WO2022106342A1

WO2022106342A1 - Evaluation device, computer program and computer-implemented method for training a neural network to determine coefficients of friction

Info

Publication number: WO2022106342A1
Application number: PCT/EP2021/081647
Authority: WO
Inventors: Mehrdad Salari; Ulrich Mair
Original assignee: Zf Friedrichshafen Ag
Priority date: 2020-11-20
Filing date: 2021-11-15
Publication date: 2022-05-27
Also published as: DE102020214620A1

Abstract

The invention relates to an evaluation device (120), to a computer program and to a computer-implemented method for training a neural network (200) to determine coefficients of friction of a motor vehicle tyre. The method is implemented on the evaluation device (120) of a motor vehicle (100) and comprises acquiring a first input dataset (210) by way of at least one first sensor arrangement (102), predictively computing a first roadway coefficient of friction (242) by way of the neural network (200) on the basis of the first input dataset (210), acquiring a second dataset (213) by way of at least one second sensor arrangement (110), determining a second roadway coefficient of friction (244), determining at least one weighting error (246) of the neural network (200) taking into consideration the first roadway coefficient of friction (242) and the second roadway coefficient of friction (244) and training the neural network (200) to minimize the at least one weighting error (246).

Description

Evaluation device, computer program and computer-implemented method for training a neural network for determining the coefficient of friction

The invention relates to a computer-implemented method for training a neural network for determining the coefficient of friction, the neural network being executed on an evaluation device of a motor vehicle.

Advances in sensor and data processing technologies are making vehicles increasingly autonomous. The introduction and improvement of assistance systems such as electronic stability control, adaptive cruise control, lane departure warning, adaptive steering systems, anti-lock braking systems, traction control, map databases, vehicle-to-vehicle communication and other assistance systems enable a vehicle to drive partially or fully autonomously. The vehicle can replace the driver in various situations.

An evaluation device interprets the input data, and in the case of vehicles in an autonomous driving mode, a control device takes over both the longitudinal guidance and the lateral guidance of the motor vehicle. The otherwise active driver of the motor vehicle becomes the passive vehicle occupant of the vehicle. Semi-autonomous and in particular autonomous vehicles use sensors to record the road ahead and the area around the vehicle.

The sensors preferably provide the evaluation device with input data, such as lane boundaries, lane edges, lane markings and other road markings.

Some vehicle systems also use input data from global positioning systems, such as GPS, vehicle-to-vehicle communications, and/or on-board data to automatically control the vehicle. Autonomous driving can be divided into five levels of automation. Automation level one corresponds to assisted driving and automation level five corresponds to fully autonomous driving. In fully autonomous driving, several input data are combined sioned to achieve higher reliability. The higher the level of automation in autonomous driving, the higher the demands on the reliability and quality of the input data. Even a small error in the calculation of driving-specific variables can result in a safety risk for the vehicle, vehicle occupants and external road users in fully autonomous driving mode.

For this reason, the determination of driving-specific variables is particularly important in a fully autonomous driving mode. The condition of the road surface and the road surface friction between the vehicle's tires and the road are necessary in order to be able to react specifically and appropriately to conditions external to the vehicle. The roadway friction coefficient between the roadway surface and at least one tire of the vehicle is important in particular in the case of high lateral acceleration in curves and during braking and acceleration processes, in order to adapt the speed of the vehicle to the conditions external to the vehicle if necessary.

Optical methods for detecting the roadway are already widely known. Road markings and other objects are detected and recognized in DE 10 2015 121 504 A1, for example. Furthermore, rudimentary methods for determining a roadway friction value are known. A road surface friction value can be determined from the tire speed. For example, motor vehicles are equipped with an anti-lock braking system to prevent the wheels from locking in a braking situation. The tire speeds of individual tires are measured using tire speed sensors and compared with one another. The brake pressure on a tire is reduced if its speed during braking falls disproportionately compared to those of the other wheels. Taking into account the change in rotational speed when accelerating or braking, a road surface condition and a road surface friction value when driving over the road can be determined.

In the case of vehicles in fully autonomous driving mode, it is not sufficient to determine the friction coefficient of the road surface at the time the vehicle crosses the road surface. In order to In order to be able to react to the conditions of the road surface, the associated road surface friction value must have been determined before driving over the road. If the roadway friction value is not already available at this point in time, it cannot be taken into account accordingly by a control unit. The vehicle cannot be steered in an adapted manner, taking into account the friction coefficient of the road ahead. A precise predictive determination of the coefficient of friction between the tires of the vehicle and the road surface is particularly necessary for fully autonomous driving without human control, so that the vehicle occupants are provided with a high level of safety and comfort. If the evaluation device in the autonomous driving mode determines roadway sections with a low coefficient of friction at an early stage, the control device can react accordingly and dangerous situations are avoided. The earlier and more reliably the coefficient of friction of a section of roadway is predictively determined, the greater the gain in safety and comfort for vehicle occupants. At the same time, anticipatory driving increases the vehicle's efficiency.

Proceeding from this, it is the object of the present invention to specify an improved method for predictively determining the roadway friction coefficient between the tires of a vehicle and a roadway. Advantageously, the method for the predictive determination of the road surface friction value should be improved during use. Furthermore, tire wear is to be recognized and taken into account when determining the road surface friction value.

In order to solve this problem, it is proposed that a computer-implemented method for training a neural network for determining the friction coefficient of the roadway be provided. The method is carried out on the evaluation device of a motor vehicle and includes the acquisition of a first input data set using at least one first sensor arrangement, the predictive calculation of a first road surface friction value using the neural network as a function of the first input data set, the acquisition of a second data set using at least one second sensor arrangement, that Determining the second roadway friction value, determining a weighting error of the neural network taking into account the first input data set and the second roadway friction value and that Train the neural network to minimize the weighting error.

When determining the coefficient of friction of the roadway, the coefficient of friction of a tribological system, preferably between at least one tire of the vehicle and the roadway, is to be determined. The term roadway friction value also includes the synonyms coefficient of friction, coefficient of friction and coefficient of friction. By determining the friction coefficient of the roadway, a parameter is assigned to the tribological system consisting of the roadway surface and the tires of the vehicle. This leads to better comparability of the values and can also be advantageous for data processing. Furthermore, the friction coefficient of the roadway is a dimensionless measure and relates to two bodies that are in contact with one another, in particular the tires of the vehicle and the roadway surface. The road surface friction value depends on the road surface and the tires of the vehicle. For example, if the road is wet, oily or icy, the friction coefficient between the tires and the road is lower than on a dry road. Impurities on the roadway also lead to a lower roadway friction coefficient between the two bodies. As described, it is the task of the sensor unit to examine the road surface for its road surface condition. The properties of the tire also influence the coefficient of friction. Using a stored tire model, tire properties can be taken into account when determining the condition of the road surface. Furthermore, the tire condition can be taken into account and also monitored when training the neural network.

In the computer-implemented method for training a neural network for determining the coefficient of road friction, a road ahead of the vehicle and to be traveled on is to be detected. In order to record this lane ahead of the vehicle, vehicle-specific data such as vehicle speed, steering angle, assistance system information with additional data such as the planned route guidance, GPS data, curves and elevation profile can be recorded and combined with one another. Based on the combined data, it can be predicted- which lane the vehicle is traveling on and for which lane the computer-implemented method for training the neural network for determining the friction coefficient of the lane is applied.

A plurality of input data sets are advantageously provided for training the neural network. At least one first input data set and at least one second data set are preferably available to the neural network for training.

The computer-implemented method for training the neural network for determining the coefficient of friction includes providing the first input data set. This first input data record is recorded using at least one first sensor arrangement. This at least one first sensor arrangement is advantageously aligned in the direction of travel and detects the road ahead. Multiple data can be taken into account and combined with each other during acquisition. In this way, the condition of the road surface, the outside temperature and other data can be recorded. To provide the first input data set, the at least one first sensor arrangement can preferably include a camera, an ultrasonic sensor, a radar system, a lidar system and/or the like. In addition to the visible spectral range, the first sensor arrangement can advantageously detect further emitted or reflected radiation in the non-visible spectral range, such as infrared radiation.

Advantageously, the roadway ahead can be subdivided into individual logical roadway sections for the predictive calculation of the first roadway friction value. The at least one first sensor arrangement can provide a first input data set for each logical roadway section. The logical lane sections to be detected can advantageously be chosen to be as small as desired. The smaller the individual logical lane sections are defined, the greater the sum of the logical lane sections for a lane. The sum of the first input data records for a lane also increases accordingly. Consequently, the amounts of data for the lane to be determined increase as the logical lane sections become smaller. For the evaluation device, the calculation of all first roadway friction values for smaller logical roadway sections is required more computationally intensive than for larger logical lane sections. Smaller logical lane sections have the advantage that the neural network can be trained more precisely. A further advantage is a finely resolved predictive determination of the roadway friction coefficient, which is made available to the control unit when controlling the vehicle.

In the predictive calculation of the first roadway friction value, the first roadway friction value between at least one tire of the vehicle and the roadway should be calculated using the first input data set. In general, the calculation of the road surface friction between a tire, such as B. the rubber of a tire and a driven road a complex function depending on the forces and the friction interface between the two media. Existing bumps, the distribution of these bumps, the material composition of the tire and the road, and the size of the contact surface of the two media affect the calculation and the driving characteristics of the vehicle on the road. Furthermore, the properties of the road surface and the tires are temperature-dependent. If the roadway is covered by a layer of snow or ice, the roadway friction coefficient between the vehicle's tires and the roadway changes.

In order to predictively calculate the first coefficient of friction on the roadway, the data from a number of sensors are advantageously recorded and made available to the evaluation device. The data can be merged and evaluated in the evaluation device. Especially in an autonomously driving vehicle, in which the control device takes over the control of the vehicle depending on the predictively calculated first road friction value of the evaluation device, a comfort and safety gain is provided compared to vehicles without predictive calculation of the road surface friction value.

The first input data set is made available to the evaluation device by the at least one first sensor arrangement. The evaluation device can process the input data record and/or other data in a number of processes within the neural network. For processing, the evaluation device advantageously includes tely a computing unit, such as a processor, which, thanks to its processor architecture, is designed to process several processes in parallel, such as those found in neural networks. A large number of processor cores enable efficient parallel processing. A graphics processor can be used particularly advantageously, since graphics processors are usually designed for parallel process calculations. Furthermore, the at least one computing unit includes a memory and/or a memory connection to an external memory. Advantageously, there are several memory levels, with a memory level close to the processor having a fast memory connection. The at least one computing unit accesses a computer program via an interface and uses a neural network to calculate a first road surface friction value, which can also be referred to as a calculated actual data set, taking into account the first input data set.

In one embodiment of the invention, the calculation of the first road surface friction includes a pattern recognition algorithm and/or an object recognition algorithm. For example, first input data sets of the front camera, the ultrasonic sensor of the radar system, and/or the like are generated and made available to the pattern recognition algorithm and/or the object recognition algorithm. Sensor data can preferably be evaluated in a fused manner by the evaluation device when a first input data set is generated, and the first roadway friction coefficient can thereby be calculated more precisely and predictively.

In the computer-implemented method for training the neural network for determining the coefficient of friction, a second data set for each roadway section is recorded by means of at least one second sensor arrangement. The second sensor arrangement records a second data record when driving over a predetermined, logical roadway section. The first input data records and the predictively calculated roadway friction values are already available for these logical roadway sections. The second set of data is recorded by a second sensor arrangement. The second data set can preferably be generated with a slip measurement. In a slip measurement, the slip which occurs on at least one tire of the vehicle is measured by the second sensor arrangement. In general, slip means the deviation between theoretical speed and actual speed. When determining the second coefficient of friction on the roadway, slip is to be understood as meaning the deviation in the speeds of elements in frictional contact with one another, in particular surfaces. In particular, the slip can be determined from the ratio of the speed of a driven tire that is subject to slip to the speed of a non-driven tire that is therefore rotating positively without tire slip. The second set of data can also include the tire speed of at least one tire. The second data record is made available to the evaluation device. This calculates a second roadway friction coefficient for each logical roadway section, taking into account the second data set.

A first input data record from the first sensor arrangement and a second data record from the second sensor arrangement are advantageously available for each logical roadway section. Assuming that the calculated second roadway friction value acts as a reference value and the first input data set acts as an actual value, a weighting error of the neural network is determined for each logical roadway section, taking into account the first input data set.

A machine learning method, in particular a reinforcing, supervised, machine learning method, can be used to train the neural network. With this learning method, for example, a backpropagation algorithm, which is also referred to as an error feedback algorithm, can be provided in order to train the neural network. For this purpose, the algorithm is provided with input data and an expected result. The algorithm links these two quantities together. Once the first training is complete, the algorithm can independently determine results for unknown data. The output of network is directly dependent on these weights. In the backpropagation method, an artificial neural network is trained by changing the weightings in order to minimize an error function.

Reinforcement learning can be used to work with a reward and/or punishment system, with weights being adjusted within the neural network and corresponding incentives being created as a result. The neural network should be trained to minimize the weighting error.

The training of the neural network by minimizing the weighting error is advantageously carried out by the evaluation device for each logical roadway section. By training the neural network, the accuracy of the predictively calculated road friction coefficient increases steadily. For the vehicle and the vehicle occupants, there is an increase in safety and comfort while at the same time increasing the efficiency of the vehicle.

The weighting error can advantageously be determined taking into account a difference between the first input data set and the second road surface friction value. If a weighting is changed as part of the training of the neural network, this has an immediate effect on all subsequent neurons within the neural network. This influences the output of the neural network. Mathematically, the weightings can be determined via an error function by calculating the partial derivative of the error function based on weighting parameters. The error function for calculating an error E depending on |i ₂ and ßi can look like this:

E is the calculated weighting error, from which adjustments to individual weightings in the neural network can be derived, n is the set of input data sets for which a road surface friction value is calculated. |i ₂ is the calculated second roadway friction value and is the predictively calculated first roadway friction value. So that small errors are included less intensively than large errors in the weighting error and thus in the training of the neural network, a medium quadratic error can be determined. This is also known as the mean squared error.

It can be considered advantageous that a control device is designed to determine and carry out vehicle actions, taking into account the predictively calculated coefficient of road friction, and in particular to provide a fully autonomous driving mode. In the autonomous driving mode, the control device takes over the longitudinal and lateral control of the vehicle.

Furthermore, it can be considered advantageous if, in order to determine the second road surface friction value, at least one tire of the vehicle is slipping and/or slip is actively generated and the second sensor arrangement detects this slip and a second road surface friction value is determined from the second data set.

In order to determine the coefficient of friction between the tires of a vehicle and a roadway in largely slip-free driving situations, as is the case for example when driving straight ahead without acceleration, slip is actively generated in an advantageous embodiment. In particular, a defined amount of slip, a target slip, is generated. If one or more tires are already slipping, the actively generated slip is additional slip. Otherwise, the actively generated slip is the total slip present. This can be made possible by accelerating each tire independently. Acceleration can advantageously be understood to mean a positive acceleration, also referred to as an acceleration process, or a negative acceleration, also referred to as a braking process.

A positive acceleration to generate slip causes a positive change in the speed of the vehicle in the direction of travel. A drive torque of a motor rotationally accelerates tires in order to positively change the speed of the vehicle. There is a minimum acceleration process at which drive slip will definitely occur. In the case of negative acceleration, the slip-generating braking process, in contrast to positive acceleration to generate slip causes a negative change in the speed of the vehicle in the direction of travel. Brakes can be used to negatively change the amount of speed of the vehicle. A minimum braking process is carried out, during which brake slip occurs in a secured manner.

Advantageously, a precise determination of the friction between the tire of the vehicle and the road surface can be made possible without the driver of the vehicle being able to perceive the active generation of the slip. In order to actively and specifically generate slip when an acceleration process is initiated, the acceleration that generates the slip must act on at least one tire. For the controlled generation of slip during acceleration, it is therefore advantageous if each tire is designed so that it can be accelerated separately and independently of other tires. With the electronic stability control, each tire can be individually accelerated negatively by means of electronic brake force distribution. Separation of all tire brake circuits is required. Similarly, with a controlled redistribution of drive torque through individual torque allocation, at least one tire is positively accelerated. In an advantageous embodiment, the acceleration of the individual tires can counteract a yawing of the vehicle.

A tire condition of at least one tire is preferably determined taking into account the first input data set and the determined second roadway friction coefficient, in particular taking into account the weighting error. If the tires of the vehicle are worn during the service life of the vehicle and the properties of the tire rubber compound change over time, the trained neural network can record this continuous change over the service life. A change in the tire properties usually results in a change in the first input data set and in the second coefficient of friction determined on the road. This results in a weighting error, which in turn is used to train the neural network. If the weighting error is stored over the useful life of the vehicle, it can be determined from the history of the weighting error, a change in the tire properties can be determined. It is known, for example, which second roadway friction coefficient tires deliver given known first input data sets. If the calculated second road surface friction coefficient changes with the same first input data set, a change in the tire properties can be determined. The control device, which uses the first predictively calculated roadway friction coefficient to control the vehicle, takes into account a change in the tire properties using the computer-implemented method for training the neural network.

A threshold value for the weighting error is preferably defined. If the threshold value is exceeded, this is detected by the evaluation device. When the threshold value in the trained neural network is exceeded, the condition of at least one tire is advantageously classified as critical. The weighting error of the neural network can be minimized while the vehicle is being driven and the computer-implemented method carried out in the process. If the evaluation device of the trained neural network detects a high weighting error that exceeds the established threshold value, the evaluation device can determine a probability of a critical tire condition, taking into account all first input data sets and second data sets. A critical tire condition of at least one tire can be, for example, a tire pressure drop as a result of tire damage. The detection of a critical tire condition is advantageously forwarded to the control device, so that it can adjust vehicle parameters accordingly. The vehicle occupants are preferably informed of a critical tire condition. In this embodiment of the invention, the safety of the vehicle and the vehicle occupants is further increased. At the same time, there is no need for complex tire sensors, which increases the vehicle's economy.

Furthermore, the invention relates to an evaluation device for a vehicle, which is designed to carry out the method described. The computer-implemented method for training the neural network is preferably executed on the evaluation device. The evaluation device is also designed to calculate individual processes within a neural network. The input data is a first input data set and a second data set. The neural network is designed to predictively calculate a road surface friction value as a function of this first sensor data and to determine a weighting error taking into account the second data set. The evaluation device preferably has an input interface in order to receive data from the at least one first sensor arrangement, preferably from the camera, radar, lidar, infrared and/or ultrasonic sensors and/or the like. Furthermore, data from the second sensor arrangement, in particular slip and/or roadway friction value data, can be transmitted. According to the invention, the input interface can connect the first and the second sensor arrangement to the evaluation device for data exchange. The exchange can also be wired or wireless. In the case of the computer-implemented method for training an artificial neural network, the artificial neural network can be executed on the evaluation device.

The invention also relates to a computer program for a vehicle. This is designed to carry out the method described. Computer programs usually include a sequence of instructions that cause hardware to carry out a specific method. When the computer program in question is used on a computer, it executes predefined command sequences. The computer program is preferably designed to carry out the predictive calculation of the roadway friction coefficient using the neural network. Furthermore, the computer-implemented method for training a neural network can be carried out.

In addition, the invention relates to a machine-readable storage medium for a vehicle. This is designed to be practicable for carrying out the method described. The machine-readable storage medium may include one or a combination of volatile memory elements and non-volatile memory elements. Volatile storage elements may include, for example, random access memory and/or erasable programmable read only memory included. Non-volatile memory elements may include, for example, hard drives and/or compact disk read-only memories. Additionally, storage may include electronic, magnetic, optical, or other types of storage media. The machine-readable storage medium may include one or more separate computer programs. The machine-readable storage medium is advantageously connected to the evaluation device.

In addition, the invention relates to a control device for a vehicle. This is designed to carry out the method described.

This control device is also designed to control the vehicle. It can be connected to the evaluation device and take into account the determined road surface friction data when controlling the vehicle.

In addition, the invention relates to a vehicle. This is designed to be able to carry out the method described. A vehicle which is designed to carry out the method described can advantageously drive fully autonomously. The decision-making process of the vehicle's control unit is based on the predictively calculated first road friction value. Measurements from sensors serve as the input variable. The main task of the sensors is to record the vehicle environment in every driving situation and to make the data available to the evaluation device. The vehicle can be optimally controlled by the control unit, taking into account the driving dynamics limits, the vehicle-specific data and the predictively calculated road surface friction value.

This not only increases the safety and comfort of the vehicle occupants, but also the economy of the vehicle.

In addition, the invention relates to a neural network. This is designed to carry out the method described. The calculation of individual processes of the neural network is carried out by the evaluation device. Measurements from sensors serve as input data. The sensors record the vehicle surroundings and tire slip in every driving situation. Further advantages, features and details of the invention result from the following description of exemplary embodiments and figures. show:

1 shows an exemplary embodiment of a vehicle according to the invention,

2 shows an exemplary embodiment of a neural network according to the invention and

3 shows an exemplary embodiment of a method according to the invention.

A vehicle 100 with an evaluation device 120 is shown in FIG. 1 . Vehicle 100 is advantageously an autonomous vehicle. Evaluation device 120 accesses a computer program 130 in order to predictively calculate a first road surface friction value 242 from a first input data set 210 via a neural network 200 . This predictively calculated first roadway friction value 242 is made available to a control device 140 .

First input data set 210 is provided by at least one first sensor arrangement 102, a vehicle communication interface 104, a vehicle sensor 106 and/or an environment sensor system 108. The at least one first sensor arrangement 102 advantageously includes a camera and is aimed at the roadway ahead.

The vehicle communication interface 104 is designed to record vehicle-external data and forward it to the evaluation device. In particular, data from other vehicles can be acquired by the vehicle communication interface 104 . Furthermore, the vehicle communication interface can call up vehicle-external data. This data is preferably stored on external servers.

The vehicle sensor 106 collects vehicle-specific data. The vehicle sensor

106 can include a number of sensors for this purpose. Vehicle sensors 106 are preferred already integrated in individual assemblies such as the drive train, transmission, engine, exhaust system, steering and electrical system of the vehicle. The vehicle-specific data are made available to the evaluation device 120 in order to determine the dynamic driving limits of the vehicle there. Furthermore, the data can be advantageous when calculating the predictive road surface friction coefficient.

Surroundings sensor system 108 captures surrounding information outside of the vehicle.

For this purpose, light sensors, moisture sensors and temperature sensors can measure, for example, the brightness, the outside temperature and/or the air humidity in the vehicle environment and forward it to the evaluation device 120 .

To train the neural network 200, a further, second data set is recorded using at least one second sensor arrangement 110.

This second sensor arrangement 110 is preferably embodied in multiple configurations, with at least one second sensor arrangement 110 being assigned to each tire 112 of the vehicle. The second sensor arrangement 110 is advantageously embodied as an active tire speed sensor in order to measure a tire speed and in particular a slip on the respectively associated tire. The second sensor arrangement 110 provides the determined data to the evaluation device 120 . Both forward and backward movements of the tire are advantageously detected via the second sensor arrangement 110 .

Evaluation device 120 includes a processing unit 122, a machine-readable storage medium 124 and at least one interface 126. Processing unit 122 includes a central processing unit, such as a processor unit and/or a graphics processing unit, the central processing unit and the graphics processing unit including a plurality of processing cores. Furthermore, the processing unit can have its own memory, such as a cache memory. Advantageously, individual processes are divided among several computing cores in order to process processes efficiently. The machine-readable storage medium 124 can be embodied as a volatile and/or non-volatile memory. The machine-readable storage medium is advantageously embodied as a temperature, shock and vibration-resistant memory in order to meet the requirements of a vehicle. The evaluation device accesses the computer program 130 and the neural network 200 via an interface 126 . The computer-implemented method for training the neural network is executed on the computing unit.

The vehicle also includes a control device 140 for controlling the vehicle 100. Vehicle actions are stored on the control device 140 for each input data set and in particular for each road surface friction value. Control device 140 selects a vehicle action, taking into account driving dynamics limits, the vehicle-specific data and in particular the roadway friction coefficient predictively calculated by neural network 200 . By training the neural network 200, the vehicle can be optimally controlled by the control unit 140.

FIG. 2 schematically shows a multi-layer neural network 200 comprising a number of interconnected neurons. The output of a neuron is always the input of the neuron connected to it. The input layer 220 is followed by a plurality of intermediate layers 230. Beginning with the input layer 220, information flows through one or more intermediate layers 230 to the output layer 240.

In the example mentioned, the first input data record 210 comprises data from one or more sensor arrangements. This data includes vehicle data 212, such as engine speed and speed of the vehicle, vehicle environment data 214, such as outside temperature and outside air humidity in the immediate vicinity of the vehicle. Furthermore, the input data set 210 includes data from a first sensor arrangement 216, it being possible for the first sensor arrangement 216 in turn to include data from one or more sensors. Camera, ultrasound, radar, lidar data and/or the like are advantageously available as data from the first sensor arrangement. Furthermore, external data 218, such as GPS or data is provided to the vehicle-to-vehicle communication interface 214 .

This first input data record 210 is provided to the input layer 220 of the neural network 200 and is received by neurons in this layer. In a further step, the data is forwarded to neurons in the intermediate layer. In this case, neurons of the input layer 220 pass on the data of the first input data record 210 to the neurons of the first intermediate layer 230 . When forwarding the data in the neural network 200, a weighting is carried out. In principle, the number of intermediate layers 230 in the artificial neural network 200 is unlimited. In practice, each intermediate layer 230 also consumes computing power that is necessary for the operation of the neural network.

The output layer 240 lies after the intermediate layers 230 of the neural network 200 in the processing of the information. The output layer represents the endpoint of the information flow in a classic artificial neural network 200 and contains the result of the information processing by the network.

When a neural network is initialized, all weightings are chosen randomly within a predetermined range of values. The result in the output layer of an untrained neural network can be a purely random value. Training the neural network improves the quality of the output layer results by minimizing the weighting error.

A supervised machine learning technique may be used to train the neural network 200 . Taking into account the first input data record 210, the information as to which roadway friction value was predictively calculated is output in the output layer 240 of the neural network 200. This information is compared with the result from the second data set, the second road surface friction value 244, and the deviation is then given as a weight error 246 is propagated backwards into the network to adjust the neurons accordingly. When adjusting, the weights between the individual neurons are adjusted. In this case, a combination of first input data sets 210 and second data sets 244 are compared as data pairs. This comparison results in weighting errors 246 which, during training after a number of training cycles, can predictively calculate a road surface friction value for all first input data sets 210 and also for new, first input data sets 210 that have not yet been known.

A back propagation method can advantageously be carried out as a method for training and learning the artificial neural network. According to the backpropagation method, the weights within the neural network are changed depending on their influence on the error. This guarantees an approximation to the desired output in an iterative process when the method is carried out again. In order to be able to display both positive and negative deviations at the same time, the mean value of the squared deviation (Squared Mean Error SME) or another error function can be determined.

FIG. 3 shows a flowchart for training a neural network 200. In a first step S1, a roadway section on a roadway ahead of the vehicle is detected. In a step S10, a first input data record 210 is provided for this detected roadway section. This first input data set 210 advantageously includes data from a number of sensors for providing data. In step S12 data of the vehicle communication interface, in step S14 vehicle data and in step S16 vehicle surroundings data are provided. These data together form the first input data set 210. Taking this first input data set 210 into account, a first roadway friction value 242 is predictively calculated in step S40 by means of the neural network 200. To train the neural network, a second sensor arrangement 110 provides a further, second data set 213 as a comparison value, which can also be referred to as a target value or reference value. The roadway section detected from S1, for which a first predictive roadway friction value is present, is detected a second time by means of the second sensor arrangement when the vehicle drives over it. According to one embodiment of the invention, slip can be actively generated on at least one tire in a further step S20 in order to determine the roadway friction value of this roadway section.

The actively generated slip is measured by the second sensor arrangement 110 on at least one tire 112 and the second data set is provided in S22. In S24, a second roadway friction value 242 for the roadway section can be determined taking into account the second data set. First input data record 210 and second roadway friction value 244 are available to evaluation device 120 .

The second roadway friction value 244 and the first input data set 210 are compared and a weighting error within the neural network 200 is calculated in S44. In S46, the weight within the neural network is corrected. In doing so, the weighting error can be changed. This correction trains the neural network and improves the quality of the predictive determination of the first predictive road surface friction coefficient 242 . In S30 the vehicle 100 is controlled by the control device 140 taking into account the first predictive roadway friction value 242 .

The flowchart from FIG. 3 can be repeated in an iterative process for each road section and the first input data record 210 recorded for it. This iterative process trains the neural network and improves the predictive determination of the road friction coefficient in the vehicle. reference sign

100 vehicle

102 First sensor arrangement

104 vehicle communication interface

106 vehicle sensor

108 environment sensors

110 Second sensor arrangement

112 tires

120 evaluation device

122 processor

124 Machine-Readable Storage Medium

126 interface

130 computer program

140 controller

200 Neural Network

210 First input data record

212 data of the first sensor arrangement

213 Second record

214 Vehicle Communication Interface Data

216 vehicle data

218 vehicle environment data

220 input layer

230 intermediate layers

240 output layer

242 First predictive road surface friction coefficient

244 Second road surface friction value

246 weight error

S1 step

S10 step

S12 step

S14 step

S16 step S18 step

S20 step

S22 step

S24 step

S30 step

S40 step

S42 step

S44 step

S46 step

Claims

Computer-implemented method for training a neural network (200) for determining the coefficient of friction of a motor vehicle tire, the method being executed on an evaluation device (120) of a motor vehicle (100), comprising the following steps:

- detecting a first input data set (210) by means of at least one first sensor arrangement (102),

- Predictive calculation of a first road surface friction coefficient (242) using the neural network (200) as a function of the first input data set (210),

- detecting a second data set (213) by means of at least one second sensor arrangement (110) and determining a second roadway friction value (244),

- determining at least one weighting error (246) of the neural network (200) taking into account the first roadway friction value (242) and the second roadway friction value (244), and

- training the neural network (200) to minimize the at least one weighting error (246). Method according to Claim 1, characterized in that the weighting error (246) is determined taking into account a deviation of the first input data set (210) and the second road surface friction value (244). Method according to one of the preceding claims, characterized in that a control device (140) taking into account the predictively calculated first road surface friction value (242) determines vehicle actions and an autonomous driving mode is provided. Method according to one of the preceding claims, characterized in that to determine the second road surface friction value (244) there is slip on at least one tire (112) of the vehicle (100) and/or slip is actively generated and the second sensor arrangement (110) detects this

23 Slip is detected and the second roadway friction coefficient (244) is determined taking this data into account. Method according to one of the preceding claims, characterized in that a tire condition of at least one tire is determined taking into account the first input data set (210) and the determined second roadway friction value (244), in particular taking into account the weighting error (246). Method according to one of the preceding claims, characterized in that a threshold value for the weighting error (246) is established and a deviation from this threshold value is detected by the evaluation device (120). Evaluation device for a vehicle (100), characterized in that the method according to one of Claims 1 to 6 can be carried out. Machine-readable storage medium on which a computer program (130) for a vehicle (100) for carrying out the method according to one of Claims 1 to 6 is stored. Control device (140) for a vehicle (100), characterized in that the method according to one of Claims 1 to 6 can be carried out. Vehicle (100) with an evaluation device (120), a control device (140), a computer program (130) and a machine-readable storage medium (124), characterized in that the method according to one of Claims 1 to 6 can be carried out.