Machine learning method for operating a vehicle component and method for operating a vehicle component

Abstract

The invention relates to a computer-implemented method for operating a vehicle component for machine learning, in particular for processing or generating a training data set, as well as a corresponding method and a control unit for operating a vehicle component. In a first step, the provided multidimensional data points are divided by dividing a plurality of data points into multidimensional clusters using a clustering algorithm. Then, a training data set is generated by selecting data points from a basic training data set. In this case, the selection includes: determining the smallest cluster with the least number of data points among the plurality of clusters. At least one subset of the data points of the smallest cluster is also provided for the training data set. In a further step, a subset of data points is selected for the training data set from each of the remaining clusters, wherein the number of selected data points of each remaining cluster corresponds to the number of selected data points of the smallest cluster.

Description

Technical Field

[0001] The invention relates to: a computer-implemented method for operating a vehicle component for machine learning, in particular for processing or generating a training data set; and a corresponding method for operating a vehicle component. A corresponding control unit and a corresponding vehicle are also provided. Background Art

[0002] Training with prior art machine learning methods has been shown to be inefficient with unevenly distributed datasets with over- or under-represented scenarios. This can result in poor performance with previously unknown data. This in turn results in the robustness of the method and applicability in safety-critical environments being significantly compromised.

[0003] Just as another technical background, we refer to the prior art, which deals with another, very special problem in the case of very extreme uneven distribution between the so-called minority class and the majority class in the output data set. In the publication "Cluster-based under-sampling approaches for imbalanced data distributions" by S. Yen et al., Expert System with Applications 36, 2009, pp. 5718-5727, doi: 10.1016 / j.eswa.2008.06.108, the weighted distribution of these rare minority classes is described. In the publication "Cluster based under-sampling for unbalanced cardiovascular data" by M. Mostafizur Rahman et al., Proceedings of the World Congress on Engineering, 2013, Vol. III, ISSN: 2078-0966 (Online), this special special case of rare minority classes is also handled. Here, the input dataset is divided into majority and minority classes, where clustering is done only with respect to the majority class, and then multiple datasets are achieved by pairing each cluster with these minority classes. The best pair of these combinations is then used as a training dataset.

[0004] In addition to these special problems, methods are implemented in the prior art in which the training performance is tested and then verified based on independent data. In this case, very large training data sets are usually used that depict the target behavior.

[0005] For large datasets, high computational costs are also incurred due to the large datasets, dataset processing such as labeling, anonymization, and increased computing resources must be provided to handle these large data volumes for the training process. Iterative, manual dataset generation is error-prone and often slow due to trial and error. Considering all data also leads to very time-consuming and ineffective training.

[0006] Furthermore, the analysis of fault cases is often based on expert knowledge. Here, the difficulties mainly lie in the coupling of different causes and the complex detection of parameters. The solution is mostly to use more data in order to achieve better performance. No objective analysis of the multidimensional space is performed. In order to expand or increase the data set, the analysis is performed by experts.

[0007] The automotive sector in particular is a safety-critical sector in which various vehicle components must be operated with high safety requirements. High performance and fast training times are particularly important in this sector. When operating vehicle components with the support of machine learning algorithms, the training data set and performance are sometimes decisive in whether serious errors occur or not. Summary of the invention

[0008] The object of the present invention is now to provide a computer-implemented machine learning method for operating a vehicle component and a corresponding method for operating a vehicle component, which solves at least some of the above-mentioned problems or at least reduces at least some of the above-mentioned problems to a minimum. One aspect of the invention is in particular to ensure good performance across all scenarios and at the same time to allow the generation of training data sets in a time-saving and resource-saving manner. In addition, the aim is to achieve the prerequisites for time-efficient and high-precision training.

[0009] In a preferred embodiment of the present invention, a computer-implemented method for machine learning, in particular for processing or generating a training data set, for operating a vehicle component is provided. In a first step, the method comprises: providing a plurality of multidimensional data points. In a further step, the method comprises: generating a basic training data set by dividing the plurality of data points into multidimensional clusters using a clustering algorithm, wherein each data point is assigned to exactly one cluster. In a further step, the training data set is generated by selecting data points from the basic training data set. In this case, the generation of the training data set comprises: determining the smallest cluster with the least number of data points among the plurality of clusters. In a further step, the generation of the training data set comprises: selecting at least a subset of the data points of the smallest cluster for the training data set. Furthermore, in a further step, a subset of data points is selected for the training data set from each of the remaining clusters, so that the number of selected data points of each remaining cluster corresponds to the number of selected data points of the smallest cluster.

[0010] Multidimensional data points can be represented by vectors and equipped with multiple entries. The collection of data points can form a time series. For example, multidimensional data points may include physical or kinematic variables, such as vehicle speed, position, direction, steering torque, lateral acceleration, yaw rate, steering wheel angular velocity, acceleration, or trajectory or trajectory segment, wherein the above examples are only exemplary enumerations and the present invention is not limited thereto. In principle, data points are measured data points related to the vehicle, such as sensor data points, that is, parameters that can be detected by sensors. For example, tire pressure can also be an entry of a data point. Data points can be sensor signals, for example. In other cases, these data points are simulated data or data provided by a server, which are obtained from a measurement process that occurred previously. The sensor signal can be detected by multiple sensors of the vehicle, especially by different sensors of the vehicle. Preferably, the k-means algorithm can be used as a clustering algorithm. However, the present invention does not depend on the clustering algorithm used. For example, EM clustering or fuzzy c-means algorithm can also be used. Each cluster is a specific scene. In addition, a scene is a clearly separable subset of the entire data set. In this case, the scenarios can be differentiated by means of a multidimensional input to a clustering algorithm. Clustering is used to automatically extract important areas of the entire unstructured database. Clustering algorithms that uniquely assign each data point to a cluster are also referred to as hard methods. Preferably, the data are physical and / or kinematic data that are necessary for determining target variables or setpoint variables for operating vehicle components.

[0011] The present invention has the advantage that: the automated uniform distribution of training data sets across different scenarios is generated. This is achieved by obtaining exactly the same number of data points from the remaining clusters for the training data set. It is advantageous to move toward the smallest cluster because the smallest cluster is a scenario that is underrepresented due to the small number of data points compared to other clusters. In the new training data set, the previously underrepresented clusters are weighted or represented in the same way as other clusters. Therefore, each cluster and thereby each scenario is advantageously considered equally based on and with respect to the number of data points of the clusters generated by the clustering algorithm through the uniform distribution generated by the clusters. By reducing the data points in a targeted manner, it is also possible to significantly save training time and also provide a data pool for test and validation data. Evenly distributed data sets also have better performance. Here and below, an objective method for considering multidimensional analysis of data is provided, which is subjectively almost impossible to grasp due to complexity and multidimensionality.

[0012] Other preferred embodiments of the present invention can be found in the following description.

[0013] Preferably, the method comprises storing the unselected data points of the smallest cluster and / or the remaining clusters in a data storage as test and / or validation data. Thus, these unused data points are a data pool that can be used to test and validate the machine learning algorithm and its performance. These data points can also be used to improve the training data set with respect to clustering performance, as further described below. If the data points in the smallest cluster are also stored, the performance of the smallest cluster can also be checked.

[0014] In a particular embodiment, the method includes: selecting the entirety of all data points of the smallest cluster for the training data set. In principle, the smallest cluster of the base data set is under-representative due to the minimum number of data points. Thus, all or nearly all data points of the cluster are usually required to scan the cluster in order to fully define or characterize the cluster.

[0015] Preferably, the method comprises: randomly selecting data points for the training data set from the smallest cluster and / or the remaining clusters. This is a particularly fast method, whereby a uniform distribution of the training set can be automatically achieved in the shortest possible time. In other embodiments, the selection can also be made based on a time criterion.

[0016] In a special embodiment, the method comprises: selecting data points from the smallest cluster and / or the remaining clusters such that the selected data points of the clusters have a maximum distance from one another. Redundant data are removed by means of this distance criterion and thus only important data for defining or covering the cluster are retained. In particular, in other words, the greatest possible expansion of the clusters is achieved. The selection of the data points is therefore carried out according to an importance criterion.

[0017] Preferably, the method comprises: training a machine learning algorithm with the aid of the generated training data set. Such training can provide information as to whether the created training data set can or must be improved. In particular, artificial neural networks come into consideration as machine learning algorithms, wherein the present invention is not limited thereto and also applies to other relevant machine learning algorithms. Preferably, supervised training with a predefined target result is applied for the training. In order to achieve an optimal operating point for the machine learning algorithm, test data of unselected data points can be used. With the test data, an overfitting of the machine learning algorithm, also called overfitting, can be avoided.

[0018] In a special embodiment, the method includes determining the performance of the individual clusters using the validation data by means of a quality criterion. Thus, by determining the performance of the individual clusters, i.e., for example, the quality or accuracy of the respective individual clusters, it can be checked whether the data points used for the respective cluster are sufficient or whether the data points for the cluster must be increased and thus densified.

[0019] Preferably, the performance of the individual clusters is determined by the data points in the group of unselected data points of the belonging cluster. If data points are extracted from a particular cluster as validation data and used for validation, the validation result may in particular provide information about the performance of the belonging cluster. Thus, the previously unselected data points are advantageously sensitive or selective to the clusters from which they were extracted. If the data points are extracted from a minimal cluster, this minimal cluster may also advantageously be checked with respect to its performance.

[0020] Preferably, the quality criterion may be the mean square error of the output values ​​relative to the target value above a threshold. In some cases, a certain number of such threshold exceedances may also be used as a quality criterion. The mean square error and especially the threshold exceedances are sensitive measures of the performance of the individual clusters.

[0021] In a special embodiment, the method comprises creating a modified training data set by adding data points to clusters that do not meet the quality criterion. In this way, individual clusters can be improved point by point or locally, that is to say the statistics or the density of data points of these clusters are improved in order to improve the performance of these clusters point by point or in a targeted manner. The advantage is that the improvement is only carried out in the case of those clusters that do not meet the quality criterion and can therefore achieve an efficient improvement, because the already superior clusters are not changed.

[0022] The added data points are preferably extracted from the group of non-selected data points of the associated cluster that do not meet the quality criteria. As a result, previously non-selected data points can be reused in a targeted manner in order to improve the value of the individual clusters by a denser or higher representation. In particular, due to these non-selected data points, particularly suitable data points are present for each cluster, so that suitable data points are always available in the data pool.

[0023] In a particular embodiment of the invention, the method comprises: repeating the modification of the training data set until the quality criterion is met for all clusters. Thus, the above steps including the performance check can be performed until all clusters meet the quality criterion.

[0024] Thus, based on the basic training dataset, a superior training dataset has been generated, which is superior in terms of quality criteria for depicting each scene or in each cluster. However, by saving redundant data points, significant savings in computing time can be achieved through the above-described systematic, automated creation of the training dataset, as already described in the previous paragraphs.

[0025] Another aspect of the invention comprises a method for operating a vehicle component. The method comprises: operating a vehicle component based on the output of a machine learning algorithm trained according to the method described in one of the above-described embodiments. Here and hereinafter, operation may include control and / or regulation. Operation may also be monitoring, maintenance, alarming, wherein the invention is not limited thereto. Purely by way of example, a vehicle component may be a steering device, a driver assistance system, a braking device, an alarm system, a lighting unit, and the like. However, the invention is not limited here to a specific vehicle component. By using a machine learning algorithm trained in this way, the high accuracy, precision and performance of the machine learning algorithm are directly utilized so as to enable sensitive and safe operation of vehicle components, for example, precisely in safety-critical areas.

[0026] In a preferred embodiment of the present invention, a method for operating a vehicle component is provided. In a first step, the method for operating a vehicle component includes providing a plurality of multidimensional data points. In another step, the method includes generating a basic training data set by dividing the plurality of data points into multidimensional clusters using a clustering algorithm, wherein each data point is assigned to exactly one cluster. In another step, a training data set is generated by selecting data points from the basic training data set. Here, the generation of the training data set includes determining a minimum cluster with the least number of data points among a plurality of clusters. In another step, the generation of the training data set includes selecting at least a subset of the data points of the minimum cluster for the training data set. In addition, in another step, a subset of data points is selected from each of the remaining clusters for the training data set, so that the number of selected data points of each remaining cluster corresponds to the number of selected data points of the minimum cluster. Other method steps can be known from the methods described above and below. The method also includes the step of operating the vehicle component based on the output of a machine learning algorithm trained using the generated training data set. Furthermore, other preferred embodiments of the method can also be combined with the method for operating a vehicle component and are therefore disclosed together. By using a machine learning algorithm trained in this way, the high accuracy, precision and performance of the machine learning algorithm can be utilized, which is important precisely in safety-critical areas.

[0027] Another aspect of the invention comprises a control unit for a vehicle, which is configured to carry out a method for operating a vehicle component or a method according to the above-described embodiments. Advantages result from the above-described embodiments.

[0028] Another aspect of the present invention comprises a vehicle comprising a control unit according to the above-described embodiment.

[0029] Unless otherwise indicated in individual cases, the various embodiments of the invention mentioned in this application can be advantageously combined with one another. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The present invention is then explained in an exemplary embodiment with reference to the accompanying drawings.

[0031] Figures 1 to 6 The steps of a computer-implemented machine learning method for operating a vehicle component according to an embodiment of the present invention and a process for executing the method for operating a vehicle component according to an embodiment of the present invention are shown; and

[0032] Figure 7 A control unit according to the invention is shown schematically. DETAILED DESCRIPTION

[0033] Figures 1 to 6 A computer-implemented method for operating a vehicle component, in particular for processing or generating a training data set, for machine learning, and a method for operating a vehicle component comprising the method are shown. If only one method is mentioned below, both methods are generally included here, unless explicitly stated otherwise.

[0034] exist Figure 1 A first method step is shown in a diagrammatic manner, in which a plurality of multidimensional data points F1, . . . , Fn are provided, for which see Figure 1 . The data points F1, ..., Fn can be measured data points, for example measured or recorded by corresponding sensors. The server can also provide previously measured measured data points. However, in principle, these data points F1, ..., Fn can also be simulated measured data points. Data points, such as data point F1, can be presented as vectors, which have corresponding kinematic or physical entries. Each entry can include a numerical value and a dimension. The present invention also discloses the case where the vector has only one component. The corresponding parameters can be, for example, vehicle speed, steering torque, lateral acceleration, yaw rate, steering wheel angular velocity, tire pressure, or else a driving trajectory or a driving trajectory segment, wherein the present invention is not limited to this. The selection of important entries for a specific application depends on the corresponding vehicle component and how the vehicle component should be operated.

[0035] In another step, see Figure 1 On the right side for illustration, a basic training data set is created. This step is achieved by partitioning a plurality of data points F1, ..., Fn into multidimensional clusters C1, ..., Cm. For this purpose, a clustering algorithm is used which assigns each data point F1, ..., Fn to exactly one cluster C1, ..., Cm. In other words, the data points F1, ..., Fn are grouped into clusters according to the identified scenarios. Thus, each cluster C1, ..., Cm forms a specific scenario. Here, a scenario can be considered as a clearly separable subset of the entire unstructured database of these data points F1, ..., Fn, for which Figure 1 The left side of the image is compared with the right side. This means that the scene can be described and differentiated by the multi-dimensional input of the clustering algorithm. Thus, the important working areas can be automatically extracted.

[0036] In the present example, which is used only for illustration, the data points F1, ..., Fn are distributed or divided into four different clusters C1, C2, C3, C4. Thus, there are four different scenarios purely by way of example. Furthermore, these clusters C1, C2, C3, C4 are disjoint, that is, each data point F1, ..., Fn is assigned exactly one cluster C1, C2, C3, C4. In practice, a number of data points and / or clusters may be processed that exceeds the above example by several orders of magnitude.

[0037] As a pair Figure 1 In addition, Figure 2 Real data are shown to further illustrate the above description. In this case, different clusters are shown, in this specific example clusters 1 to 9, which respectively indicate the data assigned to the clusters. As mentioned above, the assignment is achieved by a clustering algorithm, preferably by a k-means algorithm. In this case, each cluster C1, ..., C9 forms a specific scenario. In the current example, the vehicle speed v is shown as a physical or kinematic variable, but the present invention is not limited to this. Alternatively, depending on which vehicle component should be operated and which parameter is important for this, parameters such as tire pressure, steering torque, steering speed, etc. that have been mentioned before can be used. In addition, only this one component of the vector is shown here, while in the case of multidimensional data points F1, ..., Fn, corresponding multidimensional clusters are formed that cannot be presented graphically. It can also be seen that: Figure 1 As illustrated in , the number of data points varies from cluster to cluster. This means that some clusters are statistically over- or under-represented relative to other clusters.

[0038] Therefore, many data points in the smallest cluster, here for example cluster 9, are scenes that are underrepresented compared to other subsets (clusters) in the entire data set.

[0039] Redundancies, ie essentially multiple descriptions of data points, can be identified in the case of larger clusters, for example, see in particular clusters 3 and 7. Such redundancies are reduced or eliminated by the following steps.

[0040] Figure 3 In a diagrammatic manner, a method step is shown in which a training data set is created by selecting data points from a basic training data set. In this case, the method comprises: determining a smallest cluster. In this merely diagrammatic illustration, the smallest cluster corresponds by way of example to cluster C2 of a plurality of clusters C1, C2, C3, C4. This smallest cluster comprises the smallest number of data points. Figure 2 , for example, cluster 9 corresponds to the smallest cluster in the group of clusters 1 to 9.

[0041] In a further step, at least a subset of the data points T1, T2, T3, ..., Tk of the smallest cluster C2 is selected for the training data set, for which purpose see in a graphical manner Figure 3 In the present case, a part or subset of the data points of the smallest cluster C2 is selected for the training data set, here 90% is exemplary. This has the following advantages: the complementary part, in the present example 10%, can be stored in a corresponding data memory and used as test and / or validation data for the machine learning algorithm, for which see the description below.

[0042] For example, a subset of the data points T1, T2, T3, ..., Tk of the smallest cluster can be selected randomly, i.e., according to a random process. This is a particularly fast and time-saving selection process. However, the selection can preferably be made according to an importance measure. For example, data points T1, ..., Tk can be selected from the smallest cluster so that the selected data points T1, ..., Tk of the smallest cluster have the maximum distance from each other. As a result, the redundancy in the smallest cluster is effectively eliminated. In addition, the data points T1, ..., Tk selected according to this criterion geometrically form the multidimensional cluster appropriately in terms of the extension of the multidimensional cluster, which in turn improves the performance of the machine learning algorithm in the application. In other words, the data points T1, ..., Tk selected according to this criterion will cause the largest possible expansion of the cluster, so that the cluster is effectively defined without redundancy or at least with only a small redundancy.

[0043] In another embodiment, the entirety of all data points of the smallest cluster may also be selected for the training data set, which is not explicitly shown here. This is particularly advantageous when the number of these data points is such that each of these data points T1, ..., Tk is ultimately required for sufficient clustering performance of the smallest cluster, as described below.

[0044] In a further step, a subset of data points R1, ..., Rk is selected for the training data set from each of the remaining clusters C1, C3, C4. Here, the selection is performed so that each of the remaining clusters, in Figure 3 The number of selected data points R1, ..., Rk of the clusters C1, C3, C4 corresponds to the number of selected data points T1, ..., Tk of the smallest cluster C2. Therefore, the number of data points of each cluster C1, ..., Cm is identical to one another and therefore in particular to the number of data points T1, ..., Tk selected from the smallest cluster. Therefore, the number of selected data points of the smallest cluster identically defines the number of selected data points R1, ..., Rk of each of the remaining clusters.

[0045] The invention thus has the following advantageous properties: a uniformly distributed, automatically generated training data set is generated across different scenarios, which improves performance and is therefore also important in particular when operating vehicle components. A direction towards the smallest cluster is achieved, knowing that the scenario is underrepresented compared to other clusters due to the small number of data points. In the generated training data set, the previously underrepresented cluster is represented in the same way as the other clusters. Here, the arrangement is made solely based on the number of corresponding data points.

[0046] Reducing the data points in the remaining clusters also leads to a reduction in training time. In the case of large clusters, this reduction can be significant. In this case, depending on the cluster size, up to 40 times the computing time can be achieved according to field studies. This is also important for the operation of vehicle components.

[0047] Similar to the minimum cluster, the following process can be performed using the unselected data points. The unselected data points of the remaining clusters can be stored in a data memory. Then, these data points can also be used as test data and / or validation data. In addition to the uniform distribution of the scenes, a high amount of test and validation data is also provided by this selection.

[0048] Data points can also be randomly selected from the remaining clusters C1, ..., Ci-1, Ci+1, ..., Cm in order to save subsequent calculation time. Preferably, as described above, the data points R1, ..., Rk of the remaining clusters C1, ..., Ci-1, Ci+1, ..., Cm are selected so that the selected data points of the clusters have a maximum distance from each other. In this way, in particular redundant data points are removed and important data points that define the clusters are added to the training data set.

[0049] Preferably, using the test data, the unselected vehicle data can be used with the training data set to avoid overfitting or to find an optimal operating point in order to train a machine learning algorithm, in particular an artificial neural network.

[0050] Figures 4 to 6 Further steps of a method for improving a training data set compiled to operate a vehicle component are shown.

[0051] For this purpose, in a first step, a machine learning algorithm, for example an artificial neural network, can be trained with the aid of the generated training data set. As already described above, test data can be used during the training. These test data are used to find the optimal operating point during the training and to prevent overfitting of the neural network. Now, an advantageous subset of the group of unselected data points can be used as test data. In this case, the invention has the advantage that, through the above-described arrangement of the training data set, such test data are available in large quantities, so that the optimal time for terminating the training can be clearly defined.

[0052] Then, with the aid of the validation data, the performance, that is to say the functional capacity, of the individual clusters C1, ..., Cm can be advantageously tested, for example in Figure 4 As exemplarily shown in .

[0053] To this end, Figure 4 The results of the performance test of the first cluster C1 are shown in Figure 4 In this case, the steering torque L is determined by way of example using the validation points. The diagrams show the output of a trained artificial neural network as an example of a machine learning algorithm using validation data as input. The output of the artificial neural network is plotted together with the known setpoint or target value of the steering torque of the validation data.

[0054] In this case, the quality of the performance of the clustering can be determined with the aid of a quality criterion. In the present case, the mean square error Mean Err of the output of the machine learning algorithm ANN and the corresponding target value Target can be advantageously used. Thus, the mean error Mean Err of the machine learning algorithm is quantified. As another particularly sensitive quality criterion, the number n Err of mean square errors above a threshold value between the target value and the output value can be determined. In the present example, the chart exemplarily shows 42274 data points, an average error of 0.18177 Nm, and a threshold value exceeding number of 750, wherein these numerical examples are added only for the purpose of illustrating the present invention.

[0055] In this context, it is emphasized that the analysis of the performance of the individual clusters can be achieved by using the data points of the corresponding assigned clusters as verification data. These verification data are extracted from the group of unselected data points of these belonging clusters. Therefore, the present invention has the further advantage that a large amount of verification data can be provided by selecting and subsequently reducing the data points in these clusters when compiling the training data set. These verification data can in particular be directly assigned to each cluster or each scenario. From this, the above-mentioned cluster-related performance can be determined. Therefore, in the event of a fault, the corresponding cluster can be located or found, such as in Figure 4 As shown for cluster 1 in .

[0056] In addition, Figure 5 The examination of the cluster-related performance is further clarified in FIG. 1 , in which the performance determined for different clusters, i.e. cluster-specifically, is collected. In the above example, nine different clusters are listed, for which the performance is determined in accordance with the combination of Figure 4 The described manner and method are used to carry out a quality check. In this case, the quality parameters are indicated purely by way of example: mean square error Mean Error (first row), number of threshold value violations Number, and threshold value violation in percentage.

[0057] Figure 6 Another aspect of the method is illustrated in which further improved results for the training data set are shown.

[0058] If the quality criteria of one or more clusters are not met, additional data points can be added to the training data set for these clusters. The training data set is thus modified. For example, in one or more clusters it is ascertained according to the above-mentioned performance test that, depending on the definition of the quality criteria, a threshold value has been exceeded once or more than once in terms of the mean square error.

[0059] Then, the failure or fault situation can be assigned to a cluster or related clusters (see, for example, Figure 4 and 5 ), so the number of data points of these clusters can be increased. This can be advantageously achieved with the aid of unselected data points that are exactly from these relevant clusters. Advantageously, this can selectively improve the performance of individual clusters. This achieves automated feedback.

[0060] After adding the additional data points, the above steps of training and checking the performance of the individual clusters, in particular the relevant clusters, can be repeatedly performed, wherein reference is made to the above embodiments and descriptions.

[0061] Now, in Figure 6 The results of this improvement for a training data set for a specific cluster, here cluster 9, are shown in FIG. The left part exemplifies the performance test of the cluster of the training data set as originally arranged. The right part shows the performance of the same cluster of the training data set improved with the aid of individual data points as described above.

[0062] Compared to the originally programmed clusters, the clusters improved by retraining show in performance tests a lower mean difference Mean Err of the output values ​​ANN and the target values ​​Target. In this case, the number of threshold violations n Err is reduced by a factor of 2, as also seen in the marking of the threshold violations indicated by the triangles in the comparison. The number of threshold violations is therefore a very sensitive measure for evaluating performance.

[0063] The above modification of the training data set can be repeated successively in order to improve the performance of each cluster successively. In particular, this can be repeated until all clusters meet the set quality criteria. A high-performance algorithm is thus generated. For example, the improvement can be performed until no clusters show a threshold value exceeded in the performance test. Thus, the method enables a sequential and automated improvement of the training data set until the modified training data set meets one or more quality criteria in all clusters. The improvement cycle can be performed automatically without the need for internal intervention.

[0064] By means of the described method, a high-performance machine learning algorithm can be provided. The output of this machine learning algorithm having the above-mentioned special properties can then be advantageously used in a method for operating a vehicle component. In this case, the vehicle component is controlled or regulated or operated in response to the output of the high-performance machine learning algorithm trained according to the learning method.

[0065] Figure 7 A control unit 10 according to the invention is schematically shown. In this case, the control unit 10 is set up to implement the above-described method. The control unit 10 can in particular operate a vehicle component 20 in response to or based on the output of the above-described machine learning algorithm. The control unit 10 can also be part of the vehicle component 20. In certain applications, the vehicle component 20 can be, for example, a steering wheel, wherein the control unit 10 uses the output of the machine learning algorithm for steering torque approximation, as is required, for example, in a steer-by-wire system. In other examples, an approximation of tire parameters can be performed. However, the present invention is not limited to these examples. The training data set formed by the method improves the performance of the machine learning algorithm in its different embodiments. Therefore, aspects of the method directly lead to an improvement in the operation of the vehicle component 20.

[0066] Reference numerals list

[0067] 10 Control unit

[0068] 20 Vehicle Components

[0069] 100 Vehicles

[0070] F1,...,Fn data points

[0071] C1,...,Cm clustering

[0072] T1,...,Tk The selected data points of the first cluster

[0073] R1,...,Rk are the selected data points of the remaining clusters

Claims

1. A computer-implemented machine learning method for operating a vehicle component (20), the method comprising: - providing a plurality of multidimensional data points (F1, ..., Fn), wherein the data points (F1, ..., Fn) are measured data points related to the vehicle; - generating a basic training data set by partitioning a plurality of data points (F1, ..., Fn) into multidimensional clusters (C1, ..., Cm) using a clustering algorithm, such that each data point (F1, ..., Fn) is assigned to exactly one cluster (C1, ..., Cm); - generating a training data set by selecting data points from the basic training data set, the generating comprising: - determining the smallest cluster (Ci) among the plurality of clusters (C1, ..., Cm) having the least number of data points, - selecting for said training data set at least a subset of the data points (T1, ..., Tk) of said smallest cluster (Ci), and -Select a subset of data points (R1, ..., Rk) from each of the remaining clusters (C1, ..., Ci-1, Ci+1, ..., Cm) for the training data set, so that the number of selected data points (R1, ..., Rk) of each of the remaining clusters (C1, ..., Ci-1, Ci+1, ..., Cm) corresponds to the number of selected data points (T1, ..., Tk) of the smallest cluster (Ci).

2. The method according to claim 1, further comprising: The unselected data points of the smallest cluster (Ci) and / or the remaining clusters (C1, ..., Ci-1, Ci+1, ..., Cm) are stored in a data storage as test and / or validation data for a machine learning algorithm.

3. The method according to claim 1 or 2, further comprising: The set of all data points (R1, ..., Rk) of the smallest cluster (Ci) is selected for the training data set.

4. The method according to claim 1 or 2, wherein data points are randomly selected from the smallest cluster (Ci) and / or the remaining clusters (C1, ..., Ci-1, Ci+1, ..., Cm).

5. A method according to claim 1 or 2, wherein data points are selected from the smallest cluster (Ci) and / or the remaining clusters (C1,...,Ci-1,Ci+1,...,Cm) so that the selected data points of the clusters have a maximum distance from each other.

6. The method according to claim 1 or 2, further comprising: The machine learning algorithm is trained with the help of the generated training data set.

7. The method according to claim 6, further comprising: The performance of the individual clusters (C1, . . . , Cm) is determined using quality criteria using validation data.

8. The method according to claim 7, further comprising: The performance of the respective cluster is determined by the data points in the group of non-selected data points of the corresponding cluster.

9. The method of claim 7, wherein the quality criterion comprises a mean square error of output values ​​relative to a target value above a threshold.

10. The method according to claim 7, further comprising: A modified training data set is created by adding data points to clusters that do not meet the quality criteria.

11. The method of claim 10, wherein the added data points are extracted from the group of non-selected data points of that cluster that do not meet the quality criterion.

12. The method according to claim 10, further comprising: The modification of the training data set is repeated until the quality criterion is met for all clusters (C1, ..., Cm).

13. A method for operating a vehicle component (20), the method comprising: The vehicle component (20) is operated based on the output of a machine learning algorithm trained according to the method of any one of claims 1 to 12 above.

14. A control unit (10) for a vehicle (100), the control unit being configured to carry out the method according to claim 13 or according to any one of claims 1 to 12.

15. A vehicle (100) comprising a control unit (10) according to claim 14.

Images