Method for estimating a road friction coefficient potential for at least one road section

Abstract

The invention relates to a method for estimating road friction potentials for at least one road section, wherein a road friction potential of a road surface in the road section is measured. The invention is characterized in that environmental information is assigned to the measurement and sent together with the road friction values ​​to a backend, in which a connection between the environmental information and the road friction potentials is learned by means of machine learning, so that expected road friction potentials can be estimated. An interval with a minimum and maximum road friction value is provided as an output value for the road friction potential.

Description

The invention relates to a method for estimating a road coefficient of friction potential for at least one road section.DE 10 2015 000 394 A1 discloses a method for generating a digital map, wherein location-dependent estimated values of a physical measured variable are recorded in the digital map. Measurement data relating to the road condition and, for example, also data relating to the temperature or the road coefficient of friction can be collected via a fleet of vehicles. By means of mobile radio, these data can be transmitted to a server and processed from there. A digital map with the associated collected data may be transmitted to the vehicles so that a road condition may be provided to the driver assistance system depending on the location on the map. The driver assistance system can switch functions on or off depending on this, and thus enter into the environmental conditions.U.S. Pat. No. 2020 013 99 76 A1 describes a method for indicating a probability of a prevailing road coefficient of friction at a specific location or a specific location. A plurality of vehicles each calculate the road coefficient of friction between tire and road surface at this location and send it to a backend or a central processing unit. There, the various coefficients of friction calculated by the vehicles at this location are sorted into different pots, each pot being assigned a coefficient of friction range. A first pot is assigned a coefficient of friction range of 0 to 0.05, a second pot is assigned a coefficient of friction range of 0.05 to 0.10, a third pot is assigned a coefficient of friction range of 0.10 to 0.15, etc. It is assumed that the pot in which most of the coefficients of friction calculated by the vehicles "lie" reflects the actual or closest coefficient of friction prevailing between the tire and the road surface.It is disadvantageous in the prior art that only already measured data are fed into the system and the calculation takes place in each case in the vehicle. In particular for the adaptation of functions provided by the driver assistance system, however, a situation in the future is advantageous in order to be able to negotiate economically, sustainably and in anticipation.The object of the present invention is to specify a method for estimating track coefficient potentials which can be expected in a geolocalized manner.According to the invention, this object is achieved by a method for estimating road coefficient of friction potentials having the features of claim 1, and here in particular in the characterizing part of claim 1. Advantageous embodiments and refinements emerge from the dependent claims.In the core of the method according to the invention, the road coefficient of friction potential of a road surface is measured by a vehicle (in the case of full braking or maximum acceleration until the wheels spin), in particular with a driver assistance system, such as a slip control system, an anti-lock braking system, or the like, wherein environmental information is assigned to each measurement. In addition, driving parameters signals from other vehicles (speed, wheel speeds, acceleration, brakes, steering angle, caster, yaw, suspension travel of the chassis... ) are collected at the same location and similar time and sent to a backend, wherein a connection between the environmental information, the measurement of the road friction value potential and the measurements of other vehicles is learned by means of machine learning, so that expectable road friction value potentials can be estimated or predicted at another road surface and / or at another location. According to the invention, an interval is provided as an output value for the road coefficient of friction potential, which interval makes available a minimum and a maximum possible value of the road coefficient of friction potential depending on the location. Uncertainty in the prediction of the values can therefore be quantified and better assessed. In the systems known in the prior art, however, only the value of the friction is used as a so-called point estimator. With the method according to the invention, it can be advantageously ensured how well the estimate approximates the actual value.The method for estimating expectable road friction value potentials for at least one road section makes it possible, for example, to generate a current road friction value map, and to specify a change in the road friction values how the data of the road friction value map behave for a certain period of time. Such a map may be advantageous, for example, for road management, in particular for winter management of the road operators in order to ensure infrastructure safety. In particular, however, the functionality of vehicle assistance systems, in particular for at least partially autonomous driving, is increased. For example, an availability of the levels 3 to 5 driving functions can be significantly increased as a result, since precise information regarding the surface state and future values are available, as a result of which an early shutdown of the system can be avoided. The method can therefore advantageously reduce a current restriction of use of automated driving, as a result of which the automated functions can be used in significantly more situations and safety-critical infrastructure can be displayed early. Infrastructure operators can also advantageously maintain roads better, faster and at low cost, since conclusions can be drawn about the road state by predicting the coefficient of friction of the road over time. Consequently, previous knowledge gaps can be concluded by the high data density. This offers advantages in particular with respect to stationary measurements known to date.By using machine learning, road coefficient of friction potentials can be determined which represent or are critical for specific spatial conditions of different aggregate states, such as moisture or ice, for example. More precise estimates can thereby be made for the condition of the roadway. Consequently, both vehicle-external applications and vehicle-internal applications, such as driving assistance systems and / or autonomous driving functions, can react in an adapted manner in order to adapt the driving behavior to the state, i.e. to the coefficient of friction, of the roadway.The actual detection of the road coefficient of friction potentials on the basis of the available data can be carried out, for example, only in the backend. The detection can therefore be dependent on different requirements, such as a transmission capacity and an online computing capacity. On the basis of the collected data, a machine learning model can be trained which outputs a prediction interval by using a conformal prediction, for example, and thus predicts an uncertainty of the machine learning model. The data is output as data of one interval as described above.Training of the model for machine learning for the first time can be carried out on development vehicles with corresponding sensor systems or driver assistance systems, which can measure the coefficient of friction of the roadway. The backend is supplied with such information or data in particular by a plurality of different vehicles, so that an environment map can be created.Preferably, the track coefficient of friction potential that can be expected can be determined by means of a conformal prediction algorithm, wherein an uncertainty of the data of the machine learning is determined. Thus, the current road friction value and a temporal development of the road friction value at specific geopositions can be estimated or predicted experimentally and very accurately by a high data density. In particular, this makes use of the output of a prediction interval that qualifies the limits of machine learning. For example, the conformal prediction used can be approaches of split conformal prediction, full conformal prediction or a method which lies between the two approaches, such as a cross conformal prediction or jacknife.According to a very advantageous development of the concept, it can be provided that the environmental information items are geopositions and times. This allows the data to be assigned to a specific location at a specific time. The environmental information can originate, for example, from a measurement sensor system on the vehicle itself. Alternatively or additionally, environmental information from weather stations or from measurement data of the road operator can be used. The environmental information can additionally include states such as sun intensity, precipitation quantity, snow quantity and in particular air temperature and / or road surface temperature. For this purpose, sensors and / or cameras and a wireless connection, for example via mobile radio, can be used to communicate the data to a backend server. A total of all these data, wherein data can also be included in the German weather service, can be collected anonymously in the backend.According to an advantageous embodiment, it can be provided that further data are measured by means of an on-board sensor system of a vehicle, which data are sent to the backend together with the environmental information and the surface temperatures. This may provide improved geographic coverage of weather data. In particular, the sensor system can be located only at a very small distance from the road surface, so that in particular the data of the road surface can be specified very accurately.According to a very advantageous development of the concept, it can be provided that data of a road operator is used and likewise sent to the backend. In this case, it is likewise possible to use weather data from stationary measurements that are accessible to the public. All of these data increase the data density so that the system can operate even more accurately.According to a further advantageous embodiment, it can be provided that the conformal prediction algorithm aggregates a set of data and divides it into training set, test set and calibration set with the same distribution. In this case, according to a very advantageous development of the invention, it can be provided that the machine learning is carried out on the basis of the training set.According to an advantageous embodiment, it can be provided that a score function is defined which defines an unusualness of a data record with respect to a previous data record. Calibration scores can be calculated using the calibration amount defined above. In this case, a defined quantile function on the calibration score can be calculated. The quantile can be used in particular to calculate prediction intervals for new examples, so that in particular marginal coverage is fulfilled. In other words, this means the calculation of a probability, wherein the probability that the prediction intervals contain the correct label can be indicated by 1-alpha, wherein alpha is present in [0,1].Such approaches result in particular in the advantages that it is only necessary to subdivide it into training and test sets. In this case, for example, a AI model is trained on the training data set. A calibration data set can be dispensed with, wherein the selection of this data set is difficult. Furthermore, learning of a scoring function can be dispensed with, since this is trained or can be selected.A further advantageous embodiment is characterized in that a map is created with information about road friction values, data of a fleet of vehicles being used. Furthermore, the created map can be provided to each vehicle in the fleet, wherein these vehicles comprise, in particular, subsystems in order to transmit or receive the data contributing to the ascertainment of the map.According to a very advantageous development of the invention, it can be provided that an autonomous driving function is enabled or not depending on the current and future road friction values determined. Thus, the autonomous driving functions can be used over a larger area and have to be deactivated less frequently. This includes in particular the availability of the levels 3 to 5 driving functions, which is increased substantially. Thus, the information can be used to extend autonomous driving. Furthermore, infrastructure operators can use the information to better maintain their road network.Further advantageous embodiments of the method according to the invention for predicting expectable road friction values also result from the exemplary embodiment which is described in more detail below with reference to the single FIG.. The single FIGURE shows a schematic illustration of the sequence of the method.A possible sequence of the method is schematically illustrated in the FIG.. In a step marked 1, on-board data can be measured and stored by a plurality of vehicles, which are measured, for example, via the respective sensor system on the vehicle or by driving assistance systems. Via a communication module, this data can be combined with offboard data which is determined in a step denoted by 2. These so-called offboard data can contain data of environmental states and geopositions that are not measured by the vehicle, characterized by step 3. Back to step 2, a road coefficient of friction potential can thereby be determined, wherein the data can be determined via artificial intelligence.For example, environmental variables can be aggregated at a time t in the backend and the behavior of the road coefficient of friction potentials for future times t+n, where n is a natural number, can be predicted using a AI model. In this case, for example, the following predictions can be made: In the case of a large interval, an AI model is uncertain; in this case, the model can make a decision only poorly. To improve this, more data should be collected or data homogenized. At a small interval, the AI model is very safe, and a very good approximation of the reality can be made.In other words, a connection of environmental variables to the road coefficient of friction potential is accordingly established with the AI model, wherein an uncertainty of the model can be quantified with a conformal prediction. In summary, therefore, a method for ascertaining, providing and predicting geolocalized road coefficient of friction potentials can be provided, wherein a coefficient of friction potential at a location of the road segment is output not, i.e., an absolute value, but rather as an interval having a minimum and a maximum coefficient of friction.References included in the specificationThis list of documents cited by the applicant has been produced in an automated manner and is only included for the better information of the reader. The list is not part of the German patent application or utility model application. The DPMA does not take any adhesion for any faults or omissions.Patent Literature citedDE 10 2015 000 394 A1

[0002] US 2020 013 99 76 A1

[0003]

Claims

Method for estimating road coefficient of friction potentials for at least one road section, wherein a road coefficient of friction potential of a road surface in the road section is measured, characterized in that environmental information is assigned to the measurement and sent together with the road coefficient of friction potential to a backend, in which a connection between the environmental information and the road coefficient of friction potentials is learned by means of machine learning, such that expectable road coefficient of friction potentials can be estimated, wherein an interval having a minimum and maximum road coefficient of friction is provided as an output value for the road coefficient of friction potential.Method according to Claim 1, characterized in that the track coefficient potentials which can be expected for future points in time are determined by means of a conformal prediction algorithm, wherein an uncertainty of the data of the machine learning is determined.Method according to Claim 1 or 2, characterized in that the items of environmental information are geopositions and times.Method according to Claim 1, 2 or 3, characterized in that further data are measured by means of an on-board sensor system of a vehicle, said data being sent to the backend together with the environmental information and the surface temperatures.Method according to one of claims 1 to 4, characterised in that data of a road operator is used and is likewise sent to the backend.Method according to one of Claims 2 to 5, characterized in that the conformal prediction algorithm aggregates a set of data and divides it into training set, test set and calibration set with the same distribution.The method of claim 6, characterized in that the machine learning is performed based on the training set.Method according to one of Claims 6 or 7, characterized in that a score function is defined which defines an unusualness of a data record with respect to a previous data record.Method according to one of Claims 1 to 8, characterized in that a map is created with information on road coefficient of friction potentials, data of a vehicle fleet being used.Method according to one of Claims 1 to 9, characterized in that an autonomous driving function is enabled or not depending on the current and future road coefficient of friction potential determined.

Images