Method for operating a vehicle

Abstract

The invention relates to a method for operating a vehicle, wherein information (2) transmitted over a communication network (1) inside the vehicle and / or information (2) processed in the vehicle is examined for anomalies (3), and wherein a response is triggered as a result of at least one anomaly (3) being recognized. The method according to the invention is characterized in that - a computing unit (4) inside the vehicle, in order to identify an anomaly (3), checks whether at least one parameter (5) contained in the information (2) lies outside a defined permissible value range; and if so: - the computing unit (4) inside the vehicle replaces the value (6) of the parameter (5) with an artificially generated value (6*) that is within the permissible value range.

Description

[0001] Method for operating a vehicle

[0002] The invention relates to a method for operating a vehicle according to the type defined in the preamble of claim 1.

[0003] Many vehicle systems today are designed to be fail-safe. This means that if an error occurs, the respective system switches to a secure or emergency mode, whereby the vehicle system is still able to provide the respective vehicle function, but with a reduced or restricted scope. Depending on the severity of the error, it may also be necessary to shut down the vehicle function. A wide variety of causes are possible, such as a sensor failure, an error in data transmission via an internal vehicle communication network, e.g., a fieldbus, an internal error in a vehicle control unit, or the lack of data to be obtained from a backend external to the vehicle, and the like.For example, a temperature sensor for measuring the coolant temperature or the throttle valve control unit may fail, causing the vehicle's engine to go into emergency mode, which allows the vehicle to drive to the nearest workshop at a drastically reduced speed.

[0004] Especially in complex vehicle systems, it is desirable to be able to provide the respective vehicle function in a fail-safe manner in order to maintain vehicle operation. Since complex vehicle systems can have a particularly wide range of failure causes, the need to provide appropriate measures is particularly high.

[0005] A global automotive security system is known from EP 2 892 199 A1. The security system comprises a so-called watchman, which is integrated into the on-board electronics of a vehicle. This can be a dedicated processing unit, or the watchman can be implemented in the form of software on a processing unit already installed in the vehicle. The watchman's task is to analyze data transmitted via the vehicle's communications network and, depending on the respective characteristics of the data, to detect any abnormal behavior. This is intended to uncover cyberattacks on the vehicle. If the watchman detects abnormal behavior, countermeasures can be initiated, such as reporting the abnormal behavior to a central location or mitigating or containing the respective cause of the abnormal behavior.In response to abnormal behavior, the Watchman can adapt a state feature vector describing the state of the vehicle.

[0006] The present invention is based on the object of providing an improved method for operating a vehicle, the implementation of which enables the vehicle to be operated in a particularly fail-safe manner.

[0007] According to the invention, this object is achieved by a method for operating a vehicle having the features of claim 1. Advantageous embodiments and further developments emerge from the dependent claims.

[0008] A generic method for operating a vehicle, wherein information transmitted via an internal vehicle communication network and / or information processed in the vehicle is examined for anomalies, and wherein a reaction is triggered as a result of the detection of at least one anomaly, is further developed according to the invention in that

[0009] - an on-board anomaly detection unit checks whether at least one parameter included in the information lies outside a specified permissible range of values; and if so:

[0010] - the vehicle's internal processing unit replaces the value of the parameter with an artificially generated value within the permissible range.

[0011] Vehicle functions that can be provided by the vehicle require information as an input variable. If any of this information is faulty, the provision of the respective vehicle function is jeopardized. The method according to the invention thus provides for monitoring the respective information in the vehicle and detecting anomalies. Parameters that lie outside their permissible value range are identified and replaced with artificially generated values ​​within the permissible value range. This makes it possible to reliably ensure the correct functioning of the vehicle system processing the information. This improves the vehicle's reliability.

[0012] The vehicle can have one or more communication networks that are isolated from one another or communicatively connected to one another. Each communication network can be monitored individually. For example, each network can be a fieldbus such as a CAN bus, an Ethernet data line, or the like. There are various sources of information, such as sensors, control units, actuators, or even external sources, such as a central computing device connected to the vehicle via the Internet, e.g., in the form of a cloud server. In particular, the information transmitted and / or processed in the vehicle is available in a special data format and / or data structure. This data format can depend on the respective transport protocol of the communication network. In particular, the respective data or information is arranged or sorted in tabular form.

[0013] The information transmitted or processed in the vehicle comprises parameters. A parameter describes the respective type of information, for example the source of the information, the destination to which the information should be sent, the value of the respective parameter, and metadata such as a timestamp. A parameter can, for example, be a temperature value measured by a specific temperature sensor, a state of a vehicle control unit, a state of an actuator in the vehicle, and the like. Individual permissible value ranges can exist for each parameter. For a temperature sensor, for example, this permissible value range can be between -100° C and 500° C. Values ​​outside this value range are implausible.For this purpose, the vehicle's internal processing unit can store a corresponding database in which the respective permissible value ranges are stored for each parameter. If the processing unit detects an anomaly, i.e. at least one parameter that lies outside its permissible value range, the processing unit generates an artificial value for this parameter, which replaces the value lying outside the permissible value range. This artificially generated value can then be used for other functions in the vehicle. In general, there are various options as to how high the vehicle's internal processing unit sets the respective artificially generated value. For example, there can be various boundary conditions which specify how high the respective artificially generated value should be, depending on other information transmitted and / or processed in the vehicle.In the simplest case, fixed values ​​could also be defined, so that the same artificially generated value is always generated for a given parameter. Mixed forms are also possible, so that fixed values ​​can be artificially generated for some parameters and variable values ​​for others depending on other information.

[0014] An advantageous development of the method according to the invention provides that the computing unit considers the parameter to be outside its specified value range if the parameter assumes a parameter-specific error value. This makes it particularly clear to the computing unit that a respective anomaly exists. A wide variety of error values ​​can exist for the most diverse information sources or parameters. In particular, an error value is a particularly atypical value, such as 9999, 0, or -1. Such an error value can also be considered an error code. If, for example, a temperature sensor fails, the value of the corresponding parameter assumes -9999.

[0015] According to a further advantageous embodiment of the method according to the invention, an internal vehicle processing unit provides a function dependent on the parameter and uses the parameter with the artificially generated value to provide the function. This processing unit can be the same or a different processing unit than the processing unit that examines the information in the vehicle for anomalies and generates corresponding artificial values. In other words, artificial values ​​generated in the vehicle are read in as input data by vehicle systems to provide vehicle functions and processed like the respective parameters outside of a fault situation. This allows the respective vehicle functions to continue to be provided without restrictions.For example, if the temperature sensor for a coolant measurement fails, a value within the permissible range can be artificially generated, for example, at 90°C. Vehicle systems that process the parameter as an input, such as an engine control unit, can thus maintain normal operation. This prevents the engine control unit from switching the engine into limp-home mode.

[0016] A further advantageous embodiment of the method according to the invention further provides that the computing unit generates the value of the parameter artificially using artificial intelligence, in particular using a large language model trained for this purpose. In the simplest case, a fixed, artificially generated value is specified for each parameter. For example, if the temperature sensor in question fails, 90°C can always be specified as the artificial value. However, this entails a certain risk, since if the corresponding temperature sensor fails, the actual coolant temperature can no longer be determined. So, if there is actual damage and not just the sensor has failed, a dangerous coolant temperature can be reached. In this case, the corresponding emergency mode must be activated to avoid damage to the vehicle or vehicle components.

[0017] Thus, a further preferred embodiment of the method according to the invention further provides that the computing unit determines the level of the artificially generated parameter value depending on at least one further parameter included in the information. With reference to the aforementioned coolant case, the computing unit could, for example, record the vehicle speed, the oil temperature, the current number of engine revolutions per minute, the outside temperature, and the like, and take them into account to determine the artificially generated value. In this way, conditions can be determined from which the normally occurring level of the parameter value would result. This allows a value to be artificially generated for a parameter lying outside its permissible value range that would actually correspond to the real value without an error or would be close to it with a certain probability.For this purpose, generally fixed rules could be defined regarding how high a particular value should be set depending on other parameters. However, the use of artificial intelligence makes it particularly easy and reliable to identify corresponding patterns in the information transmitted and / or processed in the vehicle, even by analyzing a particularly large data set with a particularly large number of other parameters. Using artificial intelligence, it is therefore even more reliable to artificially generate particularly realistic parameter values. In an emergency, it could be estimated that the coolant temperature is likely excessively high, and the engine's emergency mode could be activated after all.

[0018] In particular, generative language models, such as large language models (LLMs), are used to generate artificial parameter values. Such a large language model allows the respective inputs for generating artificial parameter values ​​to be formulated in the form of natural language. The respective data structure can be adapted to the data format of the respective information. This ensures that corresponding values ​​can be reliably generated artificially for all common information transmitted and / or processed in the vehicle.

[0019] In a particularly preferred embodiment, the artificial intelligence, in particular in the form of the large language model, analyzes as much information currently transmitted and / or processed in the vehicle at the time the anomaly is detected as possible. This enables the large language model to more reliably shed light on the context of the respective error cause, i.e., the cause of the occurrence of the anomaly. In particular, the large language model can consider a respective parameter source, the target of the parameter, an operating state of the vehicle, an operating state of a vehicle function, a state of a vehicle component, for example, a sensor, a control unit, an actuator, or the like.

[0020] For example, a function can be executed in the vehicle to calculate a so-called "Eco-Score." Using the Eco-Score, the driver can be shown how sporty, sustainable, or predictive the vehicle is being driven. For example, the function call for the Eco-Score may fail. A request can then be sent to the large language model to artificially generate the Eco-Score. A corresponding request text to the large language model could contain, for example, "Driver is driving at 100 km / h on the highway, accelerator pedal position constant, brake pedal not activated, large distance to the vehicle in front. Output whether the driver is driving sporty, ecologically, or predictively and weight the three criteria." The large language model can provide the following response, for example in tabular form: sporty - less applicable; ecological - clearly applicable; predictive - probably applicable.

[0021] In addition, in-vehicle artificial intelligence, especially generative artificial intelligence, can generate additional data, such as an image. An input to the corresponding AI could be: "Draw a picture of a less sporty, significantly more eco-friendly, presumably forward-thinking driver!"

[0022] According to a further advantageous embodiment of the method according to the invention, the large language model is executed on the computing unit or on a vehicle-external computing device communicatively connected to the computing unit. The vehicle-external computing device can be, for example, a cloud server or server network. Particularly powerful hardware components can be installed on the vehicle-external computing device, allowing inputs to be processed by the large language model in a comparatively short period of time. This can reduce the latency for generating artificially generated values. It also allows the installation of less powerful hardware in the vehicle. However, if the vehicle-external computing device cannot be contacted, for example because the vehicle is located in a dead zone, there is a risk that no parameter values ​​can be generated artificially.Therefore, it may also be advantageous to run the large language model on a computing unit in the vehicle. However, this should preferably involve relatively powerful hardware components to be able to process corresponding inputs in a sufficiently short time. However, this increases the costs and energy requirements of the respective computing unit.

[0023] According to a further advantageous embodiment of the method according to the invention, the computing unit only executes the artificial intelligence when the computing unit detects an anomaly, in particular a pattern anomaly stored in an anomaly database accessible by the computing unit. Providing or executing artificial intelligence, in particular in the form of the large language model, requires a comparatively large amount of energy. Advantageously, the artificial intelligence is therefore not executed when it is not needed. This reduces the energy consumption of the computing unit and thus of the vehicle. Depending on the situation, it may not be possible to artificially generate corresponding parameter values, for example, because special boundary conditions exist or the artificial generation of parameter values ​​is not supported for this vehicle function.For example, there may be some safety-relevant vehicle functions that cannot be kept running artificially by artificially generating parameter values. To accommodate this, the respective parameters or vehicle functions and the associated anomalies can be stored in an anomaly database. This anomaly database can then contain pattern anomalies, which are anomalies for which it is permissible to artificially generate values ​​for the respective parameters. This means that upon detecting an anomaly, the computing unit first checks whether it is a pattern anomaly or not. If it is a pattern anomaly, the computing unit starts executing the artificial intelligence. If, on the other hand, it is not a pattern anomaly, the artificial intelligence is not executed.

[0024] A further advantageous embodiment of the method according to the invention further provides that the computing unit outputs an information message to a vehicle occupant when the computing unit replaces the value of the parameter with the artificially generated value. Generally, when the anomaly occurs, a malfunction or error has occurred in the vehicle. Accordingly, measures should be taken to determine the cause of the error and, if necessary, to rectify the error. A vehicle occupant, such as the person driving the vehicle or a person in a control center for autonomous vehicles, can be informed of the situation via the information message. The information message can simply contain information that a specific parameter has been artificially replaced.Advantageously, this can also include information about which vehicle functions depend on this parameter and can thus be made available accordingly. The information message can also contain a warning to alert the vehicle occupant to the potential danger associated with artificially replacing the parameter. The information message can also contain a recommendation as to how to proceed. For example, the vehicle could recommend that the person driving the vehicle visit a workshop to have the error rectified. The information message can be output via a haptic, acoustic and / or visual transmission path. For example, surfaces touched by the person driving the vehicle, such as the steering wheel or a vehicle seat, could vibrate.Warning tones could also be sounded in the vehicle, for example, by ringing a bell or emitting noises through loudspeakers. The information could also be conveyed acoustically through computer-generated speech. For visual output of the information message, warning lights could, for example, illuminate in the vehicle and / or text, possibly supplemented by pictograms, symbols, images, images generated by generative artificial intelligence, or the like, could be displayed on a display device, such as the head unit or the instrument cluster.

[0025] According to a further advantageous embodiment of the method according to the invention, when the computing unit replaces the value of the parameter with the artificially generated value, the vehicle transmits at least part of the information transmitted via the in-vehicle communications network and / or the information processed in the vehicle to a computing device external to the vehicle for analysis. As already mentioned, the computing device can be, for example, a cloud server operated by the vehicle manufacturer. However, it could also be a computer in a vehicle repair shop. This enables an assessment of the information describing the triggering of the anomaly by an external body, such as a service shop or the vehicle manufacturer itself.This not only allows for the detection of the cause of the error, but also for the verification and, if necessary, further development of the correct operation for the fail-safe provision of the respective vehicle functions—i.e., the correct implementation of the method according to the invention. For example, parameter values ​​artificially generated by the computing unit can be checked for their maximum plausibility. It is also possible to determine whether certain vehicle functions are kept running with an above-average frequency by artificially generating parameter values. This can reveal a general weakness in a particular vehicle function.

[0026] The invention further relates to a method for training a large language model that can be used in a method described above. The method for training the large language model according to the invention provides for the implementation of the following steps: - aggregating at least a subset of the information transmitted via a communication network of a training vehicle and / or the information processed in the training vehicle; and

[0027] - Feeding the information to a large language model as ground truth.

[0028] The large language model processes the corresponding input data and is therefore able to recognise relationships in the information which describe how the parameters contained in the information develop in relation to one another over time. In this way, application-specific relationships can be trained. For example, depending on the values ​​of the parameters, different values ​​or value ranges can change, a certain statistical distribution can form in the values, values ​​can influence one another, similarities can exist between values, and the like. This enables the large language model to estimate, for a specific parameter which lies outside its permissible value range, depending on the other parameters described by the information, which artificially generated value of the parameter which lies outside its permissible value range would be plausible. The values ​​transmitted or received in the respective training vehicleThe information processed can depend in particular on individual driving behavior, driving use, the vehicle environment, and the like. Such boundary conditions can also be captured as parameters during use of the training vehicle and fed into the large language model as an additional input variable.

[0029] The large language model can already be pre-trained to process tabular data. The large language model can preferably use machine learning methods such as transformers or attention-based methods.

[0030] A general language model can also be fine-tuned to the corresponding data structure of the training vehicle. This is also referred to as "fine-tuning."

[0031] An advantageous development of the method according to the invention for training the large language model further provides that a virtual vehicle running in a test environment, a real developer vehicle used in a test environment, and / or a fleet vehicle that is in the operational phase with regard to its service life is used as the training vehicle. Information generated in a vehicle, which originates from a simulated vehicle, a test vehicle, or a customer vehicle, can therefore be used to train the large language model. If the vehicle is a virtual vehicle, the test environment is a simulation running on a computer system. If the training vehicle is a real developer vehicle, the test environment can be an isolated section of the real world, such as a test track or race track, or even the real world itself.For example, developers can use the development vehicle to conduct test drives in real road traffic. In particular, unlike a fleet vehicle, the development vehicle is equipped with additional sensors and peripherals. In particular, the large language model is initially pre-trained based on information generated by a virtual vehicle and / or a development vehicle and then continuously retrained over its lifetime. For this retraining, the information generated by the vehicles in the fleet is preferably used. Additionally, information from virtual vehicles and / or development vehicles can also be used for retraining. In particular, the training of the large language model is repeated cyclically. For example, such a training iteration can occur when a certain batch size of information has been accumulated for training.

[0032] A further advantageous embodiment of the method according to the invention further provides that a large language model is trained individually based on:

[0033] - information generated by the virtual vehicle and / or the developer vehicle; and

[0034] - information generated by at least one fleet vehicle; wherein both large language models generate a value for a parameter lying outside its specified value range, the respective artificially generated values ​​are compared with one another, and at least a subset of the information generated in at least one of the vehicles is used for the further training of at least one of the large language models if the two artificially generated values ​​differ from one another by a specified amount. It is therefore possible to generate various differently trained large language models and to keep them available for use in the method according to the invention. In particular, a first large language model is trained on the basis of information generated by the vehicle manufacturer under controlled conditions, i.e. information generated by the virtual vehicle and / or the developer vehicle.Another large language model is trained using information generated by the vehicle manufacturer's fleet vehicles. These represent real-world use cases in an uncontrolled environment.

[0035] The performance of the two large language models can then be compared, particularly in a controlled environment, such as a developer environment. To do this, the two large language models are given the same task. For one and the same parameter that lies outside its permissible value range, the respective large language models are to determine an artificially generated value. This value is then compared. If the difference is greater than a certain amount, for example a squared distance, further steps can be initiated. This indicates that the boundary conditions specified under the controlled conditions deviate significantly from the real conditions encountered in the field.The large language model trained based on the information generated in the virtual vehicle and / or developer vehicle could also have a completely different structure than the large language model trained using information generated by fleet vehicles, which also needs to be analyzed. Based on the insights gained from this, a targeted, rapid, and data-efficient fine-tuning of the respective large language models can be performed. The information collected in the respective virtual vehicles or developer vehicles and / or the respective fleet vehicles can be used to further train at least one of the two, preferably both, large language models.

[0036] According to a further advantageous embodiment of the method according to the invention, at least two different large language models are trained on the basis of:

[0037] - a different subset and / or type of information; and / or

[0038] - depending on a source and / or a destination of the respective information and / or a respective vehicle function within the framework of which the respective information is generated and / or processed. This makes it possible to train tailor-made large language models for respective vehicle functions or use cases. Such tailor-made large language models are particularly suitable for delivering realistic results for the use cases underlying the respective training. For example, large language models can be created that are particularly good at recognizing statistical dependencies between different parameters. Other large language models can then, for example, be better able to recognize similarities between the values ​​of parameters. Other large language models are better suited to processing the information generated or processed in connection with a specific vehicle function.processed data, in particular with a temporal reference and the like.

[0039] A further advantageous embodiment of the method according to the invention further provides that the large language model is conditioned during training according to one of the following methods:

[0040] - feature name preconditioning;

[0041] - name value-pair preconditioning; and / or

[0042] - multiple name-value pair preconditioning.

[0043] With feature name preconditioning, only the respective feature name—in this case, an identifier for the respective parameter—is passed to the large language model, which is then able to artificially generate a suitable value for this parameter. For example, this could be the name, an ID, or another unique identifier of a specific sensor, control unit, user interface element, or the like.

[0044] Feature name-feature value combinations can also be specified, a process also known as name-value pair preconditioning. This makes it possible not only to generally generate artificial entries for a specific sensor or the like, but also to restrict the level of the respective artificially generated value to a certain range or even to a specific value. For example, special vehicle telemetry data can be artificially generated in which the engine speed assumes a certain predefined value. Furthermore, several feature name-feature value combinations can be specified as starting values ​​for generating parameter values. This is also known as multiple name-value pair preconditioning. A single output parameter can be generated depending on several input parameters, or several output parameters can be generated depending on one or more input parameters.Thus, artificially generated parameter values ​​can be made dependent on specific combinations of, for example, sensors, control units, user interface elements, and the like, and their respective current values. For example, telemetry data can be artificially generated in which the engine speed has a certain predefined value and the outside temperature measured by a certain temperature sensor has another specific predefined value. This allows suitable parameter values ​​to be artificially generated depending on the application. For the previously mentioned case of artificially replacing a coolant temperature, the following input values ​​could, for example, be specified for the large language model: vehicle speed 100 km / h; oil temperature 80°C; current engine revolutions per minute 1200 rpm; and: outside temperature 25°C.

[0045] Further advantageous embodiments of the method according to the invention for operating a vehicle and of the method according to the invention for training a large language model that can be used in such a method also emerge from the exemplary embodiment, which is described in more detail below with reference to the figure.

[0046] Figure 1 shows a schematic representation of a section of an in-vehicle communication network.

[0047] Figure 1 shows a section of an internal vehicle communication network 1. The communication network 1 can comprise several bus lines, for example, as shown, a first bus line 9.1 and a second bus line 9.2. The bus lines 9.1, 9.2 can be connected to one another via a so-called gateway 10, for example in the form of a switch, router, hub, or the like. The two bus lines 9.1, 9.2 can use the same or a different communication protocol. In particular, the data transmission rate in the respective bus lines 9.1, 9.2 can also differ. For example, it is a CAN bus and an Ethernet data line.

[0048] Various nodes are connected to the respective bus lines 9.1, 9.2, particularly in the form of sensors 11, control units 12, actuators 13, and the like. Also shown is a telecommunications unit 14, via which the respective vehicle can be communicatively connected to an external computing device 8. In particular, communication is wireless. For this purpose, the telecommunications unit 14 can connect the vehicle to the Internet, for example, via mobile radio, Wi-Fi, or the like.

[0049] Furthermore, an in-vehicle computing unit 4 is connected to the communications network 1. The computing unit 4 is tasked with examining information 2 communicated via the communications network 1 and / or information 2 processed in the vehicle for the existence of anomalies 3. This is illustrated in Figure 1 using a table. The information 2 comprises a plurality of parameters 5, to which respective values ​​6 are assigned. In particular, said parameters 5 can be arranged in a table. For example, this could be the value read or recorded by a specific sensor 11 at a specific time.

[0050] Unforeseen errors can occur during use of the vehicle. For example, a sensor 11 can fail. Other causes of errors can include, for example, an impairment of the on-board communication, an impairment of communication with the external computing device 8, an impairment or failure of a software function, for example, a program for providing a vehicle function, and the like. To detect such anomalies 3, a plurality of different so-called trigger functions are implemented in the computing unit 4. Such a trigger function can, for example, examine the value 6 of a respective parameter 5 for its permissible value range. If the respective value 6 of the parameter 5 falls outside this permissible value range, an anomaly 3 exists. In the exemplary embodiment shown in Figure 1, for example, the sensor provided for measuring the coolant temperature has failed.In this case, this sensor 11 can send an error value 7 as a signal via the communication network 1. This is detected accordingly by the computing unit 4, whereupon it replaces the value 6 of parameter 5 with an artificially generated value 6*.

[0051] This artificially generated value 6* is then used by the computing unit 4 or a respective control unit 12 to provide the vehicle function that would otherwise be restricted or terminated. For this purpose, the corresponding artificially generated value 6* can also be forwarded via the communication network 1. This can prevent, for example, the vehicle's engine from being switched into emergency mode due to a failure of the coolant temperature sensor.

[0052] According to a particularly advantageous embodiment of the method according to the invention, the computing unit 4 uses artificial intelligence, preferably a so-called large language model, to artificially generate values ​​6*. With the help of artificial intelligence, relationships between the respective pieces of information 2, i.e., the parameters 5, can be reliably and easily recognized, or with a certain probability, when anomalies 3 occur. Thus, respective relationships can be recognized in a situation-appropriate manner, and a suitable value for the value 6* can be determined for the respective situation. In particular, the temperature to be determined for the failed coolant temperature sensor can be selected depending on the current vehicle speed, oil temperature, engine revolutions per minute, and the outside temperature.If the respective values ​​are within the normal range, a typical coolant temperature such as between 80° C and 100° C, in particular 90° C, can be artificially specified.

[0053] If the computing unit 4 replaces a respective value 6 with a value 6*, a vehicle occupant can preferably be informed by issuing a corresponding message.

[0054] In addition, the computing unit 4 can only execute the artificial intelligence when it is actually required by detecting a corresponding anomaly 3. This can reduce the vehicle's energy consumption. Furthermore, the vehicle can transmit aggregated information 2 related to the detection of anomalies 3 to an external location, such as the aforementioned external computing device 8, for further analysis.

[0055] According to the invention, a method for training such a large language model is also specified. The large language model is trained using the information 2 generated during normal operation. This enables the large language model to recognize respective relationships between the individual parameters 5 of the information 2 and thus artificially generate particularly plausible values ​​6*. A virtual vehicle in a simulation environment, a real developer vehicle in a test environment, or even a fleet vehicle in use by a vehicle manufacturer can be used as the training vehicle. In particular, the large language model is continuously further trained. Central management of the large language model can be carried out by the vehicle-external computing device 8. Several large language models tailored to specific application cases can also be trained and used.These can, for example, differ depending on the task, so that a first large language model is preferably used to identify dependencies between parameters, while another large language model is used to generate special values ​​6* only for certain parameters 5 .

[0056] The 6* values ​​generated by different large language models can also be compared. This allows differences between a large language model trained on information generated in a controlled test environment and a large language model trained on real-world information generated in fleet vehicles to be identified.

[0057] A large language model can also be conditioned. Various conditioning methods are possible for this, such as feature name preconditioning, name-value-pair preconditioning, or multiple name-value-pair preconditioning.

[0058] With the help of the two methods according to the invention, the vehicle can be operated with exceptional reliability and reliability. The frequency and duration of downtimes of individual subsystems can be reduced. Since vehicle systems can thus be made available more frequently and for longer periods, road safety can also be increased. This also improves customer acceptance of technical solutions.

[0059] To provide the respective method steps, corresponding program code is stored in the individual computer systems of the vehicle, the execution of which by a processor allows the provision of the respective method steps. Thus, a computer-readable storage medium for storing the respective program code or computer program product is also part of the invention. Furthermore, the invention comprises a vehicle configured to carry out the method steps according to the invention.

Claims

Patent claims 1. A method for operating a vehicle, wherein information (2) transmitted via an internal vehicle communication network (1) and / or information (2) processed in the vehicle are examined for anomalies (3), and wherein a reaction is triggered as a result of the detection of at least one anomaly (3), characterized in that - an on-board computing unit (4) for detecting an anomaly (3) checks whether at least one parameter (5) included in the information (2) lies outside a specified permissible value range; and if so: - the vehicle-internal computing unit (4) replaces the value (6) of the parameter (5) with an artificially generated value (6*) within the permissible value range.

2. Method according to claim 1, characterized in that the computing unit (4) considers the parameter (5) to be outside its specified value range if the parameter (5) assumes a parameter-specific error value (7).

3. Method according to claim 1 or 2, characterized in that an in-vehicle computing unit (4) provides a function dependent on the parameter (5) and uses the parameter (5) with the artificially generated value (6*) to provide the function.

4. Method according to one of claims 1 to 3, characterized in that the computing unit (4) calculates the value (6*) of the parameter (5) artificially using artificial intelligence, in particular using a large language model trained for this purpose.

5. The method according to claim 4, characterized in that the large language model is executed on the computing unit (4) or on a vehicle-external computing device (8) communicatively connected to the computing unit (4).

6. Method according to one of claims 1 to 5, characterized in that the computing unit (4) determines the level of the value (6*) of the parameter (5) to be artificially generated as a function of at least one further parameter (5) included in the information (2).

7. Method according to one of claims 4 to 6, characterized in that the computing unit (4) executes the artificial intelligence only when the computing unit (4) detects an anomaly (3), in particular a pattern anomaly stored in an anomaly database accessible by the computing unit (4).

8. Method according to one of claims 1 to 7, characterized in that the computing unit (4) causes the output of an information message to a vehicle occupant when the computing unit (4) replaces the value (6) of the parameter (5) by the artificially generated value (6*).

9. Method according to one of claims 1 to 8, characterized in that when the computing unit (4) replaces the value (6) of the parameter (5) by the artificially generated value (6*), the vehicle transmits at least part of the information (2) transmitted via the vehicle-internal communication network (1) and / or the information (2) processed in the vehicle to a vehicle-external computing device (8) for analysis.

10. A method for training a large language model usable in a method according to any one of claims 4 to 9, characterized by the following steps: - aggregating at least a subset of the information (2) transmitted via a communication network (1) of a training vehicle and / or the information (2) processed in the training vehicle; and - Feeding the information (2) to a large language model as ground truth.

11. The method according to claim 10, characterized in that a virtual vehicle executed in a test environment, a real developer vehicle used in a test environment and / or a fleet vehicle in the operational phase with regard to its service life is used as the training vehicle.

12. Method according to claim 11, characterized in that a large language model is trained individually based on: - information generated by the virtual vehicle and / or by the developer vehicle (2); and - information generated by at least one fleet vehicle (2); wherein both large language models generate a value (6*) for a parameter (5) lying outside its specified value range, the respective artificially generated values ​​(6*) are compared with one another, and at least a subset of the information generated in at least one of the vehicles (2) is used for the further training of at least one of the large language models if the two artificially generated values ​​(6*) differ from one another by a specified amount.

13. Method according to one of claims 10 to 12, characterized in that at least two different large language models are trained on the basis of: - a different subset and / or type of information (2); and / or - depending on a source and / or a destination of the respective information (2) and / or a respective vehicle function within the framework of which the respective information (2) is generated and / or processed.

14. Method according to one of claims 10 to 13, characterized in that the large language model is conditioned during training according to one of the following methods: - Feature name preconditioning; - Name value-pair preconditioning; and / or - Multiple name-value pair preconditioning.

Images