Method for classifying the surface condition of a roadway, method for training a classification model, computer program product, and motor vehicle

Abstract

A method for classifying the surface condition of a roadway using a motor vehicle is described, wherein the roadway is driven on by the motor vehicle, wherein the motor vehicle records an instantaneous speed, wherein the motor vehicle records sensor data from at least one steering angle sensor and / or at least two wheel speed sensors arranged on different wheels of the motor vehicle, feeds the sensor data to a classification model trained with artificial intelligence and applies the classification model, wherein the classification model assigns a surface condition class to the roadway.

Description

[0001] A method for classifying a road surface condition, a method for training a classification model, a computer program product and a motor vehicle are described.

[0002] Methods for classifying the surface condition of a roadway, methods for training a classification model, computer program products and motor vehicles of the type mentioned above are known in the prior art.

[0003] The surface condition of a road is crucial for safety and comfort while driving on it. This includes, in particular, damage such as potholes and longitudinal undulations, i.e., transverse waves. Undulation is a fundamental measure for assessing the quality of road surfaces and, for example, deriving appropriate maintenance measures. In addition to well-known standard procedures, such as those described in ISO 8608, which require highly accurate elevation profiles recorded using special measuring vehicles, approaches now exist that achieve comparably accurate results based solely on data from the suspension sensors of production vehicles.

[0004] From DE 10 2011 120 022 A1, a method for classifying the surface profile of a roadway by means of a vehicle is known, comprising the following steps: a) determining the instantaneous speed of the vehicle and an associated level position of at least one vehicle wheel of the vehicle while driving on the roadway to be classified; b) calculating a roadway wheel elevation profile from the instantaneous speed and the associated level position; c) calculating a density spectrum of the wheel elevation profile; d) determining a modified density spectrum by masking out predetermined frequency ranges from the density spectrum calculated in step c); e) classifying the surface profile of the roadway by comparing the modified density spectrum with at least one reference spectrum.

[0005] However, the known method, in which a wheel elevation profile is recorded that simulates the height profile of the road surface, requires a chassis equipped with sensors so that spring strokes can be recorded.

[0006] The task therefore arises of further developing methods for classifying the surface condition of a roadway, methods for training a classification model, computer program products and motor vehicles of the type mentioned above in such a way that the surface condition of a roadway can be determined without spring stroke measuring sensors.

[0007] The object is achieved by a method for classifying a surface condition of a roadway according to claim 1, a method for training a classification model according to the independent claim 4, a computer program product according to the independent claim 9 and a motor vehicle according to the independent claim 10. Further embodiments and developments are the subject of the dependent claims.

[0008] A method is described for classifying a surface condition of a roadway by means of a motor vehicle, wherein the roadway is driven on by the motor vehicle, wherein the motor vehicle detects an instantaneous speed, wherein the motor vehicle detects sensor data from at least one steering angle sensor and / or at least two wheel speed sensors arranged on different wheels of the motor vehicle, feeds the sensor data to a classification model trained with artificial intelligence and applies the classification model, wherein the classification model assigns a surface condition class to the roadway.

[0009] Surprisingly, it has been shown that not only the spring travel signal from a spring travel sensor reflects road surface waviness in a measurable manner, but also other signals, particularly signals from steering angle and wheel speed sensors. However, it is more difficult to derive surface condition classes using analytical methods for such sensors, as the waviness of the road surface has a more diffuse effect on the vehicle sensor signals than, for example, on the spring travel sensor.

[0010] The physical relationships are fundamentally deterministic. For example, if a motor vehicle is moving at a speed v in the longitudinal direction, this speed is relatively constant due to the mass of the vehicle. The speed of the vehicle's wheels, on the other hand, is less constant because they sometimes have to cover different distances, particularly longer distances, due to the undulations of the road surface, while the chassis of the vehicle largely maintains its speed relative to the road. On a completely flat, undulating road, the vehicle speed and the wheel speeds are linearly dependent with a constant factor for all four wheels, whereby the factor depends primarily on the respective wheel size and the direction of travel.For example, the factor changes when cornering or when tires are at different levels of wear, when tires have different air pressures, and / or when wheel slippage occurs. This factor also changes when the wheels have to travel longer distances relative to each other than the vehicle chassis due to undulations in the road surface.

[0011] Steering angles are also affected by road surface undulation, as sudden changes in direction in the roadway exert forces on the vehicle's steering, which in turn result in small changes in the steering angle. Some of these changes can be felt through the steering wheel.

[0012] The surface quality class can be a discrete class, such as ISO 8606, or a value of a continuous classification, or a continuous quality number, such as the IRI (International Roughness Index), which generally represents a quality value or number that characterizes the longitudinal waviness of the road surface. The objective of the method is to assess the road surface waviness for one or more road sections.

[0013] However, the deflections in the corresponding wheel speed and steering angle sensors caused by the road bumps are relatively small overall, so the signal-to-noise ratio is not particularly favorable. Furthermore, different individual vehicles of a particular model exhibit certain differences, such as varying degrees of tire wear, different air pressures, different load levels, different road conditions, and so on, so that the effects of driving over the same bump with different vehicles of the same type on their sensors vary.

[0014] Therefore, the use of classification models based on artificial intelligence can achieve better results in the classification of road surfaces than conventional deterministic methods using classical algorithms.

[0015] A corresponding artificial intelligence-based classification model is an algorithm that uses machine learning to assign input data to predefined surface texture classes. The model consists of an input layer for receiving the data, one or more processing layers for data transformation, and an output layer for determining class membership.

[0016] The processing layer(s) transform the input data using trainable parameters and nonlinear functions. The output layer converts the processed data into probabilities for membership in the various surface texture classes.

[0017] This method makes it possible to classify road surfaces even using motor vehicles that are not equipped with suspension travel sensors. For vehicles equipped with the appropriate technology, classification using this method can be performed in parallel with, or alternatively to, classification based on suspension travel sensors.

[0018] In a first further embodiment, it is provided that the sensor data are pre-processed after acquisition by removing incomplete and / or erroneous data.

[0019] In this way, the evaluation quality can be improved.

[0020] In a further further embodiment, it is provided that the classification model was trained on the basis of data from at least one motor vehicle which is equipped with suspension travel sensors, at least one steering angle sensor and / or at least two wheel speed sensors arranged on different wheels.

[0021] According to a further development, the motor vehicle in question can, in particular, be a motor vehicle of the same type or model as the one in which the classification model will be used. In this case, the behavior of the motor vehicle equipped with the suspension travel sensor is generally relatively similar to the motor vehicle not equipped with it, as they have similar geometries, weights, chassis designs, and the like. Furthermore, they are relatively often equipped with the same sensor types, which results in better comparability of the sensor data.

[0022] Steering angle sensors and wheel speed sensors have shown visible signals in tests that are attributable to road irregularities such as bumps and potholes and are therefore particularly suitable for the method.

[0023] A first independent subject matter relates to a method for training a classification model for a method according to one of the preceding claims, wherein with the aid of at least one motor vehicle which is equipped on the one hand with spring travel sensors and on the other hand with at least one steering angle sensor and / or at least two wheel speed sensors arranged on different wheels, a roadway is driven on and the sensor data of the sensors are recorded, wherein a surface quality class is determined on the basis of the spring travel sensor data, wherein the classification model is trained on the basis of the data from steering angle sensors and / or the data from the wheel speed sensors on the one hand and the surface quality class on the other hand.

[0024] This method is superior to other methods that would require the use of different vehicles, e.g. those with high-precision measuring devices and others that lack the appropriate steering angle and wheel speed sensors, because the data correlates better with each other. This is because it is very difficult to precisely locate the vehicles with the relevant steering angle and wheel speed sensors along and across the road. By using one and the same vehicle to collect the necessary training data, this problem cannot arise because the data basis for training always relates to the exact same lane, resulting in higher quality data. In addition, the corresponding data is recorded almost or entirely simultaneously, meaning that the road surface cannot have changed in the meantime.

[0025] Furthermore, very large amounts of learning data can be generated this way, since the entire vehicle fleet, equipped with the appropriate sensors and suspension travel sensors, is suitable for generating training data. This is especially true for better-equipped model variants with high-quality chassis. With the amount of learning data, the quality of the classification model increases, allowing for better predictions.

[0026] Training a classification model with artificial intelligence requires several steps. These steps include data preparation, model selection and training, and model evaluation and fine-tuning.

[0027] First, the sensor data from suspension travel sensors, steering angle sensor(s), and / or wheel speed sensors, as well as any other sensors, must be collected and preprocessed. The various types of sensor data may need to be converted into a uniform format. The data then needs to be normalized or standardized to ensure that all inputs fall within the same value range.

[0028] In the next step, relevant features are extracted from the sensor data. This can be done by calculating statistical metrics such as mean and standard deviation, or by applying transformations such as transforming from a time domain to a spatial domain. This identifies meaningful features that help the classification algorithm distinguish between the different classes.

[0029] The classification model is then trained according to the respective training principles. For an artificial neural network, for example, training is performed using an algorithm called backpropagation. In this process, an error resulting from the difference between the predicted and actual classifications is propagated backward through the neural network. The weights of the connections between the neurons are adjusted to minimize the error. This process is performed iteratively over many training cycles until the network achieves a satisfactory classification accuracy.

[0030] After training, the model is evaluated on a separate test dataset to verify its performance. Key metrics for evaluating model performance are accuracy and precision. Based on the evaluation results, the model can be further optimized by adjusting hyperparameters such as the learning rate, the number of hidden layers, or the number of neurons.

[0031] In a first further embodiment, it is provided that the data from steering angle sensors and / or the data from wheel speed sensors are cleaned.

[0032] This allows for more efficient classification models to be achieved.

[0033] In a further refinement, it is provided that the cleanup includes removing invalid and / or incomplete data, normalizing the data and / or filtering the data using filters.

[0034] In a further refinement, it is provided that the filters have frequency filters.

[0035] This makes it possible, among other things, to eliminate the influence of the vehicle's own movements.

[0036] In a further refinement, the classification model comprises a regression algorithm, a random forest algorithm, an artificial neural network algorithm or a gradient boosting machine algorithm.

[0037] The goal of a regression algorithm is to find a function that describes the relationship between the input variables (features) and the target variable. Regression algorithms can be used, among other things, to predict continuous target variables, but also for classification tasks, such as logistic regression. Although the name "regression" implies that it is a conventional regression method, logistic regression is often used to predict binary target variables. Logistic regression models the probability that an observation belongs to a particular class using the logistic function.

[0038] A random forest algorithm is an ensemble learning method used for classification and regression. It is based on the aggregation of multiple decision trees created during the training process. The basic idea is to generate a large number of decision trees and combine their predictions to increase the accuracy and robustness of the model.

[0039] At the beginning of the process, a large number of decision trees are created. Each tree is trained on a randomly selected portion of the training data. This method is called bootstrap aggregation or bagging. Random sampling with replacement reduces variance and decreases the likelihood of overfitting.

[0040] Another important aspect of the Random Forest algorithm is the random selection of features when constructing decision trees. At each node of a tree, a random subset of the available features is selected, and the best split is determined from this subset. This technique increases the diversity of the trees and contributes to improving the overall performance of the model.

[0041] In classification tasks, a Random Forest model typically uses majority voting on the predictions of individual trees. This aggregation of results leads to more robust and accurate predictions compared to individual decision trees.

[0042] A random forest algorithm offers several advantages. It is relatively easy to implement and understand, requires little data preprocessing, and is less prone to overfitting. It can handle both categorical and continuous data and is capable of handling missing values.

[0043] An artificial neural network (ANN) is a machine learning model consisting of a multitude of interconnected nodes called neurons. These neurons are organized into layers: an input layer, one or more hidden layers, and an output layer.

[0044] Each neuron in one layer is connected to neurons in the next layer. These connections have associated weights that are adjusted during the training process. The input layer captures the features of the data and passes them on to the hidden layers. In the hidden layers, the inputs are transformed by activation functions that can model nonlinear relationships between the inputs and outputs.

[0045] The training process of an ANN is carried out using an algorithm called backpropagation, as described previously. An ANN can be used for various tasks, including classification, regression, and pattern recognition.

[0046] A Gradient Boosting Machine (GBM) algorithm is a machine learning technique used for classification and regression. The algorithm is based on the idea of ​​sequentially training weak learning models, typically decision trees, with each new model aiming to correct the errors of the previous ones.

[0047] At the beginning of the process, a simple decision tree, also called a weak learner, is trained on the training data. The predictions of this first model are compared with the actual values, and the errors are calculated. These errors are called residuals. A new decision tree is then trained on these residuals to correct the errors of the first model. This process is repeated iteratively, with each new tree aiming to minimize the errors of the previous trees.

[0048] By sequentially combining weak learning models and optimizing them using gradient descent, the GBM can detect and model complex patterns and relationships in the data and is therefore well suited for classification models.

[0049] A further independent subject matter relates to a computer program product comprising a computer-readable storage medium on which instructions are embedded which, when executed by at least one computing unit, cause the at least one computing unit to be configured to carry out the method of the aforementioned type.

[0050] The method can be carried out in a distributed manner on one or more computing units, so that certain method steps are carried out on one computing unit and other method steps are carried out on at least one further computing unit, whereby calculated data can be transmitted between the computing units if necessary.

[0051] Another independent subject matter relates to a motor vehicle with a computer program product of the type described above.

[0052] Further advantages, features, and details will become apparent from the following description, which – where appropriate with reference to the drawings – describes at least one embodiment in detail. Identical, similar, and / or functionally equivalent parts are provided with the same reference numerals.

[0053] They show schematically: Fig. 1 a plan view of a motor vehicle on a roadway; Fig. 2 a side view of a motor vehicle with extensive sensor technology; Fig. 3 a training procedure for training a classification model for determining road surface properties, and Fig. 4 a flowchart of a method for classifying road surface conditions.

[0054] In the exemplary embodiments described below, identical or equivalent components or elements are provided with the same reference numerals for better readability.

[0055] Fig. 1 shows a roadway 2 on which a motor vehicle 4 travels at a speed v.

[0056] Roadway 2 has various road sections 2.1 to 2.5, some of which are damaged, such as a transverse wave 6 in section 2.3 and a pothole 8 in section 2.4.

[0057] The motor vehicle 4 is equipped with a classification model 10 based on an artificial intelligence algorithm, which is described in the following figures, in particular Fig. 3 and Fig. 4, is described in more detail. Using the classification model 10, the motor vehicle 4 is able to classify the roadway 2 using sensor data from various sensors with which the motor vehicle 4 is equipped.

[0058] To warn motor vehicle 2 of damage and to enable automatic notifications to road construction authorities and / or the police, motor vehicle 4 uses classification model 10 to determine the surface condition of roadway 2, classifies it, here into surface condition classes A - C, and potentially sends the results to the relevant authorities so that roadway 2 can be repaired if necessary. In the case of serious damage, especially recently occurring damage that could potentially pose a risk to motor vehicles and / or cause accidents, this can also be used to temporarily close the corresponding section of road.

[0059] In the present embodiment illustrated, three surface quality classes A, B and C are defined, where A describes a smooth road surface, B represents a wavy road surface and C represents a damaged road surface.

[0060] Fig. Figure 2 shows a side view of a motor vehicle 4' belonging to the same model series as the motor vehicle 4 of Fig. 1, but has a more extensive range of features in terms of the sensors installed in it.

[0061] The motor vehicle 4' has a front wheel 4.1 and a rear wheel 4.2, which are movably arranged on a chassis 4.4 of the motor vehicle 4' via a running gear 4.3.

[0062] When driving on lane 2, the chassis 4.3 compresses or extends, and the relative distance between the front wheel 4.1 or rear wheel 4.2 and the chassis 4.4 changes. With the help of appropriate sensors, Fig. 2, a spring travel sensor 4.5 is shown, on the chassis 4.3 these changes can be detected so that a level position N of the chassis 4.3 is determined, which in the following example, together with a radius R of the front wheel 4.1, defines a level NK of the motor vehicle 4'.

[0063] The section of the roadway 2 shown in front of the motor vehicle 4' is very uneven, which is why the chassis 4.3 compresses and rebounds intensively.

[0064] However, since chassis 4.3 cannot ideally eliminate all shocks, chassis 4.4 will vibrate to a certain extent, causing the level NK of vehicle 4' to change. This change can be eliminated from the measurement data of the sensors on chassis 4.3 using appropriate bandpass filters, leaving a mean level of vehicle 4' and preventing the measurement values ​​from chassis 4.3 from being distorted.

[0065] To warn motor vehicle 4' of damage and to enable automatic notifications to road construction authorities and / or the police, motor vehicle 4' determines the surface condition of roadway 2, classifies it, here into surface condition classes A - C, and, if necessary, sends the results to the relevant authorities so that roadway 2 can be repaired if necessary. In the case of serious damage, especially recently occurring damage that could potentially pose a risk to motor vehicle 4' or cause accidents, this can also initiate a closure of the corresponding road section.

[0066] The motor vehicle 4' is further equipped with wheel speed sensors 4.6 and 4.7 on the front wheel 4.1 and rear wheel 4.2, respectively. The wheel speed sensors 4.6 and 4.7 can each detect the current wheel speeds of the associated front wheels 4.1 and rear wheels 4.2. Each wheel of the motor vehicle 4' can be equipped with a corresponding wheel speed sensor.

[0067] Using the wheel speed sensors 4.6 and 4.7, it is possible to detect different wheel rotational speeds on the front and rear wheels 4.1, 4.2, which may be caused by the road surface conditions. For example, the wheel speed of a front wheel 4.1 may be higher at a given time due to a bump being driven over than that of a rear wheel 4.2 in the same lane if the rear wheel 4.2 is driving over a level surface at that moment. In particular, a wheel in the same lane, e.g. the rear wheel 4.2, which is essentially in the same lane as the front wheel 4.1, can travel a longer distance when driving over the same bump a short time later than the front wheel 4.1 at the same time, provided the front wheel 4.1 is driving over a level surface.

[0068] The same applies if wheels on the same axle have different wheel speeds than would be typical for the respective journey, for example, the same wheel speeds when driving straight ahead. This makes it possible, among other things, to distinguish a continuous bump from a pothole or similar in many situations.

[0069] Furthermore, a steering angle sensor 4.8 is provided on the steered front axle of the motor vehicle 4'.

[0070] When driving through potholes or over bumps, the changed force ratios on the wheels 4.1 often result in small shocks on the steering of the motor vehicle 4', so that the steering angle sensor 4.8 registers a small deflection, particularly when cornering.

[0071] The motor vehicle 4 from Fig. 1 differs from the motor vehicle 4' from Fig. 2 in that the motor vehicle 4 is Fig. 1 does not provide a spring travel sensor, which makes it impossible for the motor vehicle 4 to derive a surface condition of the roadway 2 from a pseudo-height profile determined using the spring travel sensor 4.5 provided in the motor vehicle 4'. Otherwise, the motor vehicle 4 essentially corresponds to the motor vehicle 4'.

[0072] Therefore, as described below, data from the wheel speed sensors 4.6, 4.7 and the steering angle sensor 4.8 are used to classify the surface of the roadway 2.

[0073] Fig. 3 shows a training procedure for training a classification model for determining road surface conditions.

[0074] The training is carried out using data from the motor vehicle 4' from Fig. 2 and comparable motor vehicles. The training is fundamentally based on the merging of data from a conventional method for classifying surface textures, e.g., a method based on the use of suspension travel sensors to create a pseudo-height profile, and other sensor data, such as data from wheel speed sensors 4.6 and 4.7 and the steering angle sensor 4.8.

[0075] The goal of the training is to train the selected artificial intelligence classification model 10 to determine surface quality classes A - C of a roadway 2 based on the other sensors, i.e., the wheel speed sensors 4.6, 4.7, and / or the steering angle sensor 4.8. This serves to enable the determination of road surface quality even with a motor vehicle 4 that does not have a suspension travel sensor.

[0076] While it is fundamentally possible to use classic algorithms to determine the influence of a bump on wheel speed or steering angle, for example, these methods often reach their limits in practice and in many cases do not produce usable results, as the respective conditions of the vehicles can differ greatly from one another, for example due to different levels of tire wear, slippage, load, and the like.

[0077] Artificial intelligence algorithms have proven to be more robust than conventional algorithms in the case of such uncertainties, so the use of an appropriate artificial intelligence-based classification model 10 usually achieves better results.

[0078] For a particular road section, ie more or less simultaneously, on the one hand classification data for the particular road section and on the other hand data DL relating to the steering angle sensor 4.8 and data DR relating to the wheel speed sensors 4.6, 4.7 are recorded.

[0079] The respective data DL, DR, and the corresponding assigned surface texture classes A - C are then first prepared during the training process, i.e., cleaned, appropriately formatted, and relevant features extracted. The data is then divided accordingly, with some being used as training data, some as validation data, and the remainder as test data sets. The training data can make up approximately 70% of the data volume, with the remaining 30% being divided simultaneously between validation and testing.

[0080] Depending on the algorithm chosen, for example a regression algorithm, a corresponding training procedure is applied and over time the classification model is improved until it reaches the desired criteria in terms of precision, accuracy and the like.

[0081] For motor vehicles 4' equipped with suspension travel sensors 4.5, corresponding data can be continuously recorded. Training of the corresponding classification model 10 can also be continued during operation, and the classification model can be continuously refined. More recent versions of the classification model 10 can be distributed to the motor vehicles 4, 4' via updates.

[0082] Fig. Figure 4 shows a flowchart of the procedure for classifying the surface condition of a road using the classification model.

[0083] When driving over a section of road, the steering angle, wheel speed, and position of the motor vehicle are recorded. In particular, minor deflections in the steering angle, which are obviously not attributable to steering inputs from the driver of motor vehicle 4, as well as sudden differences in the wheel speeds of the various wheels of motor vehicle 4, become visible.

[0084] In a next step, the data DL, DR are transmitted to at least one computing unit, wherein this computing unit can be provided in the motor vehicle 4 or in a remote data center.

[0085] In the computing unit, data filtering is first carried out to filter out invalid or unreliable data, e.g. if the sensors report incorrect data.

[0086] The trained classification model is then applied and a surface condition class is output based on the corresponding input data for the road section.

[0087] Although the invention has been illustrated and explained in detail by means of preferred embodiments, the invention is not limited by the disclosed examples, and other variations may be derived therefrom by those skilled in the art without departing from the scope of the invention. It is therefore clear that a multitude of possible variations exist. It is also clear that the embodiments mentioned by way of example are truly only examples and should not be construed as limiting the scope, possible applications, or configuration of the invention in any way.Rather, the preceding description and the description of the figures enable the person skilled in the art to implement the exemplary embodiments in concrete terms, whereby the person skilled in the art, with knowledge of the disclosed inventive concept, can make various changes, for example with regard to the function or the arrangement of individual elements mentioned in an exemplary embodiment, without departing from the scope of protection defined by the claims and their legal equivalents, such as a further explanation in the description. List of reference symbols 2 lanes 2.1 - 2.5 Road section 4, 4' motor vehicle 4.1 Front wheel 4.2 Rear wheel 4.3 Chassis 4.4 Chassis 4.5 Suspension travel sensor 4.6,4.7 Wheel speed sensor 4.8 Steering angle sensor 6 Cross shaft 8 potholes 10 Classification model A - C surface finish class DL data steering angle sensor DR data wheel speed sensor N Level position NK level motor vehicle R radius v speed x Direction of travel QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] DE 10 2011 120 022 A1

[0004]

Claims

[1] Method for classifying a surface condition of a roadway (2) by means of a motor vehicle (4, 4'), wherein the roadway (2) is driven on by the motor vehicle (4), wherein the motor vehicle (4, 4') detects an instantaneous speed (v), characterized by that the motor vehicle (4) records sensor data (DL, DR) from at least one steering angle sensor (4.8) and / or at least two wheel speed sensors (4.6, 4.7) arranged on different wheels (4.1, 4.2) of the motor vehicle (4, 4'), feeds the sensor data (DL, DR) to a classification model (10) trained with artificial intelligence and applies the classification model (10), wherein the classification model (10) assigns a surface quality class (AC) to the roadway (2). [2] Method according to claim 1, characterized by that the sensor data (DL, DR) are preprocessed after acquisition by removing incomplete and / or erroneous data. [3] Method according to claim 1 or 2, characterized by that the classification model (10) was trained on the basis of data (DL, DR) from at least one motor vehicle (4') which is equipped with spring travel sensors (4.5), at least one steering angle sensor (5.8) and / or at least two wheel speed sensors (4.6, 4.7) arranged on different wheels (4.1, 4.2). [4] Method for training a classification model (10) for a method according to one of the preceding claims, characterized byin that a roadway (2) is driven on with the aid of at least one motor vehicle (4'), which is equipped on the one hand with spring travel sensors (4.5) and on the other hand with at least one steering angle sensor (4.8) and / or at least two wheel speed sensors (4.6, 4.7) arranged on different wheels (4.1, 4.2), and the sensor data (DL, DR) of the sensors (4.5, 4.6, 4.7, 4.8) are recorded, wherein a surface quality class (AC) is determined on the basis of the spring travel sensor data, wherein the classification model (10) is trained on the basis of the data from steering angle sensors (4.8) and / or the data from the wheel speed sensors (4.6, 4.7) on the one hand and the surface quality class (AC) on the other. [5] Method according to claim 4, characterized by that the data from steering angle sensors (4.8) and / or the data from wheel speed sensors (4.6, 4.7) are cleaned. [6] Method according to claim 5, characterized bythat cleaning involves removing invalid and / or incomplete data (DL, DR), normalizing the data (DL, DR), and / or filtering the data (DL, DR) using filters. [7] Method according to claim 6, characterized by that the filters have frequency filters. [8] Method according to one of claims 4 to 7, characterized by that the classification model (10) comprises a regression algorithm, a random forest algorithm, an artificial neural network algorithm or a gradient boosting machine algorithm. [9] Computer program product, comprising a computer-readable storage medium on which instructions are embedded which, when executed by at least one computing unit, cause the at least one computing unit to be adapted to carry out the method according to one of the preceding claims. [10] Motor vehicle with a computer program product according to claim 9.

Images