Method and apparatus for providing a robustized machine learning model for detecting a hands-off state

Abstract

The invention relates to a method for providing a robustized machine learning model (10, 10f) for detecting a hands-off state (41), comprising: obtaining a data set (20) of acquired measurement data and / or acquiring a data set (20) of measurement data, each comprising at least one control variable (40) and a basic truth concerning a hands-off state (41), obtaining and / or providing a machine learning model (10) already trained with a training data set (21) compiled from the data set (20), applying the trained machine learning model (10) to a test data set (22) selected from the data set (20), selecting measurement data misclassified during application from the test data set (22), determining clusters (12) based on the selected misclassified measurement data, and finding acquired measurement data in the remaining data set (23) that correspond to the determined clusters (12).Training the machine learning model (10) taking into account the acquired measurement data, and providing the machine learning model trained in this way (10f). Furthermore, the invention relates to a device (1) for providing a robustized machine learning model (10, 10f) for detecting a hands-off state (41), as well as a method and a device (51) for detecting a hands-off state (41) in a vehicle (50).

Description

The invention relates to a method and a device for providing a robustized machine learning model for detecting a hands-off state. Furthermore, the invention also relates to a method and a device for detecting a hands-off state in a vehicle. It is known to use sensors, such as a capacitive steering wheel, to monitor driver activity in vehicles. Such a capacitive steering wheel detects when a driver touches the steering wheel, and the resulting signal is transmitted to relevant functions, such as a longitudinal and / or lateral guidance assistance system. For example, the touching of the hands on the steering wheel indicates driver activity and thus attentiveness. For instance, the driver can be prompted to place their hands on the steering wheel if they have not been on it for a specified period during lateral guidance. To eliminate the additional costs of a capacitive sensor in the steering wheel, or to reduce the frequency of false warnings, methods are known that monitor driver activity, for example based on detected steering torque, using artificial neural networks. German patent DE 10 2019 211 016 A1 describes, for example, how a capacitive sensor installed in the steering wheel is used to collect training data for such a neural network and thereby develop machine learning-based hands-off detection. German patent application DE 11 2021 000 251 T5 discloses a computer-implemented method for selecting a dataset from a set of datasets for updating an artificial intelligence (AI) module. Each set of datasets comprises an input dataset and a corresponding output dataset. The computer-implemented method includes: obtaining parameter values ​​to define different clusters of the datasets; determining a metric for each dataset, where the metric depends on the degree of membership of the dataset in question in one of the clusters and on the distance of the dataset from the centroid of that cluster; and selecting at least one of the datasets from the set of datasets to update the AI ​​module based on a comparison of the metrics of the datasets. In US 2024 / 0 311 636 A1, systems and methods for using the results of a feature space analysis during the training of machine learning models are presented to improve the training process and the resulting trained model. US Patent 2012 / 0 054 184 A1 provides systems and methods for data classification and new data class detection. In one exemplary embodiment, a new class detection system or method includes receiving a data stream with multiple data points and identifying a set of filtered outliers within those data points that lie outside a decision boundary. The cohesion and separation of the filtered outlier set can be determined. Using the cohesion and separation of the filtered outlier set, a new class can be detected that may contain the filtered outlier set. DE 10 2023 206 652 A1 relates to a method for determining a hands-off situation on the steering wheel of a motor vehicle by means of a monitoring device of the motor vehicle, comprising the steps of: - detecting a moment on the steering wheel by means of a torque detection device of the monitoring device; - evaluating the detected moment by means of an electronic computing device of the monitoring device; - receiving confirmation information of the hands-off situation from a processing device in the motor vehicle that is independent of the torque detection device by means of the electronic computing device; and - determining the hands-off situation depending on the evaluation and depending on the confirmation information by means of the electronic computing device. Furthermore, the invention relates to a computer program product and a monitoring device. DE10 2023 204 053 A1 discloses a method for detecting contact between hands and a vehicle's steering wheel, wherein, based on at least one acquired and / or received state data of a vehicle's steering system, a decision is made by means of at least one trained machine learning method as to whether at least one hand is in contact with the steering wheel or not, wherein the actuation of at least one control element arranged on the steering wheel is taken into account as an input variable of the at least one trained machine learning method, and wherein a decision signal is generated and provided. The invention further relates to a device for detecting contact between hands and a vehicle's steering wheel. A major challenge of steering torque-based detection systems, for example, is identifying the driver-induced steering torque within the detected (and usually noisy) steering torque. Many factors can contribute to this noisy steering torque, particularly the position of a steering torque sensor (which is typically integrated into a steering gear or power steering system, creating a torsional vibration system due to the elasticity of the steering column in conjunction with the steering wheel, making precise detection of the driver-induced steering torque difficult); the degree of friction within the steering system; feedback from road irregularities; the weight of the steering wheel / steering system; and haptic stimulation (e.g., vibration) of the steering wheel by an assistance function, such as a lane departure warning. Furthermore, characteristics of the detected steering torque, which are used for hands-off detection, can change due to external influences, such as temperature, load, the presence of a trailer, tire type and / or tire condition, steering system changes over the lifespan, as well as road gradient, incline and / or slope. To complicate matters further, some influences cannot be measured or are not available as information, for example, a steering wheel position that is individually adjusted to the driver and causes a change in the steering column's angle of deflection, which affects the friction that occurs. Measurable influencing factors, on the other hand, are available at all times. For a machine learning process, data compilation is essential and significantly influences the robustness of the trained system. The multitude of possible combinations of influencing factors, coupled with the uncertainty surrounding the magnitude of each factor's impact, complicates data selection. In particular, this can lead to areas being omitted from the training data, resulting in "white spots"—uncovered data domains—that negatively impact the robustness of the hands-off state detection. In these areas, the machine learning model performs poorly. The invention is based on the objective of providing a method and a device for supplying a robustized machine learning model for detecting a hands-off state on a vehicle's steering wheel. In particular, the aforementioned "open gaps" are to be closed. The problem is solved according to the invention by a method with the features of claim 1 and a device with the features of claim 13. Advantageous embodiments of the invention are set forth in the dependent claims. In particular, a method for providing a robustized machine learning model for detecting a hands-off state is provided, comprising: receiving a dataset of acquired measurement data and / or acquiring a dataset of measurement data, each comprising at least one control variable and one basic truth concerning a hands-off state; receiving and / or providing a machine learning model already trained with a training dataset compiled from the dataset; applying the trained machine learning model to a test dataset selected from the dataset; selecting measurement data misclassified during the application from the test dataset; determining clusters based on the selected misclassified measurement data; and finding acquired measurement data in the remaining dataset that correspond to the determined clusters.Training the machine learning model taking into account the acquired measurement data, and providing the trained machine learning model. Furthermore, a device for providing a robustized machine learning model for detecting a hands-off state is provided, comprising: a data processing device, wherein the device is configured to receive a data collection of acquired measurement data and / or to acquire a data collection of measurement data, the measurement data each comprising at least one control variable and a basic truth concerning a hands-off state, wherein the data processing device is configured to receive and / or provide a machine learning model already trained with a training data set compiled from the data collection, to apply the trained machine learning model to a test data set selected from the data collection, to select misclassified measurement data from the test data set during application, and to determine clusters based on the selected misclassified measurement data.to locate the recorded measurement data in the remaining data collection that correspond to the specific clusters, to train the machine learning model taking into account the located recorded measurement data, and to provide the machine learning model trained in this way. The method and the device make it possible to identify "open gaps" in data domains or data dimensions based on the performance of the machine learning model and to train or retrain the machine learning model with respect to these "open gaps." This increases the robustness of the machine learning model, i.e., its performance across different data domains or data dimensions. A basic idea here is to collect misclassified measurement data and determine clusters based on this misclassified data. Starting from the determined clusters, measurement data for each cluster is retrieved from the data collection. The retrieved measurement data is then added to a training dataset. Subsequently, the retrieved measurement data is removed from the data collection. The retrieved measurement data is then used during training or retraining.Retraining of the machine learning model is taken into account. This allows the machine learning model to be additionally (re)trained with acquired measurement data that lies in the same cluster as acquired measurement data that was previously misclassified. One advantage of the method and the device is that “open gaps” in the performance of the trained machine learning model can be automatically identified and closed. A steering input is, in particular, a quantity that represents and / or describes the current state of the steering wheel. A steering input is, in particular, a torque (steering torque), which is detected, especially by means of a torque sensor on the steering wheel. However, a steering input can also be any other quantity detected directly or indirectly at the steering wheel. The minimum steering input can, for example, include one or more of the following: steering wheel torque, steering wheel angle, steering wheel angular velocity, lateral acceleration, and / or applied force. For example, it may also be possible to detect the current in an electric motor at the steering wheel and use it as a steering input.The at least one steering parameter can also include at least one further (detected and / or determined) parameter that represents and / or describes a state of the vehicle and can, in particular, influence steering behavior. For example, this could be vehicle speed. The machine learning model determines a hands-off state on the steering wheel of a vehicle based on at least one steering parameter. This at least one steering parameter is specifically a time series. The machine learning model includes, in particular, an architecture or structure for processing time series as input data. Specifically, the machine learning model comprises one or more artificial recurrent neural networks, where at each time t the input data Xt is processed and a hands-off probability ytin [0,1] is output: yt = p(xt| x0:t-1). Each recurrent neural network possesses, in particular, a so-called memory h, in which information from previous time steps is stored and which can be used for output at the current time step. The machine learning model can, in particular, comprise a long-short-term memory (LSTM) network with multiple internal layers. The machine learning model includes several inputs (e.g., 5 inputs for steering torque, steering wheel angle or steering wheel angular velocity, lateral acceleration, vehicle speed, and lateral force). For example, the first layer of the LSTM comprises 32 units, and the second layer 16 units, followed by a dense, fully connected output layer with, for example, 8 neurons (Input > LSTM (32) > LSTM (16) > Dense (1)). However, smaller LSTM networks are also possible (e.g., with only 2 x 16 units), for example, when conserving in-vehicle resources (computing power and / or memory) is a priority.To save computing resources, more layers with fewer units are particularly advantageous compared to fewer layers with more units, since the number of necessary calculations increases quadratically with the number of units. During a training phase, the machine learning model is trained. This is done using training data, which comprises pairs in which measurement data of the at least one steering variable, in particular torque data, are paired with a known hands-off state (as the baseline). 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 measured at the steering wheel, in particular time series of torques measured at the steering wheel. The training is otherwise carried out in a manner known per se, in particular by means of supervised learning. In particular, during training by means of backward projection, the parameters of the machine learning model (in particular the weights) are adjusted until a calculated error of the output to the baseline is below a predetermined threshold. The test dataset is, in particular, disjoint from the previously used training dataset. A hands-off state is, in particular, a state in which the driver does not touch the steering wheel. Specifically, none of the driver's fingers are in contact with the steering wheel. Detecting the hands-off state can, in particular, include providing a hands-off state signal. This signal may include, 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. A hands-on / hands-off state can also be provided, for example, as a hands-on / hands-off state signal with, in particular, at least two signal states (e.g., "hands-on detected" or "hands-off detected"). The determination of the clusters is carried out in particular using known methods, such as K-means, dynamic time warping, auto-encoders, etc. Dimensions in which properties of the clusters or the measurement data are located can be defined, in particular, based on the measurement data (time series) themselves and / or based on metadata and / or contextual information that is assigned or will be assigned to a respective measurement. The metadata and / or contextual information can, for example, include one or more of the following quantities and / or information: a steering wheel position; the strength of friction in the steering system; feedback from the road due to unevenness; the weight of the steering wheel / steering system; haptic feedback (e.g.,Vibration) of the steering wheel due to an assistance function, temperature, load, the presence of a trailer, tire type and / or tire condition, steering system changes over its lifespan, as well as road gradient, incline and / or slope, etc. The result of determining this is, in particular, a set of disjoint clusters. The procedure can also be carried out as a computer-implemented procedure. Parts of the devices described in this disclosure, in particular the control unit and / or the data processing unit, may be configured individually or collectively as a combination of hardware and software, for example, as program code executed on a computing unit, in particular a microcontroller or microprocessor. However, it may also be provided that parts are configured individually or collectively as an application-specific integrated circuit (ASIC) and / or a field-programmable gate array (FPGA) and / or a graphics processing unit (GPU) and / or a digital signal processor (DSP). The control unit or the data processing unit may, in particular, comprise at least one computing unit and at least one memory. In one embodiment, the determination of clusters and the retrieval of the acquired measurement data are only performed if a proportion of misclassifications exceeds a predefined threshold. This allows a minimum level of accuracy to be specified that the trained machine learning model must achieve. The proportion of misclassifications is determined, in particular, as a proportion of the total number of classifications performed. In one embodiment, a subset is selected from the specified clusters, with the selection or rejection of clusters based on at least one predefined criterion, and only clusters from the selected subset being considered during the search. This allows for the targeted selection and consideration of relevant clusters, for example, clusters that are particularly relevant due to their size and / or the number of corresponding measurement data points. For instance, it might be possible to select the n (e.g., 10) largest clusters. In a further developed embodiment, the specified criterion includes at least one criterion: that clusters are rejected if their mean center-of-mass distance exceeds a predetermined distance threshold and the number of measurement data points in the cluster falls below a predetermined minimum. This allows measurement data points with insufficient cluster density to be rejected, thereby disregarding less relevant or irrelevant clusters or measurement data points. The center-of-mass distance is defined, in particular, as the distance from the measurement data points to the center (center of mass) of the respective cluster. In a further developed embodiment, it is also provided that the specified criterion includes at least one criterion that clusters are considered which have a mean center-of-mass distance below a specified distance threshold and a number of measurement data points in the cluster exceeding a specified minimum number. This allows measurement data points with a high density within the cluster to be considered, even if the cluster itself is small. In particular, this allows clusters or measurement data points that describe rarely occurring cases, so-called corner cases, to be considered. In one embodiment, the at least one criterion includes prioritizing the selection of clusters with a greater center-of-gravity distance from other clusters. This allows for the consideration of clusters or measurement data associated with systematic errors. For example, a predetermined number of clusters can be selected, and this number can be preferentially filled by the prioritized clusters. In one embodiment, measurement data is assigned an occurrence probability value and / or a criticality value, with at least one criterion requiring that a predefined occurrence probability threshold and / or criticality threshold within the cluster must be exceeded for the cluster to be selected. This allows measurement data with a reduced occurrence probability and / or reduced criticality to be disregarded. The occurrence probability (or the associated value) can, for example, be a value from a histogram across all input signals, i.e., information about how frequently a combination of input signal values ​​occurs (e.g., measured against the training or test data set).It may be intended that the criticality value is determined based on the probability of occurrence, for example, by using a mean, median, maximum, or percentile probability of occurrence as the criticality value. Clusters can then be selected (prioritized) based on the criticality value determined in this way. For example, the criticality value for the hands-off state may be chosen to be higher than for the hands-on state, since it is more important to detect the hands-off state. In one embodiment, it is provided that a predetermined maximum distance to the center of the respective cluster is taken into account when locating the recorded measurement data. This prevents recorded measurement data that has been assigned to the nearest cluster, but which does not correspond to that cluster due to its distance from the center, from being taken into account. In particular, recorded measurement data at a distance exceeding the specified maximum distance will then not be considered. In one embodiment, it is provided that when locating the acquired measurement data, a uniform distribution of the acquired measurement data within the respective cluster is sought. This ensures the representativeness of the acquired measurement data extracted from a cluster. A uniform distribution here refers in particular to a density of acquired measurement data that lies within a predefined range with respect to the dimensions of the cluster. In one embodiment, when locating the acquired measurement data, the number of data points considered for each cluster is selected proportionally to the size of that cluster. This allows the size of the cluster to be taken into account based on the number of acquired measurement data points originating from that cluster. The rationale is that if an error occurs very frequently, training data associated with that error may also be present (relatively) frequently in the training dataset: the bias is representative. The goal is, in particular, to improve the performance of the machine learning model with regard to the most frequently occurring errors; therefore, the associated measurement data are given more weight during training than data from less frequently occurring errors.In this context, it is specifically assumed that the number of training data points of a cluster correlates with the performance of the machine learning model for that cluster during training. In one embodiment, a distinction is made between false-negative and false-positive misclassifications when determining the clusters. This allows, in particular, for a weighting of these two types of misclassification. Specifically, this makes it possible to give greater weight to false-negative misclassifications and to consider them more strongly in subsequent training, since this misclassification describes a more critical case in hands-off detection—that is, a case in which hands are detected on the steering wheel even though they are not. In one embodiment, it is provided that the determination of the clusters takes into account a distance between the value output by the machine learning model for the hands-off state and a predetermined classification threshold. This allows for a measure of the severity of misclassification. For example, clusters can be differentiated based on their spacing. The classification threshold specifically indicates when a hands-off probability is recognized as a hands-off state. For instance, if the threshold for the hands-off state is set at 0.5 and a current data point is associated with a hands-off state (i.e., a value of 1 for the baseline truth), then an output value from the machine learning model of <0.5 would indicate a misclassification.If the output value is 0.49, this is considered less critical (because the output value itself is considered uncertain, as it is far from the minimum value of 0 and the maximum value of 1) than an output value of 0.01 (= no hands-off state), which the machine learning model provides as a "very certain" output (since the value is very close to the value 0 for "no hands-off state" or "hands-on"). By taking this distance into account, such misclassifications can be reduced. The uncertainty expressed in the distance can therefore be used to improve the performance of the machine learning model. In particular, the goal is to (re)train the machine learning model more effectively for clusters with a large distance. In one embodiment, when selecting misclassified measurement data, only those data points are considered where the misclassification reaches a predefined minimum duration. This takes into account that, in hands-off detection, a short misclassification (e.g., < 100 ms) is less critical than a persistent misclassification (e.g., >= 500 ms), as the former is easier to filter. Excessive filtering, on the other hand, would lead to a large number of false negatives (at the end-user level). Therefore, only misclassifications exceeding the predefined minimum duration (e.g., >= 500 ms) or which were present for the specified minimum duration can be selected for subsequent cluster determination. In one embodiment, the determination of the clusters is provided for, at least partially, based on further processed, acquired measurement data. This allows for the consideration of features and properties contained in the acquired measurement data that can be extracted through further processing. For example, it may be possible to subject the (time series of) acquired measurement data to a Fourier or wavelet transformation and / or another type of filtering, and then to determine the clusters based on the transformed and / or filtered measurement data (e.g., based on a frequency spectrum). In one embodiment, it is provided that when locating the acquired measurement data, only those data points are considered that are located at a distance from the center of a cluster corresponding to the acquired measurement data that is below a predetermined maximum distance. This ensures a minimum similarity between the considered acquired measurement data and the respective center of the cluster. In one embodiment, it is provided that further measurement data for the data collection are acquired and / or generated if the number of acquired measurement data in the corresponding cluster is greater than the remaining number of acquired measurement data corresponding to the same cluster in the data collection. This allows, in the event of missing measurement data for a cluster, the acquisition of further measurement data in a data domain corresponding to the cluster to be initiated, thereby increasing the robustness of the trained machine learning model by ensuring that it also provides sufficient accuracy in detecting the hands-off state with respect to this cluster or data domain. In one embodiment, providing the trained machine learning model includes loading it into the memory of at least one electronic control unit (ECU) of at least one vehicle. For this purpose, a (coded) structural description and parameters (especially weights) of the trained machine learning model are loaded into the memory. The trained machine learning model can then be used in the ECU. Further features relating to the design of the device are described in the description of embodiments of the method. The advantages of the device are the same in each case as in the embodiments of the method. Furthermore, a method for detecting a hands-off state in a vehicle is provided, comprising: capturing at least one steering parameter, supplying the captured at least one steering parameter as input data to a machine learning model trained according to one of the embodiments described in this disclosure, determining the hands-off state using the trained machine learning model based on the supplied at least one steering parameter, and outputting the determined hands-off state. Furthermore, in particular, a device for detecting a hands-off state in a vehicle is provided, comprising at least one sensor configured to detect at least one steering parameter, and a control unit, wherein the control unit is configured to provide a machine learning model trained according to one of the embodiments described in this disclosure, to supply the detected at least one steering parameter to the trained machine learning model as input data, to determine the hands-off state by means of the trained machine learning model based on the supplied at least one steering parameter, and to output the determined hands-off state. The invention is explained in more detail below with reference to preferred embodiments and the figures. Figure 1 shows a schematic flowchart of embodiments of the method for providing a robustized machine learning model for detecting a hands-off state; Figure 2 shows a schematic representation of an embodiment of the device for generating training data for training a machine learning model for hands-off detection; Figure 3 shows a schematic flowchart of an embodiment of the method for detecting a hands-off state in a vehicle; Figure 4 shows a schematic representation of an embodiment of the device for detecting a hands-off state in a vehicle. Fig. 1 shows a schematic flowchart of embodiments of the method for providing a robustized machine learning model 10f for detecting a hands-off state. In process step 100, a data set 20 of recorded measurement data is obtained and / or a data set 20 of measurement data is recorded. The measurement data each comprise at least one control variable (in particular in the form of a time series) and each a basic truth regarding a hands-off state. In process step 101, the machine learning model 10 is (initially) trained using a training dataset 21 compiled from the data collection 20. Alternatively, the machine learning model can already be (initially) trained using a training dataset 21 compiled from the data collection. The (initially) trained machine learning model 10 is then made available. In process step 102, the trained machine learning model 10 is applied to a test dataset 22 selected from the data collection 20. Based on at least one control variable of the recorded measurement data in the test dataset 22, the trained machine learning model 10 infers a hands-off state, which is compared with a corresponding basic truth. An evaluation result 11 is obtained, indicating whether a classification by the trained machine learning model 10 is correct or incorrect (misclassification). In process step 103, it may be stipulated that a proportion of the misclassifications is compared with a predefined threshold. If the proportion does not exceed the threshold, the performance of the trained machine learning method 10 is deemed sufficient, and a final trained (robust) machine learning model 10f is provided. If, however, the threshold is exceeded, process step 104 is continued. In process step 104, misclassified measurement data are selected from the test dataset during application. In particular, all misclassified measurement data are selected. In process step 105, clusters are determined based on the selected misclassified measurement data. It is possible to determine the clusters based on properties of the measurement data itself. Alternatively or additionally, it is possible to determine the clusters based on metadata and / or contextual information associated with the measurement data. Well-known clustering methods can be used here, such as K-means, dynamic time warping, auto-encoders, etc. The result of process step 105 is a set of clusters 12. In a process step 106, it may be provided that a subset 13 is selected from the specified clusters 12, whereby the selection or rejection of the clusters 12 is based on at least one predefined criterion, and only clusters 12 of the selected subset 13 are considered when finding them. However, in principle, all specified clusters 12 can also be considered. In process step 107, recorded measurement data are located in the remaining data set 23, i.e., the measurement data from data set 20 excluding the previously used training data 21, which correspond to the specified clusters 12 (or the subset 13 thereof). From the remaining data set 21, recorded measurement data are extracted according to the specified clusters 12, which correspond to the specified clusters 12 or to the properties they represent. As a result of process step 107, additional recorded measurement data 24 are obtained, which are added to the training data set 21. Subsequently, in process step 101, the machine learning model 10 is trained or retrained with the training data set 21 enriched in this way. The procedure is repeated, in particular, until the check in procedure step 103 shows that the performance of the trained machine learning model 10 is sufficient. The final trained (robust) machine learning model 10f is provided in process step 108. It may be provided in process step 108 that the provision of the final trained (robust) machine learning model 10 includes loading it into a memory of at least one control unit of at least one vehicle. In process step 106, it may be stipulated that the n largest clusters are selected, while the remaining clusters 12 are discarded. A cluster size can be determined, in particular, by the number of measurement data points that were assigned to cluster 12 when determining it. It may be provided in process step 106 that the specified at least one criterion includes that clusters 12 are rejected which have a mean center of gravity distance above a specified distance threshold and a number of measurement data in cluster 12 is below a specified minimum number. In process step 106, it may be stipulated that those clusters with a greater centroidal distance from other clusters are prioritized. This is particularly relevant if the number of clusters considered is limited or becomes limited. It may be provided that measurement data is assigned an occurrence probability value and / or a criticality value, whereby at least one criterion in process step 106 includes the requirement that a predefined occurrence probability threshold and / or a predefined criticality threshold within cluster 12 must be exceeded for cluster 12 to be selected. The respective thresholds can, for example, be manually defined by experts. In process step 107, it may be stipulated that a predefined maximum distance to the center of the respective cluster 12 is taken into account when locating the recorded measurement data. In particular, in this case, only those measurement data points are considered for the cluster 12 under consideration that have a distance to the center of the cluster 12 that is smaller than the predefined maximum distance. In process step 107, it may be stipulated that when locating the recorded measurement data, an even distribution of the recorded measurement data within the respective cluster 12 is sought. The located recorded measurement data are then taken into account in such a way that they are evenly distributed within the respective cluster 12. It may be provided in process step 107 that when finding the recorded measurement data, a number of the considered recorded measurement data per respective cluster 12 is chosen proportionally to the size of the respective cluster 12. It may be provided in procedure step 105 that a distinction is made between false-negative and false-positive misclassifications when determining cluster 12. In process step 105, it may be provided that the determination of clusters 12 takes into account a distance between the value for the hands-off state output by the machine learning model 10 and a predefined classification threshold. For example, one cluster may be provided for each of several ranges of the distance. It may be provided in process step 104 that when selecting the misclassified measurement data, only those measurement data are taken into account where a misclassification reaches a specified minimum duration. In process step 105, it may be provided that the determination of clusters 12 is based, at least partially, on further processed, acquired measurement data. For this purpose, the acquired measurement data is further processed before the determination of clusters 12, for example, by means of a Fourier or wavelet transformation and / or filtering. In process step 107, it may be provided that when locating the recorded measurement data, only recorded measurement data are taken into account which have a distance to a center of a cluster 12 corresponding to the recorded measurement data that is below a specified maximum distance. In process step 109, it may be provided that further measurement data for data collection 20 are recorded and / or generated if the number of recorded measurement data in the corresponding cluster 12 is greater than the remaining number of recorded measurement data corresponding to the same cluster 12 in data collection 20. This may occur in particular if a predetermined minimum number of such clusters 12 has been reached. Fig. 2 shows a schematic representation of an embodiment of the device 1 for providing a robustized machine learning model 10f for detecting a hands-off state. The device 1 comprises a data processing unit 2. The data processing unit 2 comprises, in particular, a computing unit 2-1 and a memory 2-2. The device 1 is specifically designed to carry out the method according to the embodiment shown in Fig. 1. The device 1 is designed to obtain a data collection 20 of recorded measurement data and / or to acquire a data collection 20 of measurement data, the measurement data each comprising at least one steering parameter and a basic truth concerning a hands-off state. The data processing facility 2 is configured to obtain and / or provide a machine learning model 10 already trained with a training data set compiled from the data collection 20, to apply the trained machine learning model 10 to a test data set 22 (Fig. 1) selected from the data collection 20, to select misclassified measurement data from the test data set 22 during application, to determine clusters 12 (Fig. 1) based on the selected misclassified measurement data, to find recorded measurement data in the remaining data collection 23 (Fig. 1) that correspond to the determined clusters 12, to train the machine learning model 10 taking into account the found recorded measurement data, and to provide the (finally) trained (robustified) machine learning model 10f. Provisioning can in particular include loading the (finally) trained (robust) machine learning model 10f into a memory of a control unit 53. Fig. 3 shows a schematic flowchart of an embodiment of the method for detecting a hands-off state in a vehicle. In process step 200, at least one steering parameter is acquired, for example, a steering wheel torque, a steering wheel angle, and / or a steering wheel angular velocity. In process step 201, a hands-off state is determined using the trained machine learning model based on the supplied steering parameter. Specifically, the trained machine learning model outputs an estimate for the presence of a hands-off state. This estimate can, for example, lie in an interval between 0 and 1, where the value represents the probability of the hands-off state being present. In process step 202, the specified hands-off state is output. This specific hands-off state can, for example, be fed to a driver assistance system and / or a vehicle control system, where it can be evaluated and / or further processed. Fig. 4 shows a schematic representation of an embodiment of the device 51 for detecting a hands-off state 41 in a vehicle 50. The device 51 is, in particular, arranged in the vehicle 50. The device 51 comprises at least one sensor 52, which is configured to detect at least one steering parameter 40, for example a steering torque, and a control unit 53. The control unit 53 comprises, for example, a computing unit 53-1 and a memory 53-2. The control unit 53 is configured to provide a machine learning model 10f trained (robust) according to one of the embodiments described (final) in this disclosure, to supply the detected at least one steering parameter 40 to the trained machine learning model 10f as input data, to determine the hands-off state 41 by means of the trained machine learning model 10f based on the supplied at least one steering parameter 40, and to output the determined hands-off state 41. The specified hands-off state 41 can, for example, be fed to a driver assistance system 54 and / or a vehicle control system 55. Based on this, a warning message can, for example, be generated and issued when the driver should put their hands on the steering wheel. Reference symbol list 1 Device 2 Data processing device 2-1 Computing unit 2-2 Memory 10 Trained machine learning model 10f Final trained (robust) machine learning model 11 Evaluation result 12 Cluster 13 Subset 20 Data collection 21 Compiled training dataset 22 Test dataset 23 Remaining data collection 24 Additional acquired measurement data 40 Steering parameter 41 Hands-off state 50 Vehicle 51 Device 52 Sensor 53 Control unit 53-1 Computing unit 53-2 Memory 54 Driver assistance system 55 Vehicle control 100-108 Process steps 200-202 Process steps

Claims

Method for providing a robustized machine learning model (10, 10f) for detecting a hands-off state (41), comprising: obtaining a data set (20) of acquired measurement data and / or acquiring a data set (20) of measurement data, each comprising at least one control variable (40) and one basic truth concerning a hands-off state (41); obtaining and / or providing a machine learning model (10) already trained with a training data set (21) compiled from the data set (20); applying the trained machine learning model (10) to a test data set (22) selected from the data set (20); selecting measurement data misclassified during application from the test data set (22); determining clusters (12) based on the selected misclassified measurement data; finding acquired measurement data in the remaining data set (23) that correspond to the determined clusters (12).Training the machine learning model (10) taking into account the acquired measurement data, providing the machine learning model trained in this way (10f). Method according to claim 1, characterized in that the determination of the clusters (12) and the retrieval of the recorded measurement data is only carried out if a proportion of the misclassifications exceeds a predetermined threshold. Method according to claim 1 or 2, characterized in that a subset is selected from the specified clusters (12), wherein the selection or rejection of the clusters (12) is based on at least one predetermined criterion, wherein only clusters (12) of the selected subset (13) are considered when finding. Method according to claim 3, characterized in that the specified at least one criterion comprises that clusters (12) are rejected which have a mean center of gravity distance above a specified distance threshold and a number of measurement data in the cluster (12) is below a specified minimum number. Method according to claim 3 or 4, characterized in that those clusters (12) are prioritized for selection which have a greater center of gravity distance to the other clusters (12) compared to other clusters (12). Method according to one of claims 3 to 5, characterized in that measurement data is assigned or is assigned an occurrence probability value and / or a criticality value, wherein the at least one criterion includes that a predetermined occurrence probability threshold and / or a predetermined criticality threshold must be exceeded within the cluster (12) in order for the cluster (12) to be selected. Method according to one of the preceding claims, characterized in that a predetermined maximum distance to the center of the respective cluster (12) is taken into account when locating the recorded measurement data. Method according to one of the preceding claims, characterized in that when finding the recorded measurement data, a number of the considered recorded measurement data per respective cluster (12) is selected proportionally to a size of the respective cluster (12). Method according to one of the preceding claims, characterized in that when selecting the misclassified measurement data, only those measurement data are taken into account in which a misclassification reaches a predetermined minimum duration. Method according to one of the preceding claims, characterized in that when finding the recorded measurement data, only recorded measurement data are taken into consideration which have a distance to a center of a cluster (12) corresponding to the recorded measurement data which is below a predetermined maximum distance. Method according to one of the preceding claims, characterized in that further measurement data for the data collection (20) are recorded and / or generated if a number of recorded measurement data in the corresponding cluster (12) is greater than a remaining number of recorded measurement data corresponding to the same cluster (12) in the data collection (20). Method according to one of the preceding claims, characterized in that the provision of the trained machine learning model (10f) comprises loading it into a memory of at least one control unit (53) of at least one vehicle (50). Device (1) for providing a robustized machine learning model (10, 10f) for detecting a hands-off state (41), comprising: a data processing device (2), wherein the device (1) is configured to receive and / or acquire a data set (20) of measurement data, the measurement data each comprising at least one control variable (40) and a basic truth concerning a hands-off state (41), wherein the data processing device (2) is configured to receive and / or provide a machine learning model (10) already trained with a training data set (21) compiled from the data set (20), to apply the trained machine learning model (10) to a test data set (22) selected from the data set (20), to select misclassified measurement data from the test data set (22) during application, and to create clusters (1) based on the selected misclassified measurement data. determine,to locate the recorded measurement data in the remaining data collection (20) that correspond to the specified clusters (12), to train the machine learning model (10) taking into account the located recorded measurement data, and to provide the machine learning model trained in this way (10f). Method for detecting a hands-off state (41) in a vehicle (50), comprising: capturing at least one steering parameter (40), supplying the captured at least one steering parameter (40) to a machine learning model (10f) trained according to one of claims 1 to 12 as input data, determining the hands-off state (41) by means of the trained machine learning model (10f) based on the supplied at least one steering parameter (40), and outputting the determined hands-off state (41). Device (51) for detecting a hands-off state (41) in a vehicle (50), comprising: at least one sensor (52) configured to detect at least one steering parameter (40), and a control unit (53), wherein the control unit (53) is configured to provide a machine learning model (10f) trained according to any one of claims 1 to 12, to supply the detected at least one steering parameter (40) to the trained machine learning model (10) as input data, to determine the hands-off state (41) by means of the trained machine learning model (10f) based on the supplied at least one steering parameter (40), and to output the determined hands-off state (41).

Images