a method for predicting a vehicle-specific parameter, in particular an identity of a driving person, is provided, comprising: Acquiring sensor data by sensors designed to record vehicle-specific measured values, converting sensor data into a numerical representation, wherein the vehicle-specific measured values ​​are assigned to multiple clusters using an unsupervised clustering method, and the clusters are converted into a sequence of cluster IDs, the cluster IDs are coded, and a vocabulary is created using a tokenization algorithm, each vocabulary corresponding to a coded sequence, feeding the numerical representation to at least one trained prediction-specific model designed to output a prediction of the vehicle-specific parameter, where the trained prediction-specific model relies on a feature extraction network that was trained with stored sensor data in the numerical representation.

[0009] In the context of the invention, a vehicle-specific parameter is understood to mean any variable related to the vehicle. This does not necessarily require a direct causal connection. The identity of a person driving the vehicle depends on the vehicle itself only insofar as they use the vehicle in an individual manner, so that their identity can be inferred from this use. In principle, a vehicle can be understood to mean any land, water, or air vehicle. The vehicle can be designed as a motor vehicle, in particular as a passenger car.

[0010] In particular, it can be provided that the vehicle-specific parameter is a quantity that cannot be directly detected by a sensor, for example, the type of road surface. In this case, the method according to the invention offers the advantage that such a parameter can still be predicted.

[0011] It can also be provided that the vehicle-specific parameter can be directly measured, but measurement via a corresponding sensor is not desired for other reasons, for example, because the sensor is too expensive to purchase or operate, or the measurement is perceived as inconvenient (e.g., determining identity via a fingerprint sensor). In this case, too, the method according to the invention achieves the advantage that the measured variable can be recorded even without the corresponding sensor.

[0012] Sensor data can be understood as data recorded by a sensor that corresponds to the measurement purpose of the sensor. For example, a voltage recorded by a voltage sensor or an acceleration recorded by an acceleration sensor. Similar to vehicle-specific parameters, vehicle-specific measured values ​​only need to be indirectly related to the vehicle. For example, a sensor in a roadway can measure a vibration in the form of an acceleration caused by a vehicle. However, they can also be measured values ​​that are directly related to the vehicle, such as the vehicle's speed or its position.

[0013] It may be provided that none of the sensors is designed to directly detect the vehicle-specific parameter that is to be predicted by the method. While it is in principle possible to confirm directly detected vehicle-specific parameters using the method, a particular advantage is achieved if the vehicle-specific parameter is a quantity that cannot or should not be directly measured by a sensor.

[0014] The sensors within the meaning of the invention can, in principle, be any device suitable for recording vehicle-specific measured values. Sensors that can be used particularly advantageously are described below.

[0015] For the purposes of the invention, the conversion of sensor data into a numerical representation is understood to mean the clustering and tokenization described below. In other words, the numerical representation of the sensor data corresponds to a kind of language consisting of words and sentences that correspond to the acquired sensor data. However, this numerical representation is generally completely incomprehensible to a human reader. At most, a structure of words and sentences remains recognizable, depending on the character set used.

[0016] During the conversion to the numerical representation, the invention provides for the assignment of clusters. In other words, the sensor data is subjected to a cluster analysis. In a cluster analysis, a group of objects (here, the sensor data) is grouped in such a way that the objects in the same group (called a cluster) are (in a certain sense) more similar to each other than the objects in other groups (clusters). The clusters thus determined are converted into a sequence of cluster IDs.

[0017] Furthermore, the invention provides for encoding the cluster IDs. This means converting the data into a representation according to a specified encoding. Encoding is understood to mean the assignment of cluster IDs to graphic characters, in particular to written characters, so that these can be stored, transmitted, and converted using digital computers. However, it does not necessarily refer exclusively to written characters. In particular, any encoding can be used as long as it can be further processed by a computer. In particular, encodings with a large number of possible characters offer the advantage of enabling a compact representation, so that more sensor data can be processed, which in turn results in a more accurate prediction of the vehicle-specific parameter. It can be provided that the cluster IDs are encoded using Unicode, which comprises 144,697 characters.This enables a particularly compact representation that is also easily processed in most computer systems. Put simply, the encoding characters form the alphabet of a kind of "sensor data language."

[0018] Tokenization, as used in the method according to the invention, is to be understood as a process of delimiting, in particular classifying, sections of an input character string. The resulting tokens can then be passed on for further processing. If, as described above, a type of alphabet of the sensor data language was formed during encoding, then, in the same simplified image, individual words or vocabulary are now formed from the individual characters of the alphabet through tokenization. Just as the alphabet is structured in such a way that a single character has no perceivable connection to sensor data for a human in the rain, the same applies to the corresponding tokens or vocabulary. At most, it can be seen that a sequence of words can be formed into sentences, whereby the words do not have to have a conventional length compared to a human language.The same applies to the length and structure of the sentences.

[0019] A trained prediction-specific model is an algorithm designed to generate a prediction for one or more vehicle-specific parameters based on the input numerical representation. For example, if the trained prediction-specific model is given the sensor data corresponding to a four-minute drive in the numerical representation as input, the model can output a probability that the person driving is the vehicle owner, a registered user, or an unknown person.

[0020] To generate the prediction of the vehicle-specific parameter, the prediction-specific model relies on a feature extraction network. The feature extraction network, also called the backbone, is capable of converting the sensor data into a numerical representation through clustering and tokenization, as described above. Furthermore, the feature extraction network has trained the sensor data language using a predefined data set, as described in more detail below in connection with the further aspect of the invention.

[0021] Overall, the described method for predicting a vehicle-specific parameter offers the advantage that the compact form of the numerical representation allows the use of a language model in the form of the feature extraction network, which, through prior training, allows for the accurate prediction of a wide variety of vehicle-specific parameters. This is also due to the fact that the feature extraction network is not trained for a specific parameter, but rather learns the "language" of the sensor data itself. To predict a specific vehicle-specific parameter, only a prediction-specific model superimposed on the feature extraction network needs to be used. This achieves a particularly high degree of flexibility of the method.

[0022] Furthermore, it can advantageously be provided that at least one vehicle function or communication is triggered, wherein the triggering occurs based on the predicted vehicle-specific parameter. In other words, it can be provided that the vehicle acts accordingly in response to a prediction of the vehicle-specific parameter. The vehicle function can, for example, be implemented as a deactivation of the vehicle or a call-up of vehicle settings (e.g., sport mode, eco mode, autopilot). If, for example, it is determined that the person driving is neither the vehicle owner nor another person authorized to drive the vehicle, the vehicle can block inputs from the person driving and transfer the vehicle to a safe vehicle state (e.g., in the case of a motor vehicle, park the vehicle or, in the case of an aircraft, land the aircraft).Furthermore, it can be provided that communication is established with a vehicle manufacturer or a security authority (e.g., the police). This can, for example, be used to schedule an inspection appointment in the event of a predicted imminent system failure, or to allow the police to impound the vehicle in the event of unauthorized use. Overall, activating the vehicle function or communication can increase the safety and convenience of vehicle users.

[0023] Preferably, it can be provided that at least one of the sensors is designed to detect at least measured variables specific to the driving dynamics of the vehicle, in particular at least the applied torque, revolutions of the engine and / or the wheels, or the battery status, or measured variables non-specific to the driving dynamics of the vehicle, in particular at least one setting of a temperature control system, a temperature in the interior or in the vehicle environment, a seating position, or the state of an entertainment system. In other words, measured variables specific to the driving dynamics of the vehicle and / or measured variables non-specific to the driving dynamics of the vehicle can be detected by the sensors. By detecting variables that are non-specific and / or specific to the driving dynamics, the advantage is achieved that the accuracy of the predictions can be improved.In particular, the combination of both non-specific and specific measurement variables for driving dynamics has a beneficial effect on the accuracy of the predictions, especially when identifying a person driving.

[0024] Preferably, at least five sensors can be provided for acquiring the sensor data. In other words, a plurality of sensors, in particular at least five sensors, can acquire the sensor data. In principle, more sensors also result in greater prediction accuracy. However, with a larger data volume, the processing effort also increases. The teaching of the invention offers the advantage that even large data volumes can be processed effectively thanks to the numerical representation, and precise predictions can be made.

[0025] It is also conceivable that the sensor data is specific for at least a period of 4 minutes, in particular 8 minutes. In other words, the sensor data can be specific for a relatively long period, in particular 4 minutes or preferably even 8 minutes or more. While a larger data volume generally offers the advantage that more accurate predictions can generally be derived from it, a larger data volume is usually accompanied by a disproportionately large amount of analysis effort. However, the method according to the invention allows for data volumes that were previously unmanageable due to the numerical representation.

[0026] It is also conceivable that the unsupervised clustering method is implemented as a k-nearest neighbors algorithm. The k-nearest neighbors algorithm (KNN) takes the k nearest training examples in a dataset as input. Based on this, the output is a cluster membership (also called class membership). An object is classified by a plurality vote of its neighbors, whereby the object is assigned to the class that occurs most frequently among its k nearest neighbors (k is a positive integer, usually small). If k = 1, the object is simply assigned to the class of this one nearest neighbor. Since this algorithm relies on distance for classification, normalizing the training data can significantly improve accuracy when the features represent different physical entities or exist at very different scales.This is precisely what the method exploits, as the sensor data represents different physical units (e.g., speed and interior temperature) and can also be available at very different scales (engine revolutions may change rapidly, while a seat heating setting may not change at all over a period of minutes). Overall, the k-nearest neighbors algorithm improves the accuracy of the predictions, increasing the efficiency of the method.

[0027] Furthermore, it is conceivable that the tokenization algorithm is implemented as a byte-pair encoding algorithm. In other words, an algorithm can be used for tokenization in which the most common pair of consecutive data bytes is replaced by a byte that does not occur in that data. This achieves efficient compression of the representation.

[0028] Preferably, the prediction-specific model can be designed to determine at least the identity of a person driving or a road type traveled by the vehicle. Furthermore, further, in particular completely independent, prediction-specific models can be used. For example, the probability of component failure, electrical power or fuel consumption, or travel time on a route section can be predicted as vehicle-specific parameters using the method.

[0029] Furthermore, it is conceivable that at least the numerical representation of the sensor data has no physical meaning, or that the number of characters used to designate the cluster IDs is at least 1,000, in particular at least 10,000 characters. It can also be provided that the number of characters exceeds 100,000 characters. The more characters there are in an encoding, the more compactly the sensor data can be represented.

[0030] Furthermore, it is conceivable to provide methods for creating a model for predicting a vehicle-specific parameter, in particular an identity of a driving person, comprising: Assigning stored sensor data specific to at least one vehicle into a plurality of clusters using an unsupervised clustering method, wherein the clusters are converted into a sequence of cluster IDs, in particular by a computing unit, encoding the cluster IDs using a tokenization algorithm and creating a vocabulary, wherein each vocabulary corresponds to a coded sequence, in particular by a computing unit, and self-supervised learning of the coded sequences to obtain a trained feature extraction network, in particular by a computing unit, training a prediction-specific model with the stored sensor data, which model is designed to output a prediction of a vehicle-specific parameter, wherein the prediction-specific model relies on a feature extraction network which has been trained with stored sensor data in the numerical representation, in particular by a computing unit.

[0031] The process for creating the model is therefore similar to the process for predicting vehicle-specific parameters, whereby stored sensor data is used to create and train the feature extraction network, also known as the backbone. The prediction-specific model is also trained using this stored sensor data. The stored sensor data is previously acquired sensor data, which can, in particular, also include the vehicle-specific parameter to be predicted by the prediction-specific model, such as the identity of a driver. For example, the stored sensor data can be sensor data specific to a vehicle journey lasting 4 minutes, particularly 8 minutes. Furthermore, the sensor data can include data on the identity of a driver or the type of route traveled (highway, city, country road, etc.).).

[0032] The described method allows the creation of models for predicting vehicle-specific parameters that allow particularly accurate predictions and can also be flexibly adapted to the respective prediction task without the basic model (here the feature extraction network) having to be recreated each time.

[0033] Preferably, at least the self-supervised learning can comprise a language representation model, or the prediction-specific model can correspond to the feature extraction network with an additional output layer. In other words, the self-supervised learning can be carried out based on a method used in algorithms for (human) languages ​​and / or the feature extraction network can be supplemented with an additional output layer to generate the prediction-specific model.

[0034] Furthermore, it is conceivable that the self-supervised learning could include at least one BERT model, one ROBERT model, or one Distil-BERT model. BERT (Bidirectional Encoder Representations from Transformers) is a transformer-based machine learning method developed by Google for pre-training natural language processing. A BERT model offers the advantage that the data (or language) can be learned in two directions, and a single large text corpus, formed by the numerically represented stored sensor data set, is sufficient for learning. ROBERT and Distil-BERT models represent further developments of the BERT model, which have the advantage of enabling even more accurate predictions and / or consuming fewer resources.

[0035] According to a further aspect of the invention, a computer program product is provided, comprising instructions which, when the program is executed by a computing unit, cause the computing unit to execute at least the method according to the invention for predicting a vehicle-specific parameter or the method according to the invention for creating a model.

[0036] Thus, a computer program product according to the invention provides the same advantages as those already described in detail with reference to a method according to the invention. The method can, in particular, be a computer-implemented method.

[0037] According to a further aspect of the invention, a system is provided, comprising a computing unit which is designed to carry out at least the method according to the invention for predicting a vehicle-specific parameter or the method according to the invention for creating a model, wherein the computing unit is arranged at least in a vehicle, in a cloud computer which is in data communication with the vehicle, or in a traffic infrastructure which is in data communication with the vehicle.

[0038] Thus, a system according to the invention brings with it the same advantages as have already been described in detail with reference to a method according to the invention for predicting a vehicle-specific parameter, to a method according to the invention for creating a model, and / or to a computer program product according to the invention.

[0039] Further advantages, features, and details of the invention will become apparent from the following description, which describes embodiments of the invention in detail with reference to the drawings. The features mentioned in the claims and in the description may be essential to the invention individually or in any combination. They show schematically: Fig. 1: A simplified representation of the sequence of the method according to the invention for predicting a vehicle-specific parameter, Fig. 2: a simplified representation of the sequence of the method according to the invention for creating a model for predicting a vehicle-specific parameter, and Fig. 3: a side view of a vehicle, a cloud computer and a traffic infrastructure on which a computer program product can run on a computing unit for executing the method according to the invention for predicting a vehicle-specific parameter and / or the method for creating a model for predicting a vehicle-specific parameter.

[0040] In the following description of some embodiments of the invention, the same reference numerals are used for the same technical features even in different embodiments.

[0041] Figure 1shows a simplified representation of the sequence of the method according to the invention for predicting a vehicle-specific parameter 100. The method can be used in particular to predict the identity of a driving person 11. According to the invention, the method comprises the acquisition of sensor data 110 by sensors 20, which are designed to record vehicle-specific measured values.

[0042] In principle, the sensors 20 can be any type of sensor 20, as long as they record measured values ​​that are specific to the vehicle 10. Thus, they do not necessarily have to be located on the vehicle 10 itself. For example, a position or speed of the vehicle 10 can be recorded by a traffic infrastructure 50.

[0043] It can be provided that at least one of the sensors 20 is designed to detect at least specific measured variables for a driving dynamics of the vehicle 10, in particular at least the applied torque, revolutions of the engine and / or the wheels, or the battery status, or non-specific measured variables for a driving dynamics of the vehicle 10, in particular at least one setting of a temperature control system, a temperature in the interior or in the vehicle environment, a seating position, or the state of an entertainment system.

[0044] It may be provided that at least five sensors 20 are provided for acquiring the sensor data 110. The method according to the invention is particularly advantageous when a plurality of sensors 20 is used and can make a reliable prediction even if the sensor data 110 acquired by the sensors 20 has no direct causal relationship with the parameter 100 to be predicted.

[0045] Furthermore, it can be provided that the sensor data 110 is specific for at least a period of 4 minutes, in particular 8 minutes. In other words, a relatively large amount of data can be used for the method. This further improves the prediction quality. It is particularly noteworthy that the conversion of the sensor data 110 into the numerical representation 120, as described below, allows the algorithm to advantageously process the data in the large volume from multiple sensors 20 over several minutes.

[0046] Furthermore, the method according to the invention for predicting a vehicle-specific parameter 100 provides for converting the sensor data 110 into a numerical representation 120, wherein the vehicle-specific measured values ​​are assigned to multiple clusters 115 using an unsupervised clustering method, and the clusters 115 are converted into a sequence of cluster IDs. In other words, the conversion into a sequence of cluster IDs creates a string of values. This can be a sequence of numbers, letters, etc. Put simply, the sequence of cluster IDs corresponds to an "alphabet" of a "language" of the sensor data 110.

[0047] It can be provided that at least the numerical representation 120 of the sensor data 110 has no physical meaning, or that the number of characters used to designate the cluster IDs is at least 1,000, in particular at least 10,000 characters. This can be achieved, for example, by encoding in Unicode, where, in addition to numbers and letters, a multitude of other characters are available. Of course, the use of other encodings, in particular with a large selection of characters, is also conceivable. This allows for a particularly compact representation.

[0048] The unsupervised clustering method can be implemented as a k-nearest neighbors algorithm (also abbreviated as KKN). KNN is a type of classification in which the function is approximated locally and calculations are postponed until the function is evaluated. Since this algorithm relies on the distance between neighbors for classification, normalizing the training data can drastically improve accuracy when the features represent different physical units or are present at very different scales. This can be advantageously exploited in the described method, for example, when the measured value for the interior temperature of the vehicle 10 (which, for example, does not change or hardly changes within a few minutes) and the engine speed (which changes constantly and has no physical connection to the interior temperature of the vehicle 10) are linked.

[0049] Furthermore, the invention provides for the cluster IDs to be coded and a vocabulary to be created using a tokenization algorithm, with each vocabulary word corresponding to a coded sequence 130. Put simply, the "alphabet" of the "language" of the sensor data 110 obtained from the sequence of cluster IDs is then used to create a "dictionary" with "vocabularies" in which the sensor data 120 can be represented.

[0050] The tokenization algorithm can be implemented as a byte-pair encoding algorithm. This byte-pair encoding algorithm is a form of data compression in which the most frequent pair of consecutive data bytes is replaced by a byte that does not occur in this data. A table of replacements is required to restore the original data. This allows for particularly efficient compression and enables a representation of the sensor data 110 that can be well processed by a prediction-specific model 140, which is described below.

[0051] According to the invention, it is further provided to supply the numerical representation 120 to at least one trained prediction-specific model 140, which is designed to output a prediction of a vehicle-specific parameter 100. The trained prediction-specific model 140 uses a feature extraction network 150 that was trained with stored sensor data 160 in the numerical representation 120. In other words, a prediction model 140 adapted to a specific prediction task is used, which is based on a feature extraction network 150 that, in turn, masters the "language" of the sensor data 110.

[0052] For example, it may be provided that the prediction-specific model 140 is designed to determine at least the identity of a driving person 11 or a road type 40 traveled by the vehicle 10. For this purpose, the prediction-specific model 140 can be trained accordingly to specify the identity of a driving person 11 in the form of a probability from a series of tokenized cluster IDs in sensor data 110.

[0053] The particular advantage of the method according to the invention also lies in the fact that the respective different prediction-specific models 140 can be adapted to a whole range of conceivable prediction tasks. In addition to identifying a person 11 driving, for example, the road type 40 traveled by the vehicle 10 could be predicted. However, ultimately, almost any vehicle-specific parameter 100 can be predicted by a corresponding prediction-specific model 140.

[0054] Furthermore, the triggering of at least one vehicle function or communication can be provided, wherein the triggering occurs based on the predicted vehicle-specific parameter 100. For example, the vehicle can be switched off or deactivated if the identified person 11 driving does not correspond to any of the persons 11 registered as such for the vehicle 10 in a database. It can also be provided that the vehicle 10 is reported as stolen in this situation. Furthermore, the road type 40 on which the vehicle 10 is traveling can be communicated, for example, to an insurance company or a map manufacturer.

[0055] According to a second aspect of the invention, which is primarily Figure 2As shown, a method is provided for creating a model 140 for predicting a vehicle-specific parameter 100, in particular an identity of a driving person 11.

[0056] The method for creating a model 140 includes mapping stored sensor data 160 specific to at least one vehicle 10 into multiple clusters 115 using an unsupervised clustering method, converting the clusters 115 into a sequence of cluster IDs, encoding the cluster IDs using a tokenization algorithm and creating a vocabulary, each vocabulary corresponding to an encoded sequence 130, and self-supervised learning of the encoded sequences 130 to obtain a trained feature extraction network 150.

[0057] Furthermore, the training of a prediction-specific model 140 with the stored sensor data 160 is provided, which is designed to output a prediction of a vehicle-specific parameter 100, wherein the prediction-specific model 140 uses a feature extraction network 150 which was trained with stored sensor data 160 in the numerical representation 120.

[0058] Furthermore, it can be provided that at least the self-supervised learning comprises a language representation model, or the prediction-specific model 140 corresponds to the feature extraction network 150 with a further output layer.

[0059] Self-supervised learning can include at least a BERT model (Bidirectional Encoder Representations from Transformers), a ROBERT model, or Distil-BERT model.

[0060] During self-supervised learning, a portion 135 of the sequence may be masked out. The model then learns to predict the correct value for the masked sequence 135. The self-learning process may be repeated until a predetermined value for the reliability of predicting the correct value for the masked sequence 135 is reached.

[0061] Figure 3Finally, by way of example, shows an environment in which the methods according to the invention, the computer program product and / or the system can be used. In a vehicle 10, a computing unit 30 is shown here, on which the method for predicting a vehicle-specific parameter 100 and / or the method for creating a model can run. Likewise, the aforementioned methods could run on the traffic infrastructure 50 or a cloud computer 60. The sensors 20 can be attached both to the vehicle 10, as shown, or, for example, to the traffic infrastructure 50. By using the method for predicting a vehicle-specific parameter 100, for example, the identity of a person driving 11 or the type of road 40 being traveled on can be predicted.However, due to the flexibility of the feature extraction network 150, numerous other vehicle-specific parameters 100 can be determined by the method if the prediction-specific models 140 have been trained accordingly. List of reference symbols

[0062] 10Vehicle 11Person 20Sensors 30Computing unit 40Road type 50Traffic infrastructure 60Cloud computer 100 Parameters 110 Sensor data 115 Cluster 120 Representation 130 Sequence 135 Hidden sequence 140 Model 150 Feature extraction network 160 Stored sensor data

Claims

1. Method for predicting a vehicle-specific parameter (100), in particular an identity of a driving person (11), comprising: - acquiring sensor data (110) by means of sensors (20) which are designed to record vehicle-specific measured values, - converting sensor data (110) into a numerical representation (120), wherein i. the vehicle-specific measured values are assigned to a plurality of clusters (115) using an unsupervised clustering method, and the clusters (115) are converted into a series of cluster IDs, ii. the cluster IDs are encoded and a vocabulary is created by a tokenization algorithm, wherein each word corresponds to an encoded sequence (130), - feeding the numerical representation (120) to at least one trained prediction-specific model (140) which is designed to output a prediction of the vehicle-specific parameter (100), wherein the trained prediction-specific model (140) accesses a feature extraction network (150) which was trained with stored sensor data (160) in the numerical representation (120).

2. Method according to claim 1, characterized in that the triggering of at least one vehicle function or communication is also provided, the triggering taking place on the basis of the predicted vehicle-specific parameter (100).

3. Method according to claim 1 or 2, characterized in that at least one of the sensors (20) is designed to acquire specific measured variables at least for driving dynamics of the vehicle (10), in particular at least the applied torque, revolutions of the engine and / or the wheels, or the battery status, or to acquire non-specific measured variables for driving dynamics of the vehicle (10), in particular at least setting of a temperature control system, a temperature in the interior or in the vehicle surroundings, a seating position, or the state of an entertainment system.

4. Method according to any of the preceding claims, characterized in that at least five sensors (20) are provided for acquiring the sensor data (110).

5. Method according to any of the preceding claims, characterized in that the sensor data (110) are specific for at least a period of 4 minutes, in particular 8 minutes.

6. Method according to any of the preceding claims, characterized in that the unsupervised clustering method is implemented as a k-nearest neighbors algorithm (KKN).

7. Method according to any of the preceding claims, characterized in that the tokenization algorithm is implemented as a byte-pair encoding algorithm.

8. Method according to any of the preceding claims, characterized in that the prediction-specific model (140) is designed to determine at least the identity of a driving person (11) or a road type (40) traveled on by the vehicle (10).

9. Method according to any of the preceding claims, characterized in that at least the numerical representation (120) of the sensor data (110) has no physical meaning, or in that the number of characters used to identify the cluster IDs is at least 1,000, in particular at least 10,000 characters.

10. Method for creating a model for predicting a vehicle-specific parameter (100), in particular an identity of a driving person (11), comprising: - assigning stored sensor data (160) specific to at least one vehicle (10) into a plurality of clusters (115) using an unsupervised clustering method, wherein the clusters (115) are converted into a series of cluster IDs, - encoding the cluster IDs by means of a tokenization algorithm and creating a vocabulary, wherein each word corresponds to an encoded sequence (130), and - learning the encoded sequences (130) in a self-supervised manner to obtain a trained feature extraction network (150), - training a prediction-specific model (140) with the stored sensor data (160), which model is designed to output a prediction of a vehicle-specific parameter (100), wherein the prediction-specific model (140) accesses a feature extraction network (150) which has been trained with stored sensor data (160) in the numerical representation (120).

11. Method according to claim 10, characterized in that at least the self-supervised learning comprises a language representation model, or the prediction-specific model (140) corresponds to the feature extraction network (150) with a further output layer.

12. Method according to either claim 10 or 11, characterized in that the self-supervised learning comprises at least one BERT model (Bidirectional Encoder Representations from Transformers), ROBERT model or DistilBERT model.

13. Computer program product comprising instructions which, when the program is executed by a computing unit (30), cause the computing unit to execute at least the method according to any of claims 1 to 9 or the method according to any of claims 10 to 12.

14. System comprising a computing unit (30) which is designed to execute at least the method according to any of claims 1 to 9 or the method according to any of claims 10 to 12, wherein the computing unit (30) is arranged at least in a vehicle (10), in a cloud computer (60) which is in data communication with the vehicle (10), or in a traffic infrastructure (50) which is in data communication with the vehicle (10).