Computer-implemented method for determining a wheel braking torque for a braking device in a vehicle

Abstract

The present invention discloses a computer-implemented method (100) for determining a wheel braking torque for a braking device (92) in a vehicle (90), comprising the following steps: - generating (102) first training data (4) based on a simulated vehicle data model (2); - pre-training (104) a neural network (50) by feeding the first training data (4) as first input data (5) into the neural network (50) in order to generate first output data (8) which correspond to a wheel braking torque (94); - comparing (106) the first output data (8) with a predetermined target wheel braking torque (9) of the simulated vehicle data model (2), and if a deviation (11) of the first output data (8) from the target wheel braking torque (9) does not correspond to a predetermined quality parameter (10), repeatedly feeding (108) the first output data (8) into the neural network (50) until the subsequently newly generated second output data (11) correspond to the predetermined quality parameter (10).

Description

The present invention relates to a computer implemented method for determining a wheel braking torque for a braking device in a vehicle.Prior ArtIn the prior art, hydraulically actuated friction brakes are used in vehicles. In the case of hydraulic brakes, a pressure sensor is typically used in the hydraulic system in order to determine the clamping force of the brakes, as a result of which, inter alia, the braking torque can also be estimated and regulated. Uncertainties in the friction pairing, i.e. the variation of the coefficient of friction and thus of the brake characteristic value, represent a decisive limitation of the accuracy of the brake torque estimation.Changed boundary conditions, for example in electric vehicles, can also make the use of electromechanically actuated brakes (EMB) attractive. The question of a suitable sensor system at the actuator level is still open. In many known concepts, force measurement, e.g., the measurement of the clamping force in disc brake-based EMBs, is used in a similar manner as the pressure measurement in hydraulic brakes, whereby the accuracy of a brake torque estimation based thereon is, however, also limited. In addition, the outlay for such a sensor system is increased in comparison with hydraulic brakes, in particular since measurement has to be carried out in each wheel actuator.In particular in the case of drum brake EMBs, the direct measurement of the braking torque is an alternative which offers high accuracy, but likewise involves additional outlay for the sensor system.It is therefore the object of the present invention to provide a solution by means of which a wheel brake torque for a brake device in a vehicle can be determined in an efficient and reliable manner.Disclosure of the InventionThis object is achieved by a computer-implemented method for determining a wheel brake torque for a brake device in a vehicle having the features of the independent claim.According to a first aspect, the disclosure relates to a computer-implemented method for determining a wheel brake torque for a brake device in a vehicle, which method comprises the following steps:In a first step, first training data is generated, which are based on a simulated vehicle data model.In a second step, a pretraining of a neural network is carried out by feeding the first training data as first input data into the neural network in order to generate first output data which correspond to a wheel brake torque.In a third step, the first output data is compared with a predefined setpoint wheel brake torque of the simulated vehicle data model, and if a deviation of the first output data from the setpoint wheel brake torque does not correspond to a predefined quality parameter, in a fourth step, the first output data is repeatedly fed into the neural network until the second output data, which is subsequently newly generated, correspond to the predefined quality parameter.A basic idea of the present invention is that the wheel brake torque of the vehicle is estimated independently of sensors at the actuator level, e.g. force / torque sensors or model-based estimation on actuator motor signals (e.g. motor current or position). The estimation method is based on a sensor system at the entire vehicle level, which is already installed in the vehicle, and thus enables the replacement of physical sensors in the actuator or the plausibility check thereof in order to save redundancies. The method according to the invention is designed in particular to be robust with respect to interference variables and environmental influences which influence the relationship between sensor signals at the entire vehicle level and wheel brake torques.An essential aspect of the present invention for determining the wheel brake torque for a vehicle consists in the use of a (deep) neural network, which can be designed in particular as a long-short-term memory network (LSTM), as a recurrent neural network (RNN) or as a nonlinear autoregressive exogenous model (NARX), in order to determine or estimate wheel brake torques, for example on the basis of, inter alia, in-vehicle data, such as relevant sensor signals of the inertial sensor system, wheel speed sensor system, and of the steering angle and the pedal position (or alternatively setpoint value of a driver assistance function).The present invention thereby provides the following advantages:High accuracy of the brake torque estimationHigh robustness with respect to disturbances and environmental influencesindependence of physical sensors or signals on the actuator levelRelatively low modeling effortThis means that, as a result of the method according to the invention, apart from the sensor system already installed in present-day vehicles, no further sensors are necessary for determining the braking torque. In the present invention, therefore, the estimation of a braking torque for the vehicle is performed on the basis of total vehicle signals.In this context, it should be mentioned that the focus of the brake torque estimation in the present invention is preferably located in the comfort brake range. That is, below the wheel lock level. The estimation basically also works in the ABS range, but it is assumed that in this range the wheel is controlled in any case on the basis of the slip behavior, for which reason an accurate estimation of the braking torque is less relevant. In the comfort braking range, unequal braking torques on different wheels, which can lead, for example, to the vehicle being pulled obliquely due to unequal friction behavior of the brakes, but also due to unequal actuator behavior in the case of individually actuated wheels (for example, in the case of EMBs). As a consequence, it is not necessarily required that the algorithm can follow highly dynamic ABS modulations, which has an effect on the required speed of the estimation and sampling rate of the input signals.One embodiment of the method provides that the simulated pre-trained neural network is trained overnight by second input data being fed into the neural network, the second input data being designed as test data which were generated in a test mode using a real vehicle. This achieves the advantage that the accuracy of the determination of the braking torque is increased.One embodiment of the method provides that the simulated vehicle data model has an artificially generated data record which maps at least one simulated driving maneuver of the vehicle for at least one operating state of the vehicle in order to determine a corresponding wheel brake torque for the at least one vehicle maneuver on the basis of at least one simulation parameter to be set. As a result, a driving situation-specific braking torque determination is achieved in an efficient manner.One embodiment of the method provides that the artificially generated data set is fed with at least one adjustable influencing parameter which influences the estimation of the wheel brake torque of the vehicle. This allows a flexible and context-specific brake torque determination.One embodiment of the method provides that a corresponding parameter range for the at least one simulation parameter to be set is generated for the simulated vehicle data model on the basis of the at least one influence parameter. As a result, the determination of the braking torque by the neural network can be further improved.According to a second aspect, the disclosure relates to a computer-implemented method for determining a wheel braking torque for a vehicle using a trained neural network according to the first aspect of the present invention, having the following steps:In a first step, first measurement data is obtained, which consist of vehicle-relevant data and / or environment-related data.In a second step, the acquired first measurement data is processed in order to generate second measurement data.In a third step, the second measurement data is fed in as third input data for the neural network.In a fourth step, second output data of the neural network are generated on the basis of the first measurement data, the second output data representing an estimate of the wheel braking torque for the braking device of the vehicle.One embodiment of the method provides that the vehicle-relevant data of the vehicle have at least one of the following data: data from a driver assistance function, data from an inertial sensor system, wheel rotational speed data from at least one wheel of the vehicle, data from a drive torque of the vehicle, control device-related data which describe at least one driving parameter of the vehicle. As a result, the braking torque determination can be carried out efficiently and in a vehicle-specific manner.According to a third aspect, the disclosure relates to a computer program containing machine-readable instructions which, when executed on one or more computers and / or computer instances, cause the computer or computer instances to carry out the method according to the invention.According to a fourth aspect, the disclosure relates to a machine-readable data carrier and / or download product with the computer program.According to a fifth aspect, the disclosure relates to one or more computers and / or computer instances with the computer program, and / or with the machine-readable data carrier and / or the download product.Further measures which improve the invention are described in more detail below together with the description of the preferred exemplary embodiments of the invention on the basis of figures.Exemplary EmbodimentsIt shows: FIG. 1 is a schematic flow diagram of a computer-implemented method 100 for determining a wheel braking torque for a brake device 92 in a vehicle 90; FIG. 2 is a schematic flow diagram of a computer-implemented method 200 for determining a wheel braking torque for a vehicle 90 using a pre-trained neural network 50 according to an embodiment of the present invention; and FIG. 3 is a schematic illustration of the algorithm for determining a wheel brake torque 94, 95, 96 according to an embodiment of the present invention.FIG. 1 shows a schematic flow diagram of a computer-implemented method 100 for determining a wheel braking torque for a brake device 92 in a vehicle 90.In step 102, first training data 4 is generated, which are based on a simulated vehicle data model 2. These simulation models are usually generated during the development of the vehicle and can be used to pretraining the algorithm, i.e. the neural network 50.In step 104, a pretraining, a so-called pretraining, of a neural network 50 is carried out by feeding the first training data 4 as first input data 5 into the neural network 50 in order to generate first output data 8 which correspond to a wheel brake torque 94. The pretraining of the neural network 50 allows the advantage of the low outlay, since the wheel torque is known in simulations and does not need to be measured explicitly.The neural network 50 can be designed as a long-short-term memory (LSTM), recurrent neural network (RNN) or as a nonlinear autoregressive exogenous model (NARX).In a step 106, the first output data 8 is compared with a predefined setpoint wheel braking torque 9 of the simulated vehicle data model 2, and if a deviation 11 of the first output data 8 from the setpoint wheel braking torque 9 does not correspond to a predefined quality parameter 10, the first output data 8 is repeatedly fed into the neural network 50 in step 108 until the second output data 11 that are subsequently newly generated correspond to the predefined quality parameter 10.Optionally, the simulated trained neural network 50 is trained overnight by feeding second input data 6 into the neural network 50, wherein the second input data 6 are designed as test data which were generated in a test operation with a real vehicle.In other words: The simulated pre-trained network 50 is subsequently retrained in test operation with a real vehicle in order to compensate for inaccuracies or aspects not contained in the simulation model. For this purpose, a vehicle with a torque measuring rim can be used, for example, and the widest possible range of the driving states conceivable in real operation should be traveled through. Alternatively, it is also possible to rely on other torque estimates or measurements for training on the real vehicle in order to reduce the use of torque measurement rims. For example, the sensor system and signals at the actuator level can be used for this purpose if this delivers sufficiently precise braking torques.Optionally, the simulated vehicle data model or vehicle model has an artificially generated data record 20 which maps at least one simulated driving maneuver of the vehicle 90 for at least one operating state of the vehicle 90 in order to determine a corresponding wheel brake torque 95 for the at least one vehicle maneuver on the basis of at least one simulation parameter 22 to be set.In other words, it is proposed to generate an artificial set of driving maneuvers for the vehicle 90, which travel maneuvers cover the relevant operating states of the vehicle as completely as possible. Particular attention is paid here to so-called disturbances / variations / deviations or simply only influencing variables or influencing parameters 24 on the vehicle, which however can influence the brake torque estimate. Some important deviation parameters or disturbance parameters are listed as examples:- Tire size- Tire pressure- Tire condition- Weather temperature- vehicle massweight distributiontemperature of the brake (e.g. disc brake, drum brake etc.)wind conditionsCondition of the road surface (e.g. road inclination, condition of the road surface)Optionally, therefore, the artificially generated data set 20 is fed with at least one adjustable influencing parameter 24, which influences the estimation of the wheel brake torque 95 of the vehicle 90.Optionally, a corresponding parameter range 26 for the at least one simulation parameter 22 to be set for the simulated vehicle data model 2 can then be generated on the basis of the at least one influence parameter 24.The parameter range derived on the basis of the influence variables 24 can be very well covered in simulations. For this purpose, the parameter ranges 26 are sampled and the respective parameter variations are combined with one another. This results in a high number of maneuvers which are simulated automatically in order to train the braking torque estimator, i.e. the neural network 50.FIG. 2 shows a schematic flow diagram of a computer-implemented method 200 for determining a wheel braking torque for a vehicle 90 using a pre-trained neural network 50.In step 202, first measurement data is obtained, which consist of vehicle-relevant data 32 and / or environment-related data 34.Environmental data 34 may be, for example:Navigation data, temperature signals or windscreen wiper status, which allow conclusions to be drawn about possible coefficients of friction, driving states, inclination of the road, etc.In step 204, the acquired first measurement data 30 is processed in order to generate second measurement data 36.The optional processing of the first measurement data 30 can include at least one of the following measures: sampling the measurement data, checking the measurement data with respect to their correctness (plausibility checking), normalizing the measurement data, extracting a defined time interval or time window of the first measurement data 30, i.e. the time window comprises a start time of the braking process and an end time of the braking process from the first measurement data 30.In step 206, the second measurement data is fed in as third input data 38 for the neural network 50.In step 208, second output data of the neural network are generated on the basis of the first measurement data, the second output data 40 representing an estimate of the wheel brake torque 96 for the brake device 92 of the vehicle 90.Optionally, the vehicle-relevant data 32 of the vehicle 90 comprises at least one of the following data: data from a driver assistance function, data from an inertial sensor system, wheel rotational speed data from at least one wheel of the vehicle 90, data from a drive torque of the vehicle 90, control device-related data which describe at least one driving parameter of the vehicle 90.FIG. 3 is a schematic illustration of the algorithm for determining a wheel brake torque 94, 95, 96 according to an embodiment of the present invention.The neural network 50 is supplied with vehicle-relevant data 32 and / or environment-related data 34, wherein optionally a signal preprocessing 60 takes place, such as, for example, a normalization of the data to be supplied to the network 50, before the data are input into the network 50. The data output by the network 50 in the form of a specific or estimated wheel brake torque 94, 95, 96 for a vehicle 90 are processed accordingly for output before being output via a signal post-processing 70. This signal post-processing 70 may include, for example, cancelling normalization of the data. The rendering 60 and the rendering 70 may optionally also be learned and performed by the neural network 50.Based on FIG. 3, the method according to the invention can also be illustrated as follows:In the first step, relevant sensor signals 32, 34 are read out, which are acquired from sensors already present in the vehicle 90. These may be, for example:HMI (human machine interface) or alternatively desired signals from driver assistance functions:◯ Brake Pedal Position and / or Pressure◯ steering wheel angleinertial sensor system◯ Longitudinal Acceleration◯ Lateral Acceleration◯ yaw rateWheel speed sensors (WSS)o wheel speeds of all wheelsIn addition, signals from the motor train, for example, recuperation / drive torque, can be used.In addition, estimated values from existing control units, e.g. vehicle speed or road gradient from ESP, can be usedSignals about the environmental state can also be used, for example navigation data, temperature signals or windscreen wiper status, which allow conclusions to be drawn about possible coefficients of friction, driving states, road inclination, etc.In the second step, these signals are processed for further processing by the neural network 50 (=signal preprocessing 60). This may include, in particular:sampling the time signals: reading out all signals at defined timesplausibility checking of the sensor signals, i.e. finding out whether all signal values are in the expected range, or there are signals which indicate defective sensors.normalization of the sensor signals: To avoid numerical problems in training the algorithm, the sensor signals are normalized such that their magnitude lies between defined minimum and maximum values (typically between 0 and 1 or -1 and 1).windowing the signals: partitioning the time signals into "sensor values" on which the estimation is based and an estimated output value in the next time step.In the third step, a (deep) neural network 50 for estimating the braking torque is proposed, such as a long short-term memory (LSTM) network, a recurrent neural network (RNN) or a nonlinear autoregressive exogenous model (NARX) architecture.In the fourth step, the output of the neural network 50 is processed again (=signal processing 70), in particular the normalization of the output signals can be reversed again here, so that an estimated wheel torque in Nm is obtained as the output of the neural network 50, for example for a corresponding control device in the vehicle.

Claims

Computer-implemented method (100) for determining a wheel braking torque for a braking device (92) in a vehicle (90), comprising the following steps: - generating (102) first training data (4) which are based on a simulated vehicle data model (2); - pretraining (104) a neural network (50) by feeding the first training data (4) as first input data (5) into the neural network (50) in order to generate first output data (8) which correspond to a wheel braking torque (94); comparing (106) the first output data (8) with a predetermined setpoint wheel brake torque (9) of the simulated vehicle data model (2), and if a deviation (11) of the first output data (8) from the setpoint wheel brake torque (9) does not correspond to a predetermined quality parameter (10), repeatedly feeding (108) the first output data (8) into the neural network (50) until the second output data (11), which is subsequently newly generated, correspond to the predetermined quality parameter (10).The computer-implemented method (100) according to claim 1, wherein the simulated pre-trained neural network (50) is trained overnight by feeding second input data (6) into the neural network (50), wherein the second input data (6) are formed as test data which were generated in a test operation with a real vehicle.Computer-implemented method (100) according to Claim 1 or 2, wherein the simulated vehicle data model (2) has an artificially generated dataset (20) which maps at least one simulated driving maneuver of the vehicle (90) for at least one operating state of the vehicle (90) in order to determine a corresponding wheel brake torque (95) for the at least one vehicle maneuver on the basis of at least one simulation parameter (22) to be set.Computer-implemented method (100) according to claim 3, wherein the artificially generated data set (20) is fed with at least one adjustable influence parameter (24) that influences the estimation of the wheel brake torque (95) of the vehicle (90).The computer-implemented method (100) according to claim 4, wherein a corresponding parameter range (26) for the at least one simulation parameter (22) to be set is generated for the simulated vehicle data model (2) on the basis of the at least one influence parameter (24).Computer-implemented method (200) for determining a wheel brake torque for a vehicle (90) using a neural network (50) trained according to one of the preceding claims, having the following steps: - obtaining (202) first measurement data (30) which consist of vehicle-relevant data (32) and / or environment-related data (34); - preparing (204) the obtained first measurement data (30) in order to generate second measurement data (36); - feeding (206) the second measurement data (36) as third input data (38) for the neural network (50); and - generating (208) second output data (40) of the neural network (50) on the basis of the first measurement data (30), wherein the second output data (40) represents an estimate of the wheel brake torque (96) for the brake device (92) of the vehicle (90).Computer-implemented method (200) according to Claim 6, wherein the vehicle-relevant data (32) of the vehicle (90) has at least one of the following data: data from a driver assistance function, data from an inertial sensor system, wheel rotational speed data from at least one wheel of the vehicle (90), data from a drive torque of the vehicle (90), control device-related data which describe at least one driving parameter of the vehicle (90).A computer program comprising machine readable instructions which, when executed on one or more computers and / or compute instances, cause the computer or compute instances to perform the method of any one of claims 1 to 7.Machine-readable data carrier and / or download product comprising the computer program according to Claim 8.One or more computers and / or compute instances comprising the computer program of claim 8, and / or comprising the machine readable medium and / or download product of claim 9.

Images