Method for training a neural network, method and device for estimating a friction coefficient, and machine-readable storage medium

Abstract

The invention relates to a method for training a neural network for estimating a friction coefficient of a wheel relative to a ground, a method for estimating a friction coefficient and a corresponding device and a machine readable storage medium, wherein at least one discrete data sequence (311, 312, 313, 314) of a sensor type specific to the friction coefficient is provided (S1); a discrete data sequence (311, 312, 313, 314) of at least one driving dynamic sensor type for the wheel is provided for a limited acceleration range of the wheel (S2); the at least two discrete data sequences (311, 312, 313, 314) are synchronized (S3); an actively determined friction coefficient for the wheel is provided (S5); an input tensor for a neural network (320) is generated by means of the discrete data sequences (311, 312, 313, 314) (S4); the neural network is trained by means of the input tensor (S7) in order to estimate the friction coefficient.

Description

Technical Field

[0001] The present invention relates to a method for training a neural network and a method for estimating the friction coefficient of a wheel relative to a ground surface using the trained neural network. Background Art

[0002] Road conditions are crucial for vehicle operation and general road safety, and they play a key role in accidents. If the maximum coefficient of friction between the wheels and the roadway is exceeded, for example in hydroplaning or winter conditions such as heavy rain, snow, or ice, unstable driving conditions may result due to a loss of wheel adhesion to the roadway. For many driver assistance systems and partially automated vehicles, it is crucial not to exceed the maximum coefficient of friction in order to maintain a safe driving state or, if necessary, terminate the automated driving function.

[0003] Modern motor vehicles have control systems such as the Electronic Stability Program (ESP). Essentially, the ESP is also a type of traction control system. Safety systems such as the Anti-lock Braking System (ABS) or Traction Control System (TCS) then intervene when critical driving situations occur.

[0004] During intervention by the control system, the actual friction coefficient can be determined through measurements, resulting in a reliable value. However, since intervention by driver assistance systems is relatively rare during normal driving, there is usually little information about the currently existing minimum friction coefficient during vehicle operation. Therefore, an estimate of the maximum friction coefficient is necessary.

[0005] Typically, the aforementioned control device has a plurality of devices for determining the current friction coefficient between the wheels of the motor vehicle and the roadway. Here, these so-called friction coefficient estimators are specially designed for different driver assistance devices and are independent of each other. Summary of the Invention

[0006] The estimation of tire-roadway friction is primarily based on the slip-dependent modeling of the friction coefficient. Depending on the road surface, such as dry or wet asphalt, this dependence exhibits a very specific characteristic curve that can be used to estimate the maximum possible friction (μmax). This estimation is based on the slope of the characteristic curve in the linear range.

[0007] The present invention discloses a method for training a neural network, a method for estimating a friction coefficient using a trained neural network, a device, a computer program, and a machine-readable storage medium, which at least partially solve the above-mentioned tasks.

[0008] The present invention is based on the knowledge that, given the numerous sensors installed in modern vehicles that can provide both friction-coefficient-specific and driving dynamics data series, neural networks for estimating friction coefficients can be trained without model assumptions using actively determined friction coefficients.

[0009] Throughout the present description, the order of method steps is described so that the method is easy to understand. However, those skilled in the art will recognize that many method steps can also be performed in other orders and lead to the same or similar results. In this sense, the order of method steps can be changed accordingly.

[0010] According to one aspect, a method for training a neural network for estimating the friction coefficient of a wheel relative to a road surface is provided. In one step, at least one discrete data sequence specific to a friction coefficient sensor type is provided. In another step, at least one discrete data sequence specific to a driving dynamics sensor type for a wheel within a limited acceleration range of the wheel is provided. In another step, these at least two discrete data sequences are synchronized. In another step, an actively determined friction coefficient of the wheel, which is specific to the discrete data sequence, is provided. In another step, an input tensor for the neural network is generated using the discrete data sequences specific to the friction coefficient sensor type and the driving dynamics sensor type.

[0011] In a further step, a plurality of input tensors generated in this manner are formed using a plurality of different data sequences specific to at least the friction coefficient sensor type and the driving dynamics sensor type, each of which has an associated, actively determined friction coefficient. In a further step, a neural network is trained using the plurality of input tensors with the respectively specific, actively determined friction coefficients in order to estimate the friction coefficient using discrete data sequences specific to at least the friction coefficient sensor type and the driving dynamics sensor type.

[0012] In this context, friction coefficient-specific sensor types are understood to be those sensors that can determine factors that influence the determined friction coefficient of the wheel relative to the ground. These include, for example, external influencing factors such as temperature, additional layers on the ground (e.g., water or ice), but also the type and properties of the ground itself, as well as the material, properties, and geometry of the wheel.

[0013] Figure 1 The typical dependence of the adhesion or friction coefficient μ on the slip λ for different road surface properties is shown by way of example. Depending on the dryness 110 or snow 120 and ice 130 , different friction coefficient profiles with different maximum friction coefficients are obtained.

[0014] The driving dynamics sensor type description records the sensor type for data or data sequences that allow inferences about the dynamic driving process of a vehicle that is at least partially driven or stabilized by the wheels in question, and also includes driving dynamics data sequences of the wheels themselves. However, only data sequences obtained from a limited acceleration range of the vehicle are provided, excluding data sequences that occur, for example, when automatic safety systems intervene. The reason for this is that the vehicle's friction coefficient should be estimated during normal operation. When an automatic safety system intervenes, the friction coefficient can be determined directly, as described above according to the prior art. However, such interventions occur only occasionally during normal driving, which necessitates a method for estimating the friction coefficient.

[0015] exist Figure 1 In the first part of the curve, i.e., at low slip λ, a linear progression of the friction coefficient μ as a function of slip λ can be seen. The slope of this linear progression indicates the longitudinal tire stiffness, also known as tire rigidity. The slope of the μ-λ curve, i.e., the tire rigidity, is particularly dependent on the structure of the tire and any intermediate layer. Soft tires (e.g., winter tires) have a significantly flatter μ-λ curve than hard tires. This means that even if data sequences of driving dynamics sensors are determined within a limited acceleration range, it is possible to infer the friction coefficient using these data sequences.

[0016] Since the neural network is trained to estimate the friction coefficient using both friction-coefficient-specific sensor data or data sequences and driving dynamics sensor data sequences, model-based approaches to the complex relationship between driving dynamics and external factors influencing the friction coefficient become unnecessary. In particular, since the training can be performed using a specific vehicle, the completely specific vehicle characteristics can also be taken into account when estimating the friction coefficient.

[0017] The duration of the data series does not have to be predetermined but can be adapted to the circumstances. In particular, different sensor types can also provide data series of the same length or data series of defined time intervals for training the neural network. For some sensor types, the duration of the data series can also be determined by specifying the validity of the values ​​within a defined time interval.

[0018] Since different sensor types generate data at different times or with different frequencies, the data can be synchronized in time during the synchronization step of the method using time stamps that can be assigned to these data. However, interpolation can also be performed between data points of one sensor type in order to synchronize with another sensor type that generates data at a different frequency, for example.

[0019] Different methods for generating actively determined friction coefficients, which are provided to the method for training a neural network, are described below. When determining the corresponding friction coefficients, a temporal correlation with the discrete data sequence is necessary to establish a causal relationship between the discrete data sequence and the friction coefficient to be estimated. Depending on the method for actively determining the friction coefficients, the friction coefficients can be determined simultaneously with the discrete data sequence, in a close temporal relationship (i.e., to the extent that functional causality is preserved), or even after the discrete data sequence has been determined. In other words, it is important to assign a specific, actively determined friction coefficient to the discrete data sequence.

[0020] Neural networks provide a framework for many different algorithms for machine learning, for collaboration, and for processing complex data inputs. Such neural networks learn to perform tasks based on examples, typically without being programmed with task-specific rules.

[0021] Neural networks are based on a collection of connected units, or nodes, called artificial neurons. Each connection can transmit a signal from one artificial neuron to another. The receiving artificial neuron can process the signal and then activate another artificial neuron connected to it.

[0022] In conventional neural network implementations, the signal on the connection of an artificial neuron is a real number, and the output of the artificial neuron is calculated by a nonlinear function of the sum of the inputs of the artificial neuron. The connection of the artificial neuron typically has a weight that adapts as learning progresses. The weight increases or decreases the intensity of the signal on the connection. The artificial neuron may have a threshold so that it outputs a signal only when the total signal exceeds the threshold. Typically, multiple artificial neurons are layered and merged. Different layers may perform different types of nonlinear transformations on the input of this layer. The signal propagates from the first layer (input layer) to the last layer (output layer), having previously passed through these layers multiple times.

[0023] The architecture of an artificial neuron feedforward network or a recurrent neuron network may be one that is configured such that it receives at its input level a single data pattern corresponding to a graph or a data sequence and provides an output classification vector that assigns the pattern in the input data to different classes.

[0024] According to one aspect of the method, the neural network is a recurrent neural network (RNN). Such neural networks are called recurrent or feedback neural networks: in contrast to feedforward networks, these networks are characterized by neurons in one layer being connected to neurons in the same or previous layer. This recurrent interconnection of model neurons can be used to discover temporal information encoded in data.

[0025] RNNs are a type of neural network that is dependent on the sequential nature of its input values. They are based on the principle that the input is dependent on the previous input (the order of the values ​​in the time series) by having a "memory" that records what has been seen so far, which is responsible for the "memory" of values ​​over any time interval. Furthermore, any number of input time series can be given, which directly supports multivariate estimation. This is particularly useful when analyzing time series observations of vehicle signals from multiple sources and with multivariate properties, such as sequences of speed, acceleration, torque, etc. – all of which are related to a target variable, in this case, the maximum tire-road friction μmax.

[0026] Machine learning algorithms automatically learn from data and make predictions or estimates. For the maximum friction prediction task, this means collecting sufficient data—including dry and wet road surfaces, asphalt and concrete—to extract representative patterns. In other words, measurements must be performed on various surface coverages, from dry to wet to icy, in order to correctly estimate the contribution of these independent variables to μmax.

[0027] The neural network can be trained with the aid of training data (i.e., with the aid of training tensors) and the measured friction coefficient in order to reproduce the friction coefficient from the data sequence. With the aid of the defined structure of the neural network, each neuron connection is given, for example, a random initial weight. The input data (i.e., the training tensor) is then fed into the network, each neuron weights the input signal with its weight, and passes the result on to the neurons in the next layer. The entire result is then provided at the output layer. The error in the reproduction of the friction coefficient and the contribution of each neuron to this error can be calculated, and then the weight of each neuron is changed in the direction that minimizes the error. A recursive operation is then performed, the error is remeasured, and the weights are adapted until the error is below a predetermined limit.

[0028] According to one aspect of the method, the input tensors are generated using a provided discrete data sequence of a first plurality of friction coefficient-specific sensor types and a provided discrete data sequence of a second plurality of driving dynamics sensor types, wherein all the discrete data sequences are synchronized with one another.

[0029] The plurality of sensor types may include, for example, two or three sensor types, or may include 100 or more sensor types, with the aid of which discrete data sequences of the sensor types are used to train or use the neural network.

[0030] As shown above, the coefficient of friction of the wheel relative to the ground is affected by many external factors, which can be detected by a first plurality of friction-coefficient-specific sensor types under different conditions of wheel use. Corresponding friction-coefficient-specific sensor types detect different characteristics of these conditions, and since these first plurality of sensor types generate a corresponding number of data segments, these data segments can be provided to the neural network for training and estimating the friction coefficient to achieve the highest possible accuracy and reliability.

[0031] Compared to a single driving dynamics sensor type, the second plurality of driving dynamics sensor types makes it easier to infer the driving dynamics behavior of a vehicle equipped with such wheels and can be detected more reliably. Therefore, if data segments from the second plurality of driving dynamics sensor types are used both during training and during estimation using the neural network, the neural network's prediction of the friction coefficient is improved.

[0032] By using different sensor types, factors that might be detrimental to estimating the friction coefficient from data sequences of different sensor types can be taken into account when training the neural network, without requiring additional technical models. For example, the friction coefficient should only be determined when there are no gear shifts, no changes in engine torque, or no unfavorable road conditions. If these conditions are detected, particularly using data sequences from driving dynamics sensors, the neural network can be trained to account for these effects.

[0033] According to one aspect of the method, it is provided that the at least one driving dynamics sensor type is a longitudinal acceleration sensor, a lateral acceleration sensor, a yaw rate sensor, a wheel speed sensor, a brake pressure sensor, or a steering angle sensor.

[0034] According to one aspect of the method, the at least one friction coefficient-specific sensor type is a digital camera system, a radar system, a lidar system, an ultrasound system, an optical ground detection system, a temperature sensor, a friction coefficient map, a navigation system, or a wireless information system.

[0035] Various friction-coefficient-specific sensor types detect environmental conditions that may be correlated with the friction coefficient level. For example, ultrasonic sensors, in addition to their function as distance sensors, can also detect splashing water while driving, or their ultrasonic signals can detect different signals depending on the properties of the ground, which can influence the friction coefficient level. Radar and lidar systems can also detect the properties of the ground or intermediate layers above the vehicle and thus provide relevant information for estimating the friction coefficient.

[0036] If the type of ground surface is known based on a friction coefficient map, this further improves the estimation of the friction coefficient. Here, the position of the vehicle or wheel can be assigned to such a friction coefficient map by means of a navigation system, or other information about the ground surface can be obtained from the navigation system.

[0037] Furthermore, further information about the road surface can be transmitted to the vehicle using the wireless information system and taken into account when estimating the coefficient of friction. Furthermore, information that improves the estimation of the coefficient of friction can also be transmitted from or to other vehicles using such a wireless information system.

[0038] Direct information generated, for example, by an optical ground detection system or temperature data about the environment, the ground, the roadway, or the wheels themselves can further improve the friction coefficient estimation. It is not necessary to store a function model to use this data to estimate the friction coefficient; it is sufficient to include sufficient such data in the training dataset for the neural network. Data sequences from different sensor types can be provided via the vehicle's bus system (e.g., a CAN bus), allowing other sensors to be flexibly integrated into the method.

[0039] In other words, the method for training a neural network for estimating the friction coefficient comprises the following steps:

[0040] 1. Representative data acquisition of multiple vehicle sensor signals (e.g., vehicle acceleration, speed, slip, drive torque, etc.) is performed on different road surface types (e.g., asphalt or concrete) and under different road surface conditions or coatings (e.g., dry, wet, or icy conditions). The actively determined friction coefficient μmax can be calculated from intervention signals from, for example, ABS or ESP. These intervention signals serve as ground truth labels for training neural networks (particularly recurrent neural networks).

[0041] 2. Paired multivariate time series of sensor signals of length T and actively determined friction coefficient μmax are used to train a neural network to predict or estimate μmax. The neural network recognizes patterns in the time series of sensor signals in order to "learn" the behavior of μmax with respect to the multivariate sequential time series input.

[0042] A method for estimating the friction coefficient using a correspondingly trained neural network is used to predict or estimate the maximum friction, i.e., the friction coefficient, in multiple vehicles in real time. The estimation is based on analyzing and processing time series of data from various vehicle sensors. The neural network can be trained offline using the provided data, and the trained neural network can then be used online in different vehicles to estimate the friction coefficient. This method is therefore scalable, meaning it can be created once and used in multiple vehicles.

[0043] According to one aspect of the method, it is provided that individual data sequences are removed in a statistically controlled manner from a plurality of data sequences of a friction coefficient-specific sensor type and / or a driving dynamics sensor type during the training of the neural network.

[0044] This allows the neural network to be trained to continue estimating the friction coefficient even when data sequences from different sensor types are faulty. This is because, in the real world of at least partially automated vehicles, not all data sequences from different sensor types are always available. This can be due to system failures, which can be caused by both technical and environmental faults, but can also be caused by, for example, the lack of data such as friction coefficient maps for certain regions. To ensure safe operation under all these conditions, it is advantageous to generalize the neural network to such an extent that a reliable friction coefficient estimate can be generated even under these conditions.

[0045] This elimination of data sequences can be done statistically, if no foreseeable failure scenarios exist for data sequences of a specific sensor type. However, it is also possible to predict which data sequences of a particular sensor type or group of sensor types will at least temporarily fail during actual operation, allowing the neural network to be strategically trained for such scenarios. This makes the use of the neural network independent of external information sources, such as weather service information. For predictions using the neural network, only the signals or data sequences of the vehicle's different sensor types are required.

[0046] According to one aspect of the method, it is provided that the actively determined friction coefficient is determined for the vehicle using a driving dynamics sensor type using model-based driving dynamics measurements.

[0047] As mentioned above, if a safety system (eg, an ESP system) intervenes, such a model-based driving dynamics measurement is performed with the aid of sensor systems installed in the vehicle.

[0048] According to one aspect of the method, it is provided that the actively determined coefficient of friction is determined for the wheels by means of an active direct friction coefficient measurement.

[0049] For active direct friction coefficient measurement, a measuring vehicle equipped with friction coefficient measurement technology is used, which determines the friction coefficient using, for example, a surface friction tester. Both methods are based on force measurement. A surface friction tester is a vehicle trailer with three wheels. The third wheel is braked to the physical limit, i.e., to a defined slip. The friction force can be determined from the braking force and the braking torque required for this, and the friction coefficient can be determined using the known normal force. Since this exemplary measurement method for active direct friction coefficient measurement is complex and expensive, it is only a possible alternative to the above-mentioned method for determining the actively determined friction coefficient.

[0050] A method for estimating the friction coefficient of a wheel relative to the ground using a correspondingly trained neural network as described above is proposed. In one step of the method, at least one discrete data sequence of a friction coefficient sensor type is provided. In another step, at least one discrete data sequence of a vehicle dynamics sensor type for a limited acceleration range of the wheel is provided. Furthermore, in one step, at least the two discrete data sequences are synchronized. In one step, an input tensor for the trained neural network is generated using the discrete data sequence of the friction coefficient sensor type and the discrete data sequence of the vehicle dynamics sensor type. In another step, the friction coefficient is estimated using the input tensor and the trained neural network.

[0051] The friction coefficient can thus also be estimated without the need for intervention of safety systems and made available for further use, for example to a control system of an at least partially automated vehicle, in order to ensure safe at least partially automated driving.

[0052] According to one aspect of the method for estimating the friction coefficient, the input tensors are generated using a discrete data sequence provided by a first plurality of friction coefficient-specific sensor types and a discrete data sequence provided by a second plurality of driving dynamics sensor types, wherein all the discrete data sequences are synchronized with one another.

[0053] As already mentioned above in the training of the neural network, the friction coefficient can be determined more reliably and more accurately using a larger number of data sequences from different sensor types. In particular, it can also be provided that some sensor types are designed redundantly.

[0054] According to one aspect of the method for estimating the friction coefficient, the estimated friction coefficient is checked using a plausibility calculation. Since the friction coefficient can also be estimated using a calculation based on the identified environment, the corresponding conditions, the wheels used, and other similar knowledge, the plausibility of the friction coefficient estimated by the neural network can be checked. This allows the identification of outliers, for example due to unusual input data for which the neural network has not yet been trained.

[0055] According to one aspect of the method, a control signal for controlling the at least partially automated vehicle and / or a warning signal for warning a vehicle occupant is emitted as a function of at least one of the estimated values ​​of the friction coefficient. Thus, for example, a controller of the at least partially automated vehicle can react to a particularly low friction coefficient by adopting a correspondingly modified driving style, or the low friction coefficient can be indicated to a vehicle occupant so that the occupant's driving style can be adapted accordingly.

[0056] According to one aspect, a device is provided, which is configured to carry out one of the above-mentioned methods. With the aid of this device, the corresponding method can be easily integrated into different systems.

[0057] According to another aspect, a computer program is provided, comprising instructions which, when executed by a computer, cause the computer to carry out one of the above methods. Such a computer program enables the use of the described method in various systems.

[0058] A machine-readable storage medium is proposed, on which the above-mentioned computer program is stored. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] refer to Figures 1 to 4 An embodiment of the present invention is shown and explained in more detail below. The accompanying drawings show:

[0060] Figure 1 : Friction coefficient curves under different ground conditions;

[0061] Figure 2 : Vehicles with multiple friction-specific sensor types or driving dynamics sensor types.

[0062] Figure 3 : A simplified diagram of a neural network trained to estimate the friction coefficient;

[0063] Figure 4 : A simplified diagram of estimating the friction coefficient using a trained neural network. DETAILED DESCRIPTION

[0064] Figure 1An example of the dependence of the friction coefficient μ on the slip λ for different road conditions is shown, wherein different profiles of the friction coefficient μ as a function of the slip λ occur depending on the dryness 110 or snow and ice 120 , 130 . Figure 1 It is shown that for small slip values, there is an almost linear relationship between slip λ and friction coefficient μ. This is indicated by dashed straight lines 115, 125 and demonstrates that the friction coefficient can be estimated based on driving dynamics data even within a limited acceleration range, i.e., without traversing a large portion of the curve (e.g., when a safety system intervenes).

[0065] Figure 2 A number of different sensor types and their exemplary arrangement relative to vehicle 200 are schematically shown. Specifically, they include a pressure sensor 215 in the braking system, a wheel rotation sensor 230, a radar sensor 240, a lidar sensor 245, an ultrasonic sensor 250, a camera 260, an optical sensor 265 for detecting the ground, a communication device 275, a positioning device (GPS) 280, a navigation device 292, an RFID sensor 294 for identifying the tire type, a steering angle sensor 296, a rack force sensor 298, and an ESP system 210. All sensor types are coupled to a bus (e.g., a CAN bus) represented by a single line in a signal relationship. Additionally, a display device 270 for transmitting information to the vehicle occupants is schematically depicted.

[0066] Figure 3 A method 300 for training a neural network 320 for estimating the friction coefficient of a wheel relative to the ground is described. In a step S1, at least one discrete data sequence 311, 312, 313, 314 of a sensor type specific to the friction coefficient is provided. In a further step S2, a discrete data sequence 311, 312, 313, 314 of at least one wheel driving dynamics sensor type for a limited acceleration range of the wheel is provided. In a further step S3, these at least two discrete data sequences are synchronized. In a further step S5, an actively determined wheel friction coefficient is provided, which is specific to the discrete data sequence 311, 312, 313, 314. This specificity or correlation is in Figure 3 It is indicated by a dotted line 310 .

[0067] In a further step S4, input tensors for a neural network 320 are generated using discrete data sequences 311, 312, 313, 314 specific to the friction coefficient sensor type and the driving dynamics sensor type. In a further step S6, a plurality of input tensors generated in this manner are formed, along with the respectively associated, actively determined friction coefficient, using a plurality of different data sequences 311, 312, 313, 314 specific to at least the friction coefficient sensor type and the driving dynamics sensor type. In a further step S7, neural network 300 is trained using the plurality of input tensors with the respectively associated, actively determined friction coefficients in order to estimate the friction coefficient using the discrete data sequences (311, 312, 313, 314) specific to at least the friction coefficient sensor type and the driving dynamics sensor type. In particular, when training S7 the neural network, individual data sequences 311 , 312 , 313 , 314 can be removed S8 in a statistically or strategically controlled manner from a plurality of data sequences 311 , 312 , 313 , 314 of a friction coefficient sensor type and / or driving dynamics sensor type, as already explained in more detail above.

[0068] Figure 4 A method for estimating the friction coefficient of a wheel relative to the road surface using a neural network 320 trained as described above is described. In step S11 of the method, at least one discrete data sequence 311, 312, 313, 314 of a friction coefficient sensor type is provided. In another step S12, at least one discrete data sequence 311, 312, 313, 314 of a wheel driving dynamics sensor type for a limited acceleration range of the wheel is provided. Furthermore, in step S13, at least two discrete data sequences 311, 312, 313, 314 are synchronized. In step S14, an input tensor for trained neural network 320 is generated using discrete data sequences 311, 312, 313, 314 of the friction coefficient sensor type and the driving dynamics sensor type. In another step S15, the friction coefficient 350 is estimated using the input sensors and the trained neural network. Estimated friction coefficient 350 can be checked, in particular, using a plausibility calculation (step S16).

Claims

1. A method for training a neural network (320) for estimating the friction coefficient of a wheel relative to a ground surface, wherein: The method comprises: providing at least one discrete data sequence of a friction coefficient-specific sensor type; providing a discrete data sequence of at least one driving dynamics sensor type for the wheel in a limited acceleration range of the wheel; synchronizing at least two discrete data sequences; providing an actively determined friction coefficient for the wheel, the friction coefficient being specific to the discrete data sequence; generating an input tensor for the neural network (320) with the aid of the friction coefficient-specific sensor type and the discrete data sequence of the driving dynamics sensor type; forming a plurality of input tensors generated in this way with the aid of a plurality of different data sequences of at least the friction coefficient-specific sensor type and the driving dynamics sensor type, the input tensors having respectively associated, actively determined friction coefficients; training the neural network with the aid of the plurality of input tensors having respectively specific, actively determined friction coefficients in order to estimate the friction coefficient with the aid of at least the discrete data sequence of the friction coefficient-specific sensor type and the discrete data sequence of the driving dynamics sensor type.

2. The method according to claim 1, wherein The input tensors are each generated using a discrete data sequence provided by a first plurality of friction coefficient-specific sensor types and a discrete data sequence provided by a second plurality of driving dynamics sensor types, wherein all the discrete data sequences are synchronized with one another.

3. The method according to claim 1 or 2, wherein: The at least one driving dynamics sensor type is selected from the group consisting of a longitudinal acceleration sensor, a lateral acceleration sensor, a yaw rate sensor, a wheel speed sensor, a brake pressure sensor, and a steering angle sensor.

4. The method according to claim 1 or 2, wherein: The at least one friction coefficient-specific sensor type is selected from the group consisting of digital camera systems, radar systems, lidar systems, ultrasound systems, optical ground detection systems, temperature sensors, friction coefficient maps, navigation systems, and wireless information systems.

5. The method according to claim 1 or 2, wherein: During the training of the neural network (320), individual data sequences are removed in a statistically controlled manner from a plurality of data sequences from a friction coefficient-specific sensor type and / or a driving dynamics sensor type.

6. The method according to claim 1 or 2, wherein: The actively determined coefficient of friction for the wheel is determined by means of a driving dynamics sensor type using a model-based driving dynamics measurement.

7. The method according to claim 1 or 2, wherein: The actively determined coefficient of friction for the wheel is determined by means of an active direct coefficient of friction measurement.

8. A method for estimating the coefficient of friction of a wheel relative to the ground using a neural network (320) trained according to the method of any one of claims 1 to 7, wherein: The method comprises: providing at least one discrete data sequence of a sensor type specific to the friction coefficient; providing at least one discrete data sequence of a driving dynamics sensor type for the wheel in a limited acceleration range of the wheel; synchronizing the at least two discrete data sequences; generating an input tensor for the trained neural network (320) with the aid of the discrete data sequence of the sensor type specific to the friction coefficient and the discrete data sequence of the driving dynamics sensor type; estimating the friction coefficient with the aid of the input tensor and the trained neural network (320).

9. The method according to claim 8, wherein The input tensors are each generated using a discrete data sequence provided by a first plurality of friction coefficient-specific sensor types and a discrete data sequence provided by a second plurality of driving dynamics sensor types, wherein all the discrete data sequences are synchronized with one another.

10. The method according to claim 8 or 9, wherein: The estimated friction coefficient is checked with the aid of a plausibility calculation.

11. The method according to claim 8 or 9, wherein: Depending on at least one of the estimated values ​​of the coefficient of friction, an actuation signal for actuating the at least partially automated vehicle and / or a warning signal for warning a vehicle occupant is emitted.

12. An apparatus configured to perform the method according to any one of claims 1 to 11.

13. A computer program comprising instructions which, when the computer program is run by a computer, arrange the computer to carry out the method according to any one of claims 1 to 12.

14. A machine-readable storage medium having stored thereon the computer program according to claim 13.

Images