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

The method and device improve hands-off state detection on a vehicle's steering wheel by enriching the training dataset with clustering techniques, addressing noisy torque signals and external influences, thus enhancing detection accuracy and reducing costs.

WO2026233582A1PCT designated stage Publication Date: 2026-11-12VOLKSWAGEN AG
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/EP2026/052351
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-08
Filing Date
2026-01-29
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing systems for detecting a hands-off state on a vehicle's steering wheel face challenges due to noisy steering torque signals from factors like torsional vibrations, friction, road irregularities, and external influences, leading to false detections and the need for additional sensors to account for vertical excitations, which increases costs and limits scalability.

Method used

A method and device that enhance a machine learning model for hands-off detection by identifying dominant features in misclassified data without additional sensor information, using clustering techniques to enrich the training dataset with data from vehicles without such sensors, thereby improving robustness against vertical excitations.

Benefits of technology

The enhanced machine learning model effectively detects hands-off states without requiring additional sensors, reducing costs and enhancing performance across various driving conditions, ensuring consistent detection accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure EP2026052351_12112026_PF_FP_ABST
    Figure EP2026052351_12112026_PF_FP_ABST
Patent Text Reader

Abstract

The invention relates to a method for providing a trained machine learning model (10) for detecting a hands-off state (41), involving: obtaining a collection (20) of measurement data, each piece of measurement data comprising at least one steering variable (40) and a ground truth (43), wherein some of the measurement data additionally comprises at least one respective piece of additional sensor information (42); applying a trained machine learning model (10) to a selected test data set (31); selecting measurement data that was incorrectly classified during the application process; identifying the incorrectly classified measurement data that comprises the at least one piece of additional sensor information (42); determining first clusters (C1); determining dominant features (22) in the first clusters (C1); extracting the first clusters (C1) for which features of the at least one piece of additional sensor information (42) are dominant; determining second clusters (C2) on the basis of the steering variable (40) for the measurement data contained in the extracted first clusters (C1); assigning measurement data of another data collection (23) without the at least one piece of additional sensor information (42) to the second clusters (C2); enhancing the training data set (10) with measurement data of the other data collection (23) which has a distance to the second clusters (C2) that falls below a distance threshold value (D); training the machine learning model (10) using the enhanced training data set (30+); and providing the trained machine learning model (10).
Need to check novelty before this filing date? Find Prior Art

Description

[0001] Description

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

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

[0004] 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 longitudinal and / or lateral guidance systems. For example, the touching of the hands on the steering wheel is interpreted as driver activity and thus as attentiveness.

[0005] For example, the driver may be reminded to put their hands on the steering wheel if their hands have not been on the steering wheel for a specified period of time during a lateral maneuver.

[0006] 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 a detected steering torque, using artificial neural networks. German patent DE 102019211 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.

[0007] 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.

[0008] 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.

[0009] For a machine learning process, data composition is essential and significantly influences the robustness of the trained system. In particular, an unintended bias can lead to a systematic error, as it is also learned. For this reason, it is important to balance the training data accordingly. For example, during natural driving, the hands are kept on the steering wheel and only very rarely, and then only very briefly, removed. A balanced dataset would thus contain, for example, 99% data with a hands-on state. An optimization tool could, in the worst case, completely ignore data with a hands-off state. On the other hand, it can be useful to deliberately introduce a certain (but not excessive) bias to improve performance in a specific data domain (Operational Design Domain, ODD) (e.g., for hands-off driving).This is just one example of an influencing factor, and the sheer number of possible combinations of influencing factors is enormous. Furthermore, it is usually unknown how the data must be related to each other to create the most robust machine learning model possible. In this context, robustness refers specifically to the property that the machine learning model performs as consistently well as possible across different domains and does not exhibit a performance bias in, for example, one domain (e.g., hands-on).

[0010] Besides maintaining a balance between hands-on and hands-off operation, situations that are potentially underrepresented and difficult to describe can lead to critical system performance. In particular, disturbances from the road surface, for example, can transmit very large moments to the steering system, which can lead to false detections in the hands-off state. To describe these scenarios and represent them with sufficient balance in the training and test data, information on vertical excitation of the steering system is relevant. This allows for an objective description of critical states and robustness of hands-off detection against disturbances resulting from poor road conditions.

[0011] One way to gather relevant information regarding vertical excitation is through special vehicle equipment options with additional sensors that provide adaptive chassis control. With adaptive chassis control, implemented through adjustable dampers, the suspension can typically be adjusted in predefined increments at the touch of a button, for example, to achieve normal, sporty, or comfortable settings. The electrically adjustable damping is adapted based on driver input, driving situation, and road surface conditions. Adaptive chassis control improves handling and allows the driver to actively experience different suspension settings within the vehicle. This requires additional sensors that provide the chassis control system with data on body acceleration and vehicle height at various measuring points.However, the additional sensors represent an additional cost factor and are largely unavailable in vehicles with basic equipment. Consequently, it is not possible to use information from vehicles with basic equipment for the entire state space of the machine learning-based application described above. This significantly limits scalability and entails additional costs for training and testing to map all variant variations and describe the entire action and state space.

[0012] The invention is based on the objective of improving a method and a device for providing a trained machine learning model for recognizing a hands-off state on a vehicle's steering wheel.

[0013] In particular, a method for providing a trained machine learning model for detecting a hands-off state on a vehicle's steering wheel is provided, comprising: receiving a data set of acquired measurement data and / or acquiring a data set of measurement data, each comprising at least one steering parameter and a basic truth regarding a hands-off state, wherein a portion of the measurement data in the data set additionally includes at least one supplementary sensor piece of information; receiving and / or providing a machine learning model already trained with a training data set compiled from the data set, neglecting the at least one supplementary sensor piece of information; applying the trained machine learning model to a test data set selected from the data set; selecting measurement data from the test data set that were misclassified during the application; and identifying those misclassified measurement data.which include at least one additional sensor information, determining first clusters for the misclassified measurement data, determining dominant features in the determined first clusters, extracting those first clusters for which features of the at least one additional sensor information are dominant features, determining second clusters based on the at least one control variable for the misclassified measurement data contained in the extracted first clusters, assigning measurement data from another data set without the at least one additional sensor information to the determined second clusters, enriching the training data set with those assigned measurement data from the other data set which have a distance to the second clusters that is below a predefined distance threshold, training the machine learning model using the enriched training data set, providing the trained machine learning model.

[0014] Furthermore, in particular a device for providing a trained machine learning model for detecting a hands-off state on a steering wheel of a vehicle is provided, comprising a data processing device, wherein the data processing device is configured to perform the following steps:

[0015] 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 regarding a hands-off state, wherein a portion of the measurement data in the dataset additionally includes at least one additional sensor information; obtaining and / or providing a machine learning model already trained with a training dataset compiled from the dataset, neglecting the at least one additional sensor information; 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; identifying those misclassified measurement data that include the at least one additional sensor information; determining initial clusters for the misclassified measurement data; determining dominant features in the determined initial clusters.Extracting those first clusters for which features of at least one additional sensor information are dominant, determining second clusters based on the at least one control variable for the misclassified measurement data contained in the extracted first clusters, assigning measurement data from another data set without the at least one additional sensor information to the determined second clusters, enriching the training data set with those assigned measurement data from the other data set that have a distance to the second clusters below a predefined distance threshold, training the machine learning model using the enriched training data set, and providing the trained machine learning model.

[0016] The method and the device enable the provision of an improved machine learning model for detecting a hands-off state on a vehicle's steering wheel. In particular, the performance of the machine learning model can be improved in the presence of vertical excitations in the steering system, even if no sensors are present to detect these excitations. The method specifically provides that, based on at least one additional sensor information, measurement data from a data set is identified that lack this additional sensor information. These data must exhibit characteristic features in the at least one steering parameter they contain that are associated with or correlate with vertical excitation. This measurement data is then used to (re)train the machine learning model.

[0017] The advantage of the method and the device is that, due to improved training, a machine learning model can be provided that is particularly robust against vertical excitations. Additional sensors, especially for detecting vertical excitations, are unnecessary, thus saving costs.

[0018] This involves obtaining and / or acquiring a dataset of recorded measurement data, each comprising at least one control parameter and a basic truth regarding a hands-off state, with a portion of the measurement data in the dataset additionally including at least one supplementary sensor piece of information. This supplementary sensor information may, in particular, include sensor data representing vertical excitation. A machine learning model, already trained on a training dataset compiled from the dataset while neglecting the at least one supplementary sensor piece of information, will be obtained and / or provided. The trained machine learning model will be applied to a test dataset selected from the dataset. During application, only the at least one control parameter will be considered as input data.In other words, the machine learning model only has inputs for at least one steering parameter; at least one additional sensor information is not considered. During application, misclassified measurement data is selected from the test dataset. Specifically, those measurement data are selected for which the machine learning model's output does not correspond to a given basic truth (no hands-off state or hands-off state). From the selected measurement data, those misclassified measurement data that include at least one additional sensor information are identified. This can be done, for example, via an assignment or linking, which could be based on vehicle information and / or vehicle type information.

[0019] Initial clusters are determined for the misclassified measurement data. In this process, all misclassified measurement data are specifically considered. The determination of these initial clusters can be performed using known clustering methods, such as k-means or time-series clustering. The clustering is based primarily on the measurement data and / or on features derived from the measurement data itself. Specifically, at least one steering parameter is used directly as a feature. Furthermore, at least one additional sensor piece of information can serve as a feature (e.g., vertical accelerations, which can also be referred to as Dynamic Chassis Control (DCC) accelerations). For example, five steering parameters can be considered, meaning five dimensions in which groups can be formed using a clustering method. Since time series are being considered, the dimensionality can be increased arbitrarily across the time steps (e.g.,Each journal has 5 values ​​for the exemplary control variables x 10 time steps = 50 dimensions). Clustering methods attempt to create groups based on a distance measure between each data point and every other data point, where the distance within the group is as small as possible. Clustering methods are specifically classified as unsupervised learning methods.

[0020] Dominant features are identified within the initial clusters. Dominant features are those that strongly influence or are responsible for cluster formation. "Dominant" means that a cluster is primarily defined or formed by this feature. Proximity [in this dimension(s)] between the data points thus enabled clusters to be formed / found. For example, if measurement data were always recorded at 30 km / h and 100 km / h, this would result in two clusters with speed as the dominant feature. Therefore, the focus is on identifying (initial) clusters that have "found" each other due to their proximity in the dimensions of at least one additional sensor information (especially the vertical accelerations or DCC values). Other dimensions are not relevant to the disclosed method and device.If the (first) clusters are examined and, for example, it is found that a (first) cluster has a small variance and large values ​​in the at least one additional sensor information (especially in the vertical accelerations or DCC setup accelerations), while the velocity values ​​are distributed across the entire range, it can be assumed (because of the distance minimization) that the at least one additional sensor information (especially the vertical accelerations) was dominant for the cluster formation.

[0021] From the first clusters, those clusters are extracted for which features of the sensor data representing at least one additional sensor information, in particular the features of the vertical excitations, are dominant features.

[0022] Second clusters are then determined based on at least one control parameter for the misclassified measurement data contained in the extracted first clusters. The at least one additional sensor information is specifically not taken into account in this process.

[0023] In other words, the misclassified measurement data contained in the extracted first clusters are used, neglecting the additional sensor information, to determine the second clusters based on at least one steering parameter. Since this is only done for the extracted first clusters, the measurement data are correlated with the additional sensor information, particularly with respect to the at least one steering parameter.

[0024] The measurement data from a further data set, lacking at least one additional sensor piece of information (which is independent of the data set itself), are assigned to specific second clusters. This assignment is based primarily on the respective values ​​of the features corresponding to the clusters in the measurement data.

[0025] The training dataset is enriched with those assigned measurement data from the further data collection that show a distance to the second clusters below a predefined distance threshold. This results in an enriched training dataset. The distance is determined in particular by means of a distance measure, specifically through an internal evaluation, that is, solely based on the measurement data assigned to each cluster.

[0026] The machine learning model is trained using the enriched training dataset. This is done in a known manner, particularly through supervised learning. The trained machine learning model is then provided. The data collection includes, in particular, measurement data from test drives, which can vary in length. This data collection includes measurement data acquired with vehicles (e.g., basic models) that do not capture at least one additional sensor information, and measurement data acquired with vehicles (e.g., vehicles with optional equipment such as adaptive suspension) that capture at least one additional sensor information using suitable sensors.

[0027] 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.

[0028] The machine learning model determines a hands-off state on a vehicle's steering wheel 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 includes one or more artificial recurrent neural networks, where the input data X is available at each time t. t be processed and a hands-off probability (or a probability of the hands-off state existing) y toutput in [0,1]: y t = p(x t | xo :ti). In this case, a recurrent neural network possesses, in particular, a so-called memory h, in which information from previous journals is stored and which can be used for the output of the current journal. The machine learning model includes, in particular, an input and an output for the hands-off probability for each steering parameter considered. The machine learning model can, in particular, comprise a long-short-term memory (LSTM) network with several internal layers. The machine learning model includes, in particular, several inputs (e.g., 5 inputs for a steering torque, a steering wheel angle or steering wheel angular velocity, a lateral acceleration, a vehicle speed, and a lateral force). A first layer of the LSTM comprises, for example, 32 units, and a second layer, for example, 16 units.units), followed by a dense "fully connected" layer with, for example, 8 neurons as the output layer with one output neuron (input > LSTM (32) > LSTM (16) > Dense (1)). However, smaller LSTM networks are also possible (e.g., with only 2 x 16 units), for example, if in-vehicle resources (computing power and / or memory) need to be conserved. 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.

[0029] During a training phase, the machine learning model is trained. This is done using training data in a training dataset based on the data collection, where the training data 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 acquired at the steering wheel and / or elsewhere, in particular time series of torques acquired 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 via backward projection, the parameters of the machine learning model (especially the weights) are adjusted until a calculated error of the output to the baseline is below a predetermined threshold.

[0030] Before training, the data collection is divided, in particular, into at least one training data set, which is used to train the machine learning model, and a separate test data set, which is used to test the trained machine learning model according to the procedure.

[0031] 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").

[0032] It may be possible to divide the measurement data into data snippets of equal duration. This division is achieved, in particular, by decomposing the measurement data into smaller time intervals (time series) of a predetermined duration (e.g., a few seconds). Specifically, it is intended that the time base of various control parameters contained within a single measurement data set, and, if applicable, the at least one additional sensor information, is aligned, so that a common and synchronized time base is subsequently available for all control parameters. In this case, the procedure is carried out using the data snippets.

[0033] The procedure can also be carried out as a computer-implemented procedure.

[0034] The at least one additional sensor piece of information was acquired, in particular, by means of additional sensors installed in the vehicles for this purpose. Vehicles that provide the at least one additional sensor piece of information include, in particular, corresponding additional sensors.

[0035] 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.

[0036] It may be stipulated that the distance threshold is chosen based on an averaged centroidal distance of all measurement data within a cluster. The centroidal distance here refers specifically to the distance of a measurement data point from the centroid, i.e., the center of the cluster. For example, the distance threshold may be chosen such that it corresponds to an averaged centroidal distance of all measurement data points within a cluster. The averaged centroidal distance here specifically refers to the mean value of all distances of the measurement data points from the centroid.

[0037] In one embodiment, the distance threshold is determined based on the distance of a predefined percentile of the measurement data assigned to a given cluster. This allows those measurement data from a cluster to be selected for enrichment that are closest to (or within the respective percentile of) a centroid, i.e., the center point of the cluster. For example, the distance threshold can be chosen to correspond to the 10th percentile (1st decile) with respect to the distance of the measurement data from the centroid.

[0038] In one embodiment, measurement data from the data collection or measurement data from a further data collection containing at least one additional sensor information are assigned to specific first clusters and specific second clusters. A respective distance threshold of a centroidal distance for assignment to the respective second clusters is defined such that the number of measurement data points not also assigned to the extracted first clusters (i.e., in particular, measurement data points assigned only to second clusters but not to the extracted first clusters) is minimized during assignment to the second clusters. This allows the distance threshold for measurement data without additional sensor information, or for all measurement data, to be selected based on a centroidal distance of those measurement data points with additional sensor information. This, in particular, allows the determination of...Optimizing a respective distance threshold for assigning data to the second clusters minimizes the number of measurement data points where at least one additional piece of information (i.e., vertical acceleration) is not dominant. The distance thresholds determined in this way for the second clusters are then used, in particular, when assigning the measurement data from the further data collection, excluding at least one additional sensor information, to the designated second clusters.

[0039] In one embodiment, the at least one additional sensor information component includes at least one vertical motion information component of the vehicle. This allows the method to be performed, particularly with regard to vertical excitations from the road, so that the machine learning model is robusted, especially with regard to these vertical excitations. The vertical motion information can, for example, include one or more of the following parameters: respective body accelerations, respective suspension travels, and / or rotation angles (pitch angle, yaw angle, roll angle).

[0040] In one embodiment, the machine learning model comprises an ensemble of several machine learning models. When selecting the misclassified measurement data from the test dataset, those misclassified measurement data are considered where the outputs of the multiple machine learning models match or where a predetermined proportion of the outputs of the multiple machine learning models show a match. Using an ensemble allows the machine learning model to be designed more robustly with regard to performance. The goal here is to identify systematic misclassifications, i.e., data points (measurement data) that are incorrectly classified despite the different learning methods of the multiple machine learning models. If the variance of the outputs of the multiple machine learning models within the ensemble is large (e.g., 50 / 50 correct / incorrect), this is not considered systematic behavior.However, if a large majority (e.g., > 90%) of the multiple machine learning models misclassify data, this may indicate systematic behavior. A single model might select data points that would be correctly classified in another model—thus not representing a systematic error. An ensemble model, however, attempts to avoid this and therefore identify the most problematic cases.

[0041] In one embodiment, dominant features are determined, at least in part, based on differences between the feature means of the first clusters. If the mean of a feature in one cluster deviates significantly from the mean of the feature in the other clusters, it is likely that this feature is particularly important for defining the boundaries of that cluster.In a simple example, if the two feature dimensions of velocity and vertical acceleration (as additional sensor information) are used, and two clusters are found, both showing a large variance in velocity but a much smaller variance in vertical acceleration, and exhibiting one small and one significantly larger mean value for vertical acceleration, then vertical acceleration can be determined as the dominant feature in the cluster with the significantly larger mean value. In contrast, the vertical acceleration of the other cluster, due to its significantly smaller mean value, is not suitable as the dominant feature. In other words, the cluster with the small mean value can be discarded based on expert knowledge, particularly because the excitation from the road was only weak.

[0042] In one embodiment, it is provided that dominant features are determined, at least partially, based on the respective variance of the features within the first clusters. This provides a simple measure for determining the dominant features. In particular, it is provided that features with lower variance within the clusters, but higher variance between the clusters, are considered more relevant or dominant. These two factors can, for example, be weighted to determine the dominant features.

[0043] In one embodiment, dominant features are determined, at least partially, based on a feature-dependent silhouette coefficient of the respective first clusters. This provides a meaningful measure for determining the dominant features. The silhouette coefficient quantifies how well a data point (measurement data) fits its assigned (first) cluster. To determine the relevance of a feature (for clustering), the silhouette coefficient of a cluster with respect to a specific feature can be calculated. If the feature is omitted or added and this significantly changes the value of the silhouette coefficient (i.e., above a predefined threshold), it can be assumed that this feature is more relevant than another feature whose value changes less when omitted.It may be provided that a predetermined number of those characteristics which show the greatest change in the coefficient are taken into account.

[0044] In one embodiment, dominant features are determined at least partially based on a weighted distance to the respective cluster centroids. The K-means method minimizes the quadratic distance to the cluster centroids. Features that contribute more to the distance have a greater influence than features that contribute less and are therefore more relevant. In particular, the influence of each feature dimension on the cluster assignment can be adjusted. In the example described above, where velocity and vertical acceleration (or at least one additional sensor piece of information) are considered feature dimensions in the clustering process, vertical acceleration (or at least one additional sensor piece of information) can, for instance, be weighted more heavily than velocity.Changing the weighting allows the distance to the cluster, up to which a data point is assigned to the cluster, to be varied in the respective feature dimension. In other words, with two feature dimensions, the boundary for cluster assignment is typically a circle if the dimensions are equally weighted. With differently weighted feature dimensions, however, the boundary is typically an ellipse.

[0045] In one embodiment, dominant features are determined, at least partially, based on principal component analysis. This allows the significance of the features to be determined based on the respective variance in the principal components. Those principal components that contribute most to cluster separation, i.e., those with the lowest variance, contain the most important features.

[0046] In one embodiment, dominant features are determined using a hierarchical clustering method. This hierarchical clustering method, for example using the Iterative Dichotomizer (ID3) algorithm, yields a feature importance for each feature dimension.

[0047] For example, the separation is first made via feature dimension A instead of feature dimension B, etc. This allows dominant features or feature dimensions, or a dominance of at least one additional sensor information (especially vertical acceleration), to be identified.

[0048] Further features for the design of the device result from 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.

[0049] Furthermore, a method for detecting a hands-off state on a vehicle's steering wheel 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.

[0050] Furthermore, in particular, a device for detecting a hands-off state on a vehicle's steering wheel 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.

[0051] The invention is explained in more detail below with reference to preferred embodiments and the figures. These show:

[0052] Fig. 1 shows a schematic flowchart of embodiments of the method for providing a trained machine learning model for detecting a hands-off state on a vehicle's steering wheel;

[0053] Figs. 2a to 2c are schematic representations to illustrate embodiments of the method for providing a trained machine learning model for detecting a hands-off state on a vehicle's steering wheel;

[0054] Fig. 3 shows a schematic representation of an embodiment of the device for providing a trained machine learning model for detecting a hands-off state on a vehicle's steering wheel;

[0055] Fig. 4 shows a schematic flowchart of an embodiment of the method for detecting a hands-off state in a vehicle;

[0056] Fig. 5 shows a schematic representation of an embodiment of the device for detecting a hands-off state in a vehicle.

[0057] Fig. 1 shows a schematic flowchart of embodiments of the method for providing a trained machine learning model for detecting a hands-off state on a vehicle's steering wheel.

[0058] In a process step 100, a data collection of recorded measurement data is obtained and / or a data collection of measurement data is recorded, each comprising at least one control parameter and one basic truth concerning a hands-off state, wherein a portion of the measurement data in the data collection additionally includes at least one additional sensor information.

[0059] In process step 101, a machine learning model already trained with a training data set compiled from the data collection, neglecting at least one additional sensor information, is obtained and / or provided.

[0060] In process step 102, the trained machine learning model is applied to a test data set selected from the data collection.

[0061] In process step 103, misclassified measurement data are selected from the test data set during application.

[0062] In process step 104, those misclassified measurement data that include at least one additional sensor piece of information are identified. This can be done, for example, by the presence of at least one additional sensor piece of information in the measurement data and / or by identification information that marks the measurement data for easy and quick identification.

[0063] In process step 105, initial clusters are determined for the misclassified measurement data. This is done using well-known clustering methods, such as the k-means method.

[0064] In process step 106, dominant features in the identified initial clusters are determined. This is done in particular by evaluating the clusters or by assessing the relevance of the respective features for the cluster or for separating the respective cluster from the other clusters.

[0065] In a process step 107, those first clusters are extracted for which features of at least one additional sensor information form dominant features.

[0066] In process step 108, second clusters are determined based on the at least one control variable for the misclassified measurement data contained in the extracted first clusters. The at least one additional sensor information is specifically disregarded in this step. This is done using known clustering methods, such as the k-means method. In process step 109, measurement data from a further data set, excluding the at least one additional sensor information, are assigned to the determined second clusters. This further data set is structurally identical to the first data set in terms of its properties; only the values ​​of the measurement data differ. Furthermore, the measurement data in this further data set does not include any additional sensor information.

[0067] In process step 110, the training data set is enriched with those associated measurement data from the further data collection which have a distance to the second clusters that is below a specified distance threshold.

[0068] In process step 111, the machine learning model is (re)trained using the enriched training data set.

[0069] In process step 112, the trained machine learning model is provided.

[0070] Figures 2a to 2c show schematic representations illustrating embodiments of the method for providing a trained machine learning model 10 for detecting a hands-off state 41 on a vehicle's steering wheel. Figures 2a to 2c also refer in particular to the reference numerals of the process steps in the flowchart shown in Figure 1.

[0071] The starting point is a data collection 20 of recorded measurement data, each comprising at least one steering parameter 40 and a basic truth 43 concerning a hands-off state 41. The measurement data 11 of the data collection 20 were partly recorded by vehicles 50 of a vehicle fleet 60 which have no additional sensors, i.e., which provide measurement data 11 without state sensor information 42, and partly by vehicles 50+ of a vehicle fleet 60+ which have additional sensors (in particular for recording vertical excitations), i.e., which provide measurement data 11 with state sensor information 42.

[0072] It may be provided that the at least one additional sensor information 42 includes at least one vertical motion information of the vehicle 50+. This can, for example, include one or more body accelerations, one or more suspension travels, and / or one or more rotation angles. A machine learning model 10, already trained with a training dataset 30 compiled from the data collection 20 while neglecting the at least one additional sensor information 42, is applied to a test dataset 31 selected from the data collection 20. The respective output of the machine learning model 10 is compared with the respective basic truth 43 (hands-off state or no hands-off state).

[0073] The measurement data 11 from the test data set 31 that were misclassified during application are selected (procedure step 103) and collected in an intermediate storage 21.

[0074] In the example shown in Fig. 2a, the trained machine learning model 10 is an ensemble of several machine learning models. In this case, when selecting the measurement data 11 that were misclassified during application from the test data set 31, those misclassified measurement data 11 are taken into account where the outputs of the several machine learning models match or where a predefined proportion of the outputs of the several machine learning models show a match.

[0075] Those misclassified measurement data 11 that include at least one additional sensor information 42 are identified (Fig. 2b). This can be done, for example, by identifying those measurement data 11 to which a vehicle identification information ID is assigned that is associated with the vehicle fleet 60+ with additional sensors. Alternatively, it can also be checked directly whether additional sensor information 42 is present in the measurement data 11.

[0076] Initial clusters C1 are determined for the misclassified measurement data 11 stored in the intermediate memory 21. Both measurement data 11 with at least one additional sensor information 42 and measurement data 11 without at least one additional sensor information 42 are taken into account.

[0077] In the initial clusters C1, dominant characteristics 22 are determined (procedure step 106). For this purpose, the characteristics are evaluated, in particular with regard to their respective relevance for the division into clusters C1.

[0078] First clusters C1 are extracted for which features of the at least one additional sensor information 42 form dominant features 22 (process step 107). Second clusters C2 are determined based on the at least one control parameter 40 for the misclassified measurement data 11 contained in the extracted first clusters C1 (process step 108). In other words, for those measurement data 11 (with the at least one additional sensor information 42) contained in the extracted first clusters C1, clusters are again determined, which are designated as second clusters C2, whereby only the at least one control parameter is considered during clustering, but not the respective at least one additional sensor information 42.

[0079] Subsequently, measurement data 11 of a further data collection 23 (Fig. 2c) without at least one additional sensor information 42 are assigned to the determined second clusters C2 (procedure step 109).

[0080] The training data set 30 (Fig. 2a), with which the machine learning model 10 was trained, is enriched with those associated measurement data 11 from the further data collection 23 (procedure step 110) which have a distance to the second clusters C2 that is below a specified distance threshold D, so that an enriched training data set 30+ is available.

[0081] The machine learning model 10 (Fig. 2a) is (re)trained using the enriched training data set 30+ and the (re)trained machine learning model 10 is provided.

[0082] It may be provided in a process step 109a that the distance threshold D is determined starting from the clusters C1, C2.

[0083] In process step 109a, it may be provided that the distance threshold D is determined based on a distance of a predefined percentile P of the measurement data 11 assigned to a respective cluster C1, C2. All measurement data 11 from the further data collection 23 that are assigned to a second cluster C2 within this distance are taken into account during enrichment.

[0084] In process step 109a, it can alternatively be provided that measurement data 11 from data collection 20 or measurement data 11 from another data collection with at least one additional sensor information 42 are assigned to the specified first clusters C1 (process step 109b-1) and the specified second clusters C2 (process step 109b-2), whereby a respective distance threshold D of a centroid distance for assignment to the respective second clusters C2 is set such that the number of measurement data 11 that were not also assigned to the extracted first clusters C1 is minimized when assigning them to the second clusters C2. In a process step 109b-3, measurement data in the second clusters C2 that were also assigned to the extracted first clusters C1 are specifically sought. In process step 109a, a process step 109a-x then provides that the centroid distance is set according to the above specification.

[0085] It may be provided that the machine learning model 10 comprises an ensemble of several machine learning models, whereby when selecting the misclassified measurement data 11 from the test data set 31, those misclassified measurement data 11 are taken into account where the outputs of the several machine learning models match or a predetermined proportion of the outputs of the several machine learning models show a match (prediction overlap).

[0086] It may be provided in process step 106 that dominant features 22 are determined at least partially based on differences between feature means of the first clusters C1.

[0087] It may be provided in process step 106 that dominant features 22 are determined at least partially based on a respective variance of the features within the first cluster C1.

[0088] In process step 106, it may be provided that dominant features 22 are determined at least partially based on a feature-dependent silhouette coefficient of the respective first clusters C1.

[0089] In process step 106, it may be provided that dominant features 22 are determined at least partially based on a weighted distance to the respective cluster centroids of the first cluster C1.

[0090] In process step 106, it may be provided that dominant features 22 are determined at least partially based on a principal component analysis of the respective first clusters C1. Fig. 3 shows a schematic representation of an embodiment of the device 1 for providing a trained machine learning model 10 for detecting a hands-off state on a vehicle's steering wheel.

[0091] 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 one of the embodiments shown in Fig. 1.

[0092] The data processing unit 2 is configured to perform the following steps: Receiving 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 and one basic truth regarding a hands-off state, wherein a portion of the measurement data in the data set 20 additionally includes at least one additional sensor information; Receiving and / or providing a machine learning model 10 already trained with a training data set compiled from the data set 20, neglecting the at least one additional sensor information; Applying the trained machine learning model 10 to a test data set selected from the data set 20; Selecting measurement data from the test data set that were misclassified during the application;

[0093] Identifying those misclassified measurement data that include at least one additional sensor information; determining initial clusters for the misclassified measurement data; determining dominant features in the determined initial clusters;

[0094] Extracting those first clusters for which features of the at least one additional sensor information form dominant features; determining second clusters based on the at least one control variable for the misclassified measurement data contained in the extracted first clusters; assigning measurement data from a further data set 23 without the at least one additional sensor information to the determined second clusters;

[0095] Enriching the training dataset with those associated measurement data from the further data collection 23 which have a distance to the second clusters that is below a specified distance threshold, training the machine learning model 10 using the enriched training dataset, providing the trained machine learning model 10.

[0096] The provisioning can in particular include loading the trained machine learning model 10 into a memory of a control unit 53. Fig. 4 shows a schematic flowchart of an embodiment of the method for detecting a hands-off state on a steering wheel of a vehicle.

[0097] In process step 200, at least one steering parameter is recorded, for example a steering wheel torque, a steering wheel angle and / or a steering wheel angular velocity.

[0098] In process step 201, a hands-off state is determined using the trained machine learning model based on the input of at least one control variable. 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 occurring.

[0099] 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.

[0100] Figure 5 shows a schematic representation of an embodiment of the device 51 for detecting a hands-off state 41 on the steering wheel of a vehicle 50, in particular a motor vehicle. The device 51 is, in particular, arranged in the vehicle 50.

[0101] 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.

[0102] 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 10 trained according to one of the embodiments described in this disclosure, 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 10 based on the supplied at least one steering parameter 40, and to output the determined hands-off state 41.

[0103] 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

[0104] device

[0105] Data processing facility

[0106] -1 Computing unit

[0107] -2 memory

[0108] 0 Machine learning model

[0109] 1. Measurement data

[0110] 0 Data collection

[0111] 1 intermediate storage

[0112] 2 dominant traits

[0113] 3 additional data sets (without additional sensor information) 4 additional data sets (with additional sensor information) 0 training data sets

[0114] 0+ enriched training data set

[0115] 1 test data set

[0116] 0 Steering size

[0117] 1 Hands-Off State

[0118] 2 Additional sensor information

[0119] 3 Fundamental Truths

[0120] 0 vehicles

[0121] Vehicle 0+ (with additional sensors)

[0122] 1 Device

[0123] 2 Sensor

[0124] 3 Control unit

[0125] 4 Driver assistance systems

[0126] 5 Vehicle control

[0127] 0 vehicle fleet

[0128] 0+ vehicle fleet (with additional sensors)

[0129] 00-112 Procedure steps

[0130] 00-202 Procedure steps

[0131] 1 first cluster

[0132] 2 second ClusterD distance threshold

[0133] Vehicle identification information (VIN)

Claims

- 26 - Patent claims 1. Method for providing a trained machine learning model (10) for detecting a hands-off state (41) on a steering wheel of a vehicle (50), comprising: receiving a data set (20) of acquired measurement data (11) and / or acquiring a data set (20) of measurement data (11), each comprising at least one steering parameter (40) and one basic truth (43) concerning a hands-off state (41), wherein a portion of the measurement data (11) of the data set (20) additionally comprises at least one additional sensor information (42), Obtaining and / or providing a training data set (30) already compiled from the data collection (20) neglecting at least one additional sensor information (42), applying the trained machine learning model (10) to a test data set (31) selected from the data collection (20), Selecting misclassified measurement data (11) from the test data set (31) when applying the test data set (31), Identifying those misclassified measurement data (11) that include at least one additional sensor information (42), Determining first clusters (C1) for the misclassified measurement data (11), determining dominant features (22) in the determined first clusters (C1), extracting those first clusters (C1) for which features of at least one additional sensor information (42) form dominant features (22), Determining second clusters (C2) based on the at least one control variable (40) for the misclassified measurement data (11) contained in the extracted first clusters (C1), Assigning measurement data (11) from another data set (23) without at least one additional sensor information (42) to the specified second clusters (C2), enriching the training data set (10) with those assigned measurement data (11) from the other data set (23) which have a distance to the second clusters (C2) that is below a specified distance threshold (D), training the machine learning model (10) using the enriched training data set (30+), Provision of the trained machine learning model (10).

2. Method according to claim 1 , characterized in that the distance threshold (D) is determined based on a distance of a predetermined percentile (P) of the measurement data assigned to a respective cluster (C2).

3. Method according to claim 1, characterized in that measurement data (11) of the data collection (20) or measurement data (11) of a further data collection with the at least one additional sensor information (42) are assigned to the specific first clusters (C1) and the specific second clusters (C2), wherein a respective distance threshold (D) of a centroid distance for an assignment to the respective second clusters (C2) is set such that a number of measurement data (11) that were not also assigned to the extracted first clusters (C1) is minimized when assigning to the second clusters (C2).

4. Method according to one of the preceding claims, characterized in that the at least one additional sensor information (42) comprises at least one vertical movement information of the vehicle (50).

5. Method according to one of the preceding claims, characterized in that the machine learning model (10) comprises an ensemble of several machine learning models, wherein when selecting the measurement data (11) misclassified during application from the test data set (31), those misclassified measurement data (11) are taken into account where the outputs of the several machine learning models match or a predetermined proportion of the outputs of the several machine learning models show a match.

6. Method according to one of the preceding claims, characterized in that dominant features (22) are determined at least partially based on differences between feature means of the first clusters (C 1 ).

7. Method according to one of the preceding claims, characterized in that dominant features (22) are determined at least partially based on a respective variance of the features within the first clusters (C 1 ).

8. A method according to any of the preceding claims, characterized in that dominant features (22) are determined at least partially based on a feature-dependent silhouette coefficient of the respective first clusters (C1).

9. A method according to any of the preceding claims, characterized in that dominant features (22) are determined at least partially based on a weighted distance to the respective cluster centroids.

10. Method according to one of the preceding claims, characterized in that dominant features (22) are determined at least partially based on a principal component analysis.

11. Device (1) for providing a trained machine learning model (10) for detecting a hands-off state (41) on a steering wheel of a vehicle (50), comprising: a data processing unit (2) wherein the data processing unit (2) is configured to perform the following steps: Receiving a data set (20) of recorded measurement data (11) and / or recording a data set (20) of measurement data (11), each comprising at least one steering parameter (40) and one basic truth (43) concerning a hands-off state (41), wherein a portion of the measurement data (11) of the data set (20) additionally comprises at least one additional sensor information (42), Obtaining and / or providing a training data set (30) already compiled from the data collection (20) neglecting at least one additional sensor information (42), applying the trained machine learning model (10) to a test data set (31) selected from the data collection (20), Selecting misclassified measurement data (11) from the test data set (31) when applying the test data set (31), Identifying those misclassified measurement data (11) that include at least one additional sensor information (42), Determining first clusters (C1) for the misclassified measurement data (11), determining dominant features (22) in the determined first clusters (C1), extracting those first clusters (C1) for which features of at least one additional sensor information (42) form dominant features (22), Determining second clusters (C2) based on the at least one control variable (40) for the misclassified measurement data (11) contained in the extracted first clusters (C1), Assigning measurement data (11) to another data collection (23) without the minimum - 29 - an additional sensor information (42) for the specified second clusters (C2), enriching the training data set (30) with those associated measurement data (11) of the further data collection (23) which have a distance to the second clusters (C2) that is below a specified distance threshold (D), training the machine learning model (10) using the enriched training data set (30+), Deploying the trained machine learning model (10).

12. Method for detecting a hands-off state (41) on a steering wheel of a vehicle (50), comprising: Capture at least one steering variable (40), Feeding the detected at least one steering parameter (40) to a machine learning model (10) trained according to one of claims 1 to 10 as input data, determining the hands-off state (41) using the trained machine learning model (10) based on the supplied at least one steering parameter (40), Issuing the specified hands-off state (41).

13. Device (51) for detecting a hands-off state (41) on a steering wheel of a vehicle (50), comprising: at least one sensor (42) 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 (10) trained according to one of claims 1 to 10, and to supply the detected at least one steering variable (40) to the trained machine learning model (10) as input data, to determine the hands-off state (41) using the trained machine learning model (10) based on the supplied at least one control variable (40), and to output the determined hands-off state (41).