Method and device for providing a trained machine learning model for detecting a hands-on / hands-off state of a steering wheel of a vehicle

Abstract

The invention relates to a method for providing a trained machine learning model for detecting a hands-on / hands-of state (40) of a steering wheel of a vehicle (50), having the steps of: training at least one general teacher model (20) using a training data set (10); training at least one teacher model (21-x) which is specialized with respect to a context (11-x), wherein for this purpose, training data (10-x) for the context (11-x) is selected from the training data set (10) and / or the training data set (10) is supplemented with training data (10-x) for the context (11); training a student model (22) using the general teacher model (20) and / or the at least one specialized teacher model (21-x) in the form of a teacher-student training process, the process of selecting which teacher model(s) (20, 21-x) is / are used to train the student model (22) being made on the basis of the context (11-x) corresponding to the respective piece of training data (10-x); and providing the trained student model (22). The invention additionally relates to a device (1) for providing a trained machine learning model for detecting a hands-on / hands-off state (40) of a steering wheel of a vehicle (50).

Description

[0001] Description

[0002] Method and apparatus for providing a trained machine learning model for detecting a hands-on / hands-off state on a steering wheel of a vehicle

[0003] The invention relates to a method and a device for providing a trained machine learning model for detecting a hands-on / hands-off state on a steering wheel of a vehicle. Furthermore, the invention relates to a method and a device for detecting a hands-on / hands-off state on a steering wheel of a vehicle.

[0004] Sensors such as a capacitive steering wheel are used in vehicles to monitor driver activity. Such a steering wheel detects when the driver has touched (“hands-on”) or not touched (“hands-off”) the steering wheel using a capacitive sensor. The result is transmitted to the functions being used, such as a longitudinal and / or lateral guidance assistance system. The contact of the hands on the steering wheel can be used to determine driver activity and the driver’s level of attention. For example, the system can remind the driver to put their hands on the steering wheel if it is detected that their hands have not been on the steering wheel for a specified time during lateral guidance.

[0005] To save additional costs for a capacitive sensor in the steering wheel, it is known to monitor driver activity using machine learning models, in particular artificial neural networks, based on a torque (hand torque) detected on the steering wheel. Such a method is known, for example, from DE 102019211 016 A1.

[0006] A major challenge of steering torque-based detection is identifying the driver-induced steering torque in the measured (noisy) steering torque. Many factors can lead to noisy steering torque, in particular the position of the sensor (this usually forms part of the steering gear or steering assistance, which, via the elasticities of the steering column in conjunction with the steering wheel, creates a system capable of torsional vibration, whose inherent dynamics make precise measurement of the driver-induced torque difficult); the strength of friction in the steering system; re-excitation from the road due to unevenness; the dead weight of the steering wheel (I of the steering system); and vibration of the steering wheel due to an assistance function (e.g., due to haptic feedback when leaving the lane).

[0007] In addition, the characteristics (used for hands-off detection) of the measured steering torque can change due to external influences, e.g. temperature, vehicle load, the presence of a trailer, tire type and / or tire condition, steering system changes over the lifetime, road gradient / inclination / slant, etc.

[0008] Due to the number of influences, it is necessary to equip the machine learning model, especially the neural network, with enough degrees of freedom, i.e., weights, to allow the model to represent all influences. However, this results in a large model and thus requires very large memory and computing resources. In particular, the relationship between the number of weights and the quality of representation of the desired objective function is not linear, but increases almost exponentially.

[0009] The use of a teacher-student training is, for example, from the

[0010] DE 11 2020 001 663 T5. Furthermore, the use of teacher-student training is known from G. Hinton et al., Distilling the Knowledge in a Neural Network, arXiv: 1503.02531 v1 [stat.ML], March 9, 2015.

[0011] The invention is based on the object of improving a method and a device for providing a trained machine learning model for detecting a hands-on / hands-off state on a steering wheel of a vehicle. Furthermore, the invention is based on the object of improving a method and a device for detecting a hands-on / hands-off state on a steering wheel of a vehicle.

[0012] The object is achieved according to the invention by a method having the features of patent claim 1 and a device having the features of patent claim 9. Furthermore, the object is achieved by a method having the features of patent claim 8 and a device having the features of patent claim 10. Advantageous embodiments of the invention emerge from the subclaims.

[0013] In particular, in a first aspect, a method is provided for providing a trained machine learning model for recognizing a hands-on / hands-off state on a steering wheel of a vehicle, comprising training at least one general teacher model using a training data set, training at least one teacher model specialized with regard to a context, wherein for this purpose training data of the context is selected from the training data set and / or the training data set is enriched with training data of the context, training a student model using the general teacher model and / or the at least one specialized teacher model by means of teacher-student training, wherein a selection of which of the teacher models is or are used to train the student model is made based on a context corresponding to the respective training date, and providing the trained student model.

[0014] Furthermore, in particular in a second aspect, a device is provided for providing a trained machine learning model for recognizing a hands-on / hands-off state on a steering wheel of a vehicle, comprising a data processing device, wherein the data processing device is configured to train at least one general teacher model using a training data set, to train at least one teacher model specialized with regard to a context, and for this purpose to select training data of the context from the training data set and / or to enrich the training data set with training data of the context, to train a student model using the general teacher model and / or the at least one specialized teacher model by means of teacher-student training, and a selection of which of the teacher models is or are used to train the student model,based on a context corresponding to the respective training date and to provide the trained student model.

[0015] Furthermore, in particular in a third aspect, a method for detecting a hands-on / hands-off state on a steering wheel of a vehicle is provided, comprising detecting at least one steering variable on the steering wheel of the vehicle, using a student model trained according to the embodiments of the first and / or second aspect to detect the hands-on / hands-off state, wherein the detected at least one steering variable is supplied to an input of the trained student model for this purpose, and providing the hands-on / hands-off state estimated by the trained student model.

[0016] Finally, in a fourth aspect, a device for detecting a hands-on / hands-off state on a steering wheel of a vehicle is provided, comprising at least one sensor for detecting at least one steering variable on a steering wheel, a control unit, wherein the control unit is configured to receive the detected at least one steering variable, to provide and use a student model trained according to the embodiments of the first and / or second aspect to detect the hands-on / hands-off state, and to supply the detected at least one steering variable to an input of the trained student model for this purpose, and to provide the hands-on / hands-off state estimated by the trained student model.

[0017] The method and device according to the first and second aspects make it possible to provide a model that is improved in terms of memory and computing resource consumption. For this purpose, a student model is provided which, in terms of structural design, is particularly simpler or less complex than the teacher models used. For example, the student model can have fewer neurons or fewer inner layers and thus fewer parameters (in particular weights) than the respective teacher models. To prevent a reduction in functional quality, it is provided that, in addition to a general teacher model, at least one teacher model specialized for a context is trained.During teacher-student training, the training of a student model is then provided using the general teacher model and / or the at least one specialized teacher model, with the selection of which of the teacher models is used to train the student model being made based on a context corresponding to the respective training date. The student model trained in this way is provided, in particular output. The trained student model is used, in particular, in a vehicle to detect a hands-on / hands-off state. Based on this, an assistance system can be operated, for example.

[0018] A particular advantage of the method and device is that a student model with improved functional quality can be provided. This allows a machine learning model with reduced memory and computing resource requirements to be provided, particularly with improved functional quality.

[0019] The models (the teacher models and the student model) are or will be trained in particular to estimate a hands-on / hands-off state based on at least one steering variable.

[0020] The at least one general teacher model, the at least one specialized teacher model, and the student model are designed, in particular, as artificial neural networks, in particular as deep neural networks, which in particular comprise several inner layers. The models are, in particular, artificial recurrent neural networks that, at each time point t, process the input data X tprocess and at least a hands-off probability y t output in [0,1]: y t = p(x t | XO :M). Preferably, the neural networks output a probability for a hands-on state and a hands-off state at each time t. In particular, the neural networks have a so-called memory h in which information from previous magazines is stored and which can be used for the output for the current magazine. The output is further processed, for example, by filtering before the decreasing functions (e.g. a lateral guidance assistant) process it. In particular, it can be provided that, based on a comparison of the hands-off probability with a predetermined threshold value, a binary hands-off signal is provided (with the two states "hands-off detected" and "hands-off not detected"). The same can be done for the hands-on state, or a joint signal can be provided with the two signal states "hands-on" or "hands-off".

[0021] During training (i.e. during the respective training phase), the models are trained using training data comprising pairs in which data of the steering variable, in particular torque data, are each paired with a hands-on / hands-off state as ground truth. The data of the at least one steering variable, in particular the torque data, are in particular time series of the at least one steering variable detected at the steering wheel, in particular time series of torques detected at the steering wheel. The training data is obtained in particular with the aid of test drives and / or in simulators. In principle, the provision of training data can be carried out in particular according to the method described in DE 102019211 016 A1. Training is carried out in a manner known per se, in particular by means of supervised learning. The detection and / or collection of the training data of the training data set can be part of the first and second aspects.The hands-on / hands-off state is detected in particular by means of additional sensors (e.g. a capacitive steering wheel).

[0022] During training, the loss function calculates an error per weight as part of the backpropagation and thus optimizes the error. In particular, binary cross entropy can be used to detect the hands-on / hands-off state:

[0023] L = -y * log(Y) + (1-y)*log(1-Y), where Y is the ground truth and y is the output of the model.

[0024] When training the student model, a customized loss function is used:

[0025] L = A * student loss + (1-A) * distillation loss, where student loss is the binary cross-entropy between the output of the student model and the ground truth, and distillation loss is the binary cross-entropy between the output of the student model and the teacher model. The weighting factor A indicates how strong the influence of the teacher model is compared to the influence of the original training data (i.e., the ground truth). The weighting factor A is determined, for example, using empirical test series.

[0026] During teacher-student training (teacher-student learning), the student model is trained, in particular, with the help of at least one general teacher model and / or at least one specialized teacher model. The same training data is fed to both the teacher and student models as input data. In addition to a ground truth, an output from the teacher model or a combination of teacher models is used to adjust parameters (particularly weights) of the student model and thus train the student model. These two variables are taken into account using the adjusted loss function described above. When combining teacher models, the outputs of the teacher models are considered in a weighted manner.

[0027] The method for providing a trained machine learning model for detecting a hands-on / hands-off state on a steering wheel of a vehicle can in particular be carried out as a computer-implemented method.

[0028] A steering variable is, in particular, a variable that represents and / or describes a current state of the steering wheel. A steering variable is, in particular, a torque, which is detected in particular by means of a torque sensor on the steering wheel. In principle, however, a steering variable can also be another variable detected directly or indirectly on the steering wheel. For example, it can be provided to detect a current at an electrical machine on the steering wheel and use it as a steering variable. The hands-on / hands-off state can be detected exclusively on the basis of the steering variable detected on the steering wheel, in particular a detected torque. However, it is also possible, in particular, for the machine learning model to be provided with further (steering) variables that are detected on the steering wheel (e.g., a steering wheel angle and / or a steering wheel angular velocity, etc.).) and the trained machine learning model recognizes the hands-on / hands-off state taking this additional variable into account. Furthermore, variables that are not recorded at the steering wheel can also be taken into account, such as vehicle speed, lateral acceleration, yaw rate, wheel ticks, damper information, and / or other driving dynamics variables, etc. In particular, however, no capacitive sensor is provided on the steering wheel.

[0029] A hands-off state is, in particular, a state in which the driver does not touch the steering wheel. In particular, none of the driver's fingers are in contact with the steering wheel. Detecting the hands-off state can, in particular, comprise providing a hands-off state signal. This comprises, for example, a hands-off probability or coded signals for the states "hands-off detected" and "hands-off not detected." A hands-on state is, in particular, a state in which the driver touches the steering wheel. In particular, a hands-on / hands-off state is provided, for example as a hands-on / hands-off state signal with, in particular, two signal states (e.g., "hands-on detected" or "hands-off detected").

[0030] A context refers to or includes, in particular, properties of a situation in which the hands-off state is to be detected and / or in which one or more values ​​of the at least one steering variable were recorded. Examples of properties that can determine a context are: an outside temperature (e.g., specialized teacher models can be provided for extremely hot or extremely cold temperatures), an inside temperature, a steering wheel vibration, a load and / or a weight of the vehicle, the presence of a trailer, the presence of snow chains, cobblestones, potholes, maximum steering interventions (e.g., steering vibration), speed bumps, properties (e.g., identity, gender, age, weight) of the driver, etc. The current context is detected and / or determined, in particular, based on recorded sensor data.For example, when using the trained student model, it may be provided to query such sensor data via a vehicle's CAN bus and / or to obtain it from sensors and / or a vehicle control system. A subdivision of the contexts and the associated provision of specialized teacher models for each can be based, for example, on expert knowledge. In particular, the need for specialized teacher models for different contexts or aspects of a context can be determined based on expert knowledge.

[0031] In principle, however, automated classification procedures can also be used for this purpose.

[0032] Parts of the devices, in particular the data processing devices, can be implemented individually or collectively as a combination of hardware and software, for example, as program code executed on a microcontroller or microprocessor. However, it can also be provided that parts are implemented individually or collectively as an application-specific integrated circuit (ASIC) and / or a field-programmable gate array (FPGA).

[0033] In one embodiment, the trained student model is loaded into a memory of at least one control unit of a vehicle after training as part of the provision process. This allows the trained student model to subsequently be used in the vehicle to provide hands-on / hands-off detection. Loading into the memory includes, in particular, loading a structural description and parameters (weights, hyperparameters, etc.) of the student model into the memory. In particular, an assistance system of the vehicle can be operated based on a detected hands-on / hands-off state.

[0034] In one embodiment, the at least one specialized teacher model is initialized at the beginning of training using parameters of the trained general teacher model. This allows the training time of the at least one specialized teacher model to be reduced, since it is already pre-trained due to the initialization.

[0035] In one embodiment, it is provided that the selection is made based on metadata linked to the respective training data item. In this way, the respective context in which the training data item is located can be directly determined based on the metadata linked to the respective training data item. Such metadata can, for example, directly describe the context, e.g. with reference to the circumstances under which the training data item, in particular the at least one control variable, was acquired. Examples of properties of the context have already been described above. In one embodiment, it is provided that the selection is made at least partially by means of a decision tree. In this way, one or more teacher models can be directly selected depending on properties of the context.It can also be stipulated that only the general teacher model is used in teacher-student training, for example if no specialized teacher model is available on a particular training date.

[0036] In one embodiment, the selection is performed at least partially by means of a trained additional machine learning model, wherein the trained additional machine learning model is trained to select the teacher model(s) most suitable for teacher-student training based on a training datum supplied as input data. This allows suitable teacher models or combinations of teacher models to be found for all training data.

[0037] In a further embodiment, it is provided that the further machine learning model is trained using the trained teacher models, wherein the teacher model which has / have the highest quality in estimating the hands-on / hands-off state is assigned as ground truth to a respective training data item.

[0038] Further features of the device design will become apparent from the description of embodiments of the respective method. The advantages of the respective device are the same as those of the respective method.

[0039] Furthermore, a steering system is also provided, comprising a device for detecting a hands-on / hands-off state on a steering wheel of a vehicle according to one of the described embodiments.

[0040] Furthermore, a vehicle is also provided, comprising a steering system according to one of the described embodiments and / or a device for detecting a hands-on / hands-off state on a steering wheel of a vehicle according to one of the described embodiments.

[0041] The invention is explained in more detail below using preferred embodiments with reference to the figures. Herein: Fig. 1 shows a schematic representation to illustrate embodiments of the device for providing a trained machine learning model for recognizing a hands-on / hands-off state on a steering wheel of a vehicle;

[0042] Fig. 2 is a schematic flow diagram illustrating embodiments of the method for providing a trained machine learning model for detecting a hands-on / hands-off state on a steering wheel of a vehicle;

[0043] Fig. 3 is a schematic diagram illustrating embodiments of the method for providing a trained machine learning model for detecting a hands-on / hands-off state on a steering wheel of a vehicle.

[0044] Fig. 1 shows a schematic representation to illustrate embodiments of the device 1 for providing a trained machine learning model for detecting a hands-on / hands-off state 40 on a steering wheel of a vehicle 50. The device 1 is designed to carry out the method described in this disclosure.

[0045] The device 1 comprises a data processing device 2. The data processing device 2 comprises at least one computing device 2-1 and at least one memory 2-2. The computing device 2-1 can perform computing operations on data stored in the memory 2-2 to implement the method.

[0046] The data processing device 2 is configured to train at least one general teacher model 20 using a training data set 10, to train at least one teacher model 21-x specialized with respect to a context 11-x, and to select training data 10-x of the context 11-x from the training data set 10 and / or to enrich the training data set 10 with training data 10-x of the context 11-x, and to train a student model 22 using the general teacher model 20 and / or the at least one specialized teacher model 21-x by means of teacher-student training. Furthermore, the data processing device 2 is configured to select which of the teacher models 20, 21-x are used to train the student model 22 based on a context 11-x corresponding to the respective training data 10-x.The student model 22 trained in this way is provided by the data processing device 2. It can be provided that the trained student model 22, after training, is loaded into a memory 53 of at least one control unit 52 of a vehicle 50 as part of the provision process. The control unit 52 is part of a device 51 for detecting a hands-on / hands-off state 40 on a steering wheel (not shown) of a vehicle 50. The device 51 further comprises at least one sensor 54 for detecting at least one steering variable 55 on the steering wheel. The at least one steering variable 55 comprises, in particular, a torque.The control unit 52 is configured to receive the detected at least one steering variable 55, to provide and apply the trained student model 22 to detect the hands-on / hands-off state 40, and to supply the detected at least one steering variable 55 to an input of the trained student model 22 for this purpose, and to output the hands-on / hands-off state 40 estimated by the trained student model 22. The hands-on / hands-off state 40 can then be supplied to another control unit 56 of the vehicle 50 for further processing.

[0047] In particular, it may be provided to operate an assistance system of the vehicle 50 based on the hands-on / hands-off state 40. This may, for example, be a lateral guidance assistant.

[0048] Figure 2 shows a schematic flowchart illustrating embodiments of the method for providing a trained machine learning model for detecting a hands-on / hands-off state on a steering wheel of a vehicle. The method can be implemented, for example, using a device as described by way of example with reference to Figure 1.

[0049] In a measure 100, at least one general teacher model is trained using a training data set. In particular, a (single) general teacher model is trained. The training is carried out in a conventional manner, in particular by means of supervised learning.

[0050] In a measure 101, training data is selected from the training dataset according to its context. Furthermore, training data can be enriched according to its context.

[0051] In a measure 102, at least one teacher model specialized with respect to a specific context is trained. In particular, a plurality of specialized teacher models can be trained, each specialized for a specific context. In a measure 103, a student model is trained using the general teacher model and / or the at least one specialized teacher model by means of teacher-student training. For this purpose, measure 103 comprises measures 103a to 103d.

[0052] In measure 103a, a training date is selected from the training dataset. This can be done randomly and / or in groups. In measure 103b, a teacher model (general or specialized) or a group of teacher models (general and / or specialized) suitable for the selected training date is selected. This is done taking into account the context of the training date. In particular, properties of the context are taken into account. For example, a context can have the characteristics already described in the general description. In this way, the most suitable teacher model or the most suitable combination of teacher models for a given context of the training date can be used for each training date. In measure 103c, the student model is trained using teacher-student training in a conventional manner.In step 103d, a check is made to determine whether the training of the student model has already been completed, i.e., whether a specified number of training runs and / or training data has been reached. Furthermore, the quality of the student model can be checked and, depending on the quality achieved, the training of the student model can be continued or terminated. If the check in step 103d reveals that training should not be terminated yet, the program returns to step 103a and continues training with additional training data. If, however, the check in step 103d reveals that the training of the student model has been completed, the program continues with step 104.

[0053] In measure 104, the trained student model is provided. In particular, it is provided that the trained student model is loaded into the memory of at least one control unit of a vehicle during the provision process. Subsequently, a vehicle assistance system can be operated based on the outputs of the trained student model.

[0054] It can be provided that the at least one specialized teacher model 21 -x (Fig. 1) is initialized at the beginning of the training by means of parameters of the trained general teacher model 20.

[0055] It may be provided that the selection is based on the respective

[0056] This is done using metadata 12-x linked to training date 10. Using metadata 12-x, properties and characteristics of context 11-x can be directly described (e.g., trailer present / absent; weather conditions, etc.). Starting from metadata 12-x, corresponding metadata 12-x can be searched for that are assigned to teacher models 20, 21-x, allowing a selection of teacher model 20, 21-x, or teacher models 20, 21-x.

[0057] It can be provided that the selection is carried out at least partially by means of a decision tree. This allows a suitable teacher model 20, 21-x or a suitable combination of teacher models 20, 21-x to be directly selected for a training data item 10-x, based on the context 11-x, which is described, for example, in the form of metadata 12-x.

[0058] It can be provided that the selection is carried out at least partially by means of a trained additional machine learning model 13, wherein the trained additional machine learning model 13 is trained to select the teacher model 20, 21-x or the teacher models 20, 21-x most suitable for the teacher-student training based on a training datum 10-x supplied as input data. The trained additional machine learning model 13 is stored in particular in the memory 2-2, in particular in the form of a structural description and associated parameters, and is provided by the data processing device 2.

[0059] It can be further provided that the further machine learning model 13 is trained using the trained teacher models 20, 21-x, wherein the teacher model 20, 21-x is assigned as ground truth to a respective training date 10-x, or those teacher models 20, 21-x are assigned as ground truth which have(s) the highest quality in estimating the hands-on / hands-off state 40.

[0060] Figure 3 shows a schematic representation to illustrate embodiments of the method for providing a trained machine learning model for recognizing a hands-on / hands-off state on a vehicle steering wheel. The schematic sequence of the teacher-student training is shown. A trained general teacher model 20 and several trained specialized teacher models 21-x are available to train the student model 22.

[0061] During teacher-student training, models 20, 21-x, 22 are fed training data 10-x, i.e., at each training step, models 20, 21-x, 22 are fed a training data 10-x, which is the same for all models 20, 21-x, 22. For example, a training data 10-x comprises the input data xi to x n , which in particular comprise measured values ​​of at least one control variable for various consecutive journals. The training data 10-x corresponds to a context 11-x in which this training data 10-x was recorded or in which this training data 10-x was generated (e.g., by means of a simulation).

[0062] For example, it can be provided that the specialized teacher model 21-1 is a teacher model that is specialized for the context 11-x in which the vehicle has a trailer coupled to a trailer hitch. Furthermore, for example, it can be provided that the specialized teacher model 21-2 is a teacher model that is specialized for a context 11-x in which a vibration is present at the steering wheel, which vibration is applied to the steering wheel, for example, by an actuator to provide a driver with haptic feedback.

[0063] During teacher-student training, a suitable teacher model 20, 21-x or a suitable combination of teacher models 20, 21-x is selected, taking into account the context 11-x corresponding to the respective training date 10-x. For example, if context 11-x corresponds to a training date 10-x, in which a trailer is coupled, the specialized teacher model 21-1 is selected. In the example of context 11-x, in which there is a vibration in the steering wheel, the specialized teacher model 21-2 is selected accordingly. If the context 11-x is not quite so clear, the general teacher model 20 or a combination of teacher models 20, 21-x can be selected. For example, if a trailer is coupled to the vehicle and there is a vibration in the steering wheel, both specialized teacher models 21-1, 21-2 can be used, whereby the outputs 14-x are taken into account in a weighted manner.

[0064] The student model 22 is trained using the output 14-x of the selected teacher model 20, 21-x or the combination of selected teacher models 20, 21-x. For this purpose, the same training data 10-x is fed to the student model 22, and based on this, the student model 22 generates an output 15. This output 15-x is then compared with the output 14-x (by means of knowledge distillation 17) and a ground truth 16-x, which is available for the training data 10-x. Based on a loss function, as described in the general description section, the parameters (weights) of the student model 22 are adjusted. In particular, it is therefore intended that the teacher models 20, 21-x are selected for each training step, taking into account the respective context 11-x of the training data 10-x present in this training step.

[0065] List of reference symbols

[0066] device

[0067] Data processing device -1 Computing device -2 Memory 0 Training data set 0-x Training date 1-x Context 2-x Metadata 3 Further machine learning model 4-x Output (teacher model(s)) 5-x Output (student model) 6-x Ground truth 7 Knowledge distillation 0 General teacher model 1-x Specialized teacher model 2 Student model 0 Hands-on / hands-off state 0 Vehicle 1 Device 2 Control unit 3 Memory 4 Sensor (steering variable) 5 Steering variable 6 Further control unit (assistance system) 00-104 Measures of the procedure

Claims

Patent claims 1. A method for providing a trained machine learning model for detecting a hands-on / hands-off state (40) on a steering wheel of a vehicle (50), comprising: training at least one general teacher model (20) using a training data set (10), Training at least one teacher model (21-x) specialized with respect to a context (11-x), wherein for this purpose training data (10-x) of the context (11-x) are selected from the training data set (10) and / or the training data set (10) is enriched with training data (10-x) of the context (11), Training a student model (22) using the general teacher model (20) and / or the at least one specialized teacher model (21-x) by means of teacher-student training, wherein a selection of which of the teacher models (20, 21-x) are used to train the student model (22) is made based on a context (11-x) corresponding to the respective training date (10-x), Providing the trained student model (22).

2. Method according to claim 1, characterized in that the trained student model (22) is loaded into a memory (53) of at least one control unit (52) of a vehicle (50) after the training as part of the provision.

3. Method according to claim 1 or 2, characterized in that the at least one specialized teacher model (21-x) is initialized at the beginning of the training by means of parameters of the trained general teacher model (20).

4. Method according to one of the preceding claims, characterized in that the selection is carried out based on metadata (12-x) linked to the respective training data (10-x).

5. Method according to one of the preceding claims, characterized in that the selection is carried out at least partly by means of a decision tree.

6. Method according to one of the preceding claims, characterized in that the selection is carried out at least partially by means of a trained further machine learning model (13), wherein the trained further machine learning model (13) is trained to select the teacher model (20, 21-x) or the teacher models (20, 21-x) most suitable for the teacher-student training on the basis of a training data item (10-x) supplied as input data.

7. The method according to claim 6, characterized in that the further machine learning model (13) is trained using the trained teacher models (20, 21-x), wherein that teacher model (20, 21- x) is assigned as ground truth to a respective training datum (10) or those teacher models (20, 21- x) are assigned as ground truth which have(s) the highest quality in estimating the hands-on / hands-off state (40).

8. A method for detecting a hands-on / hands-off state (40) on a steering wheel of a vehicle (50), comprising: Detecting at least one steering variable (55) on the steering wheel of the vehicle (50), using a student model (22) trained according to one of claims 1 to 7 to detect the hands-on / hands-off state (40), wherein the detected at least one steering variable (55) is supplied to an input of the trained student model (22) for this purpose, and Providing the hands-on / hands-off state (40) estimated by the trained student model (22).

9. Device (1) for providing a trained machine learning model for recognizing a hands-on / hands-off state (40) on a steering wheel of a vehicle (50), comprising: a data processing device (2), wherein the data processing device (2) is configured to train at least one general teacher model (20) using a training data set (10), to train at least one teacher model (21-x) specialized with respect to a context (11-x), and for this purpose to select training data (10-x) of the context (11) from the training data set (10) and / or to enrich the training data set (10) with training data (10-x) of the context (11-x), to train a student model (22) by means of the general teacher model (20) and / or the at least one specialized teacher model (21-x) by means of a teacher-student training, and to make a selection as to which of the teacher models (20, 21-x) are used to train the student model (22) based on a context (11-x) corresponding to the respective training date (10-x), and to provide the trained student model (22).

10. A device (51) for detecting a hands-on / hands-off state (40) on a steering wheel of a vehicle (50), comprising: at least one sensor (54) for detecting at least one steering variable (55) on a steering wheel, a control unit (52), wherein the control unit (52) is configured to receive the detected at least one steering variable (55), to provide and use a student model (22) trained according to one of claims 1 to 7 in order to detect the hands-on / hands-off state (40), and to supply the detected at least one steering variable (55) to an input of the trained student model (22) for this purpose, and to provide the hands-on / hands-off state (40) estimated by the trained student model (22).

Images