DE102019214564A1 - System for training at least one neural control network for an adaptive speed control system of a vehicle - Google Patents

Abstract

A system and a method for training at least one neural control network for an adaptive speed control system of a vehicle are described, the neural control network serving to determine a control parameter for the adaptive speed control system for a vehicle that influences the driving behavior of the vehicle, the system using at least one neural training network comprises, wherein the system is set up to train the at least one neural training network, the at least one neural training network being supplied as input information at least for controlling the speed of a vehicle, the at least one neural training network being set up to provide a training model of the vehicle behavior To determine the vehicle from the vehicle information that can be used at least for cruise control, the system comprising at least one reference model, the at least ei ne reference model is set up to determine at least one reference parameter as output, the system being set up to train the at least one neural control network at least on the basis of the output of the reference model and the output of the neural training network in order to determine the driving behavior of the vehicle at least one output influencing control parameters.

Description

background

A system and a method for training at least one neural control network for an adaptive cruise control system (ACC) of a vehicle are described. In particular, the neural control network is used to determine a control parameter for the adaptive cruise control system that influences the vehicle behavior of the vehicle.

Conventionally, a precontrol and a controller have to be used to determine a control parameter that influences the vehicle behavior of the vehicle for an adaptive cruise control system (ACC). The pre-control is based on a vehicle model that calculates a control parameter from the currently estimated driving resistances. The controller, for example a PID controller, is used to compensate for the remaining control error that results from the model error and external interference.

Underlying task

The vehicle parameters for the precontrol and the control parameters can only be determined with great effort. In addition, it cannot be ensured that the desired vehicle behavior is regulated on the basis of the pilot control and the controller. There is therefore a need for a system with which the parameterization effort for determining a control parameter for an adaptive speed control system can be reduced.

Suggested solution

A system for training at least one neural control network for an adaptive speed control system of a vehicle is proposed. The neural control network is used to determine a control parameter influencing the driving behavior of the vehicle for the adaptive speed control system for a vehicle. The system comprises at least one neural training network. The system is set up to train the at least one neural training network, the at least one neural training network being supplied as input information at least for regulating the speed of a vehicle. The at least one neural training network is set up to determine a training model of the vehicle behavior of the vehicle from the vehicle information that can be used at least for speed control.

The system includes at least one reference model. The at least one reference model is set up to determine at least one reference parameter as an output.

The system is further configured to train the at least one neural control network at least on the basis of the output of the reference model and the output of the neural training network in order to output the at least one control parameter that influences the driving behavior of the vehicle.

With the system for training a neural control network, first a training model or a controlled system can be identified on the basis of the input information about the vehicle behavior with the at least one neural training network. In other words, the neural training network is used to generate a training model from an existing control vehicle model. The training model determined by the neural training network can then be used together with the reference model for training the neural control network. This can significantly reduce the parameterization effort. With the system it is possible to train the neural control network fully automatically to determine a control parameter that influences the driving behavior of the vehicle.

The system can be set up to determine a model error in the training model. The determined model error can be fed to the neural training network as an input into the training network. The system can determine the model error between the output of the existing vehicle model and the output of the neural training network.

The system can be set up to reduce the model error of the training model. In this context, the system can be set up to feed the determined model error to the neural training network via a predetermined learning method. Such a learning method can be, for example, the back propagation method (error feedback method). The backpropagation method can serve to minimize an error. An artificial neural control network can be learned in with the backpropagation method by changing the weights, since the output of the neural control network is directly dependent on them.

The system can be set up to complete the training of the at least one training model when predetermined states of the model error arise. The system can train the at least one training model complete when the model error falls below at least a predetermined limit value. Alternatively, the system can terminate the training of the at least one training model if the model error no longer decreases.

The neural training network can output at least one control parameter as output, which can be used by a speed control system to control the vehicle behavior. The at least one control parameter can be a vehicle acceleration, for example. The vehicle acceleration can be positive or negative. In particular, the control parameter can be a vehicle acceleration estimated by the training network.

The system can also be configured to use one or more of the following information as input information: road gradient, vehicle speed, vehicle acceleration, actual engine torque, target engine torque, actual wheel torque, target wheel torque, vehicle mass, gear ratio. These can be measured or estimated values. For example, the vehicle acceleration can be a measured value. The control of a vehicle can provide further information that can express the vehicle behavior. This information can in particular be available on a CAN bus of the vehicle. Depending on the availability of the information on a CAN bus of the vehicle, the system can be set up to use further information as input information for the neural training network.

After the training of the at least one neural training network has been completed, the system can be configured to train the neural control network to determine a control parameter that influences the driving behavior of the vehicle. The neural control network can be trained at least on the basis of the output of the reference model and the output of the neural training network.

The system can be set up to supply the neural control network for training the neural control network with input information provided by the neural training network. The input information for training the neural control network can be generated by the neural training network based on the training model and input into the neural control network. The system can accordingly be set up to use input information for training the neural control network that is generated by the neural training network. For example, based on the training model, the neural training network can generate one or more of the following input information for input into the neural controller network: road gradient, vehicle speed, vehicle acceleration, actual engine torque, target engine torque, actual wheel torque, target wheel torque, vehicle mass, gear ratio.

The system can be set up to determine a control error between the output of the neural training network and the at least one reference model for training the neural control network. This control error determined by the system can be fed to the neural control network. The neural control network can be trained on the basis of the control error determined and the input information described. The system can be set up to minimize the control error. The training of the at least one neural control network can be continued until the control error falls below a predetermined limit value or does not decrease any further. If one of these two states occurs, the training of the neural control network can be ended.

The reference model can have a transfer function. The transfer function can map a reference system behavior. The reference system behavior can express the desired system behavior. The reference model can output a reference rule parameter. The reference control parameter can be a reference acceleration, for example.

• Trainieren eines ein neuronalen Trainingsnetzes, wobei dem wenigstens einen neuronalen Trainingsnetz als Eingabeinformationen zumindest zur Geschwindigkeitsregelung eines Fahrzeugs verwendbare Fahrzeuginformationen zugeführt werden, wobei das wenigstens eine neuronale Trainingsnetz ein Trainingsmodell des Fahrzeugverhaltens des Fahrzeugs aus den Eingabeinformationen ermittelt,
• Ermitteln wenigstens eines Referenzparameters als Ausgabe wenigstens eines Referenzmodells, und
• Trainieren des wenigstens einen neuronalen Regelnetzes zumindest auf Basis der Ausgabe des Referenzmodells und der Ausgabe des neuronalen Trainingsnetzes, um den wenigstens einen das Fahrverhalten des Fahrzeugs beeinflussenden Regelparameter auszugeben.

A method is proposed for training at least one neural control network for an adaptive cruise control system for a vehicle, the neural control network serving to determine a control parameter for the adaptive cruise control system for a vehicle that influences the driving behavior of the vehicle. The procedure consists of the following steps:

• Training a neural training network, the at least one neural training network being supplied with vehicle information that can be used at least for speed control of a vehicle as input information, the at least one neural training network determining a training model of the vehicle behavior of the vehicle from the input information,
• Determining at least one reference parameter as the output of at least one reference model, and
• Training the at least one neural control network at least on the basis of the output of the reference model and the output of the neural training network in order to output the at least one control parameter that influences the driving behavior of the vehicle.

A model error in the training model can be determined and fed to the neural training network. The model error of the training model can be reduced, and the model error determined can be fed to the neural training network via a predetermined learning method.

The training of the at least one training model can be completed when the model error falls below at least a predetermined limit value or the model error no longer decreases.

After the training of the at least one neural training network has been completed, the neural control network can be trained to determine a parameter that influences the driving behavior of the vehicle. To train the neural control network, the neural training network can provide input information with which the neural control network is trained.

To train the neural control network, a control error between the output of the neural training network and the at least one reference model can be determined. This control error can be fed to the neural control network.

Furthermore, a neural control network for an adaptive cruise control system of a vehicle, a controller for an adaptive cruise control system of a vehicle, which has at least one such neural controller, and a vehicle with such a controller are proposed.

The aspects, variants and implementation options described above can be combined with one another regardless of the combinations described and regardless of the order described. The present disclosure therefore also encompasses aspects, variants and implementation options that are not explicitly described.

Figure list

• 1 zeigt schematisch ein System zum Trainieren wenigstens eines neuronalen Reglernetzes für ein adaptives Geschwindigkeitsregelsystem für ein Fahrzeug.
• 2 zeigt schematisch eine Steuerung für ein adaptives Geschwindigkeitsregelsystem.

Further goals, features, advantages and possible applications emerge from the following description of non-restrictive exemplary embodiments with reference to the associated drawings. All of the features described and / or shown in the figures, individually or in any combination, show the subject matter disclosed here, regardless of their grouping in the claims or their references. The dimensions and proportions of the components shown in the figures are expressly not to scale.

• 1 shows schematically a system for training at least one neural control network for an adaptive speed control system for a vehicle.
• 2 Figure 3 shows schematically a controller for an adaptive cruise control system.

Detailed description of the drawings

Components and features that are comparable or the same and have the same effect are each provided with the same reference symbols in the figures. In some cases, for reasons of clarity, reference symbols for individual features and components have been omitted in the figures, these features and components having already been provided with reference symbols in other figures. The components and features that are not described again with reference to the other figures are similar in their design and function to the corresponding components and features according to the other figures.

1 Figure 3 shows a schematic view of the system 100 for training a neural control network 10 to determine a parameter influencing the driving behavior of the vehicle. In 1 is the training operation of the system with dotted lines 100 shown. The actual regular operation with the trained neural control network 10 is shown with solid lines.

The system 100 includes a neural training network 12th . First is the neural training network 12th trained to use a training model of an existing control vehicle model 14th to create. This is done using the neural training network 12th Input information x _veh , states supplied, which are available for a cruise control system of a vehicle. The input information x _veh , states can be one or more of the following information: road gradient, vehicle speed, vehicle acceleration, actual engine torque, target engine torque, actual wheel torque, target wheel torque, vehicle mass, gear ratio. It is measured or otherwise determined vehicle information that indicates “real” driving conditions of a vehicle. Accordingly, the information additionally present on the CAN bus of a vehicle can also be used as input information for training the neural training network 12th be used. With this input information x _veh , states, the neural training network 12th identify a training model or a controlled system that the existing control vehicle model 14th should map.

At the point 16 can be a model error between the output of the vehicle model 14th and the training model of the neural training network 12th be determined. The neural training network 12th can determine _{an estimated variable} for a vehicle acceleration a estmd from the input information x _{veh, states.} The model error can be the difference between the output a _estmd , x _{veh_sts} , estmd of the neural training network 12th and the output of the vehicle model 14th a _real , x _{veh_sts} be. The model error determined is returned to the neural training network 12th guided to minimize the model error. A predetermined learning method such as backpropagation (error feedback method) can be used for this purpose. The training of the neural training network 12th is continued until the model error falls below a predetermined limit value or no longer decreases, ie until the model error remains constant.

After the training of the neural training network has been completed, the neural control network becomes 12th with the training model of the neural training network 12th trained. To do this, the neural controller network 10 the vehicle _acceleration a estmd and input information x _{veh_sts} , estmd, such as the speed and the drive torque, from the neural training network 12th or supplied by the training model of the neural training network. The input information a _estmd , x _{veh_sts} , _estmd mentioned is obtained from the training model of the neural training network 12th generated.

The system 100 assigns a reference model 18th for training the neural controller network 12th on. The reference model 18th includes a transfer function that indicates the desired vehicle behavior. In other words, the reference model can 18th specify the vehicle behavior between the actual acceleration and the target acceleration. The reference model 18th the vehicle _{acceleration a set is} supplied, which can be requested by the cruise control system and / or indirectly by a driver. The reference model 18th outputs a _{veh, des} as output.

At the point 20th becomes a control error between the output of the neural training network 12th and the output of the reference model 18th determined. The control error corresponds to the difference from the neural training network 12th output vehicle _acceleration a estmd and the output a _{veh, of} the reference model 18th . The rule error is from the system 100 back to the neural control network 10 guided and entered into this. The output of the neural controller network 10 becomes the neural training network 12th and supplied by the training model of the neural training network 12th processed to generate a corresponding output. The training of the at least one neural control network 10 is continued until the control error falls below a predetermined limit value or does not decrease any further.

In the normal mode shown with solid lines, the output of the neural control network 10 the vehicle model 14th supplied, which generates a corresponding output. In normal operation, the neural control network 10 the desired vehicle _{acceleration a set} and the input information x _veh , states supplied. The trained neural controller network 10 can determine a target torque for the drive train on the basis of the input information x _veh , states described above and the desired vehicle _{acceleration a set.} The target torque for the drive train represents a control parameter influencing the driving behavior of the vehicle. The target torque for the drive train can be a target wheel torque or a target engine torque as a function of the input information.

2 shows a schematic view of a controller 200 for an adaptive cruise control system. The control 200 includes the neural control network 10 with the system described above 100 was trained. The control 200 can input information from a CAN bus 202 retrieve or receive a vehicle. In addition, the controller can 200 also of controls 204 further vehicle systems receive input information. The control 200 is with a powertrain 206 of the vehicle connected to the powertrain 206 to be able to control accordingly. The neural controller network 10 The input information supplied includes the desired vehicle acceleration a _set and several other parameters such as the road gradient, vehicle speed, actual engine torque, target engine torque, actual wheel torque, target wheel torque. On the basis of this input information, the trained controller network 10 a target torque for the drive train 206 be determined.

It goes without saying that the exemplary embodiments explained above are not exhaustive and do not restrict the subject matter disclosed here.

Claims (20)

System (100) for training at least one neural control network (10) for an adaptive speed control system of a vehicle, the neural control network (10) for determining the driving behavior of the vehicle influencing control parameters for the adaptive speed control system, the system (100) comprising at least one neural training network (12) and at least one reference model (18), the system (100) being set up to train the at least one neural training network (12) , the at least one neural training network (12) being supplied as input information (x _{veh, states} ) that can be used at least for speed control of a vehicle, the at least one neural training network (12) being set up to generate a training model of the vehicle behavior from the input information (x _veh, states), the at least one reference model (18) being set up _{to determine at least one reference parameter (a veh, des} ) as output, the system (100) being set up to control the at least one neural controller network (10 ) at least on the basis of the output (a _{veh, des} ) of the reference model (18) and d _{To train the} output (a estmd, x _{veh_sts} , estmd) of the neural training network (12) in order to determine the at least one control parameter that influences the driving behavior of the vehicle. System (100) according to Claim 1 , wherein the system (100) is set up to determine a model error of the training model and to feed it to the neural training network (12). System (100) according to Claim 2 , wherein the system (100) is set up to reduce the model error of the training model, wherein the system (100) is further set up to supply the neural training network (12) with the determined model error via a predetermined learning method. System (100) according to Claim 3 , the system (100) being set up to complete the training of the at least one neural training network (12) when the model error falls below at least a predetermined limit value or the model error no longer decreases. System (100) according to one of the Claims 1 to 4th , wherein the neural training network (12) outputs at least one control parameter as an output, the at least one control parameter in particular being a vehicle _acceleration (a estmd). System (100) according to one of the Claims 1 to 5 , wherein the system (100) is set up to use one or more of the following information as input information: vehicle speed, vehicle acceleration, actual engine torque, target engine torque, actual wheel torque, target wheel torque, vehicle mass, gear ratio. System (100) according to one of the Claims 1 to 6th , wherein the system (100) is set up to train the neural control network (10) to determine at least one parameter influencing the driving behavior of the vehicle after the training of the at least one neural training network (12) has been completed. System (100) according to Claim 7 _{, the system (100) being set up to supply input information (a estmd} , x _{veh_sts} , estmd) provided by the neural training network (12) to the neural controller network (10) for training the neural controller network (10). System (100) according to Claim 7 or 8th , the system (100) being set up to determine a control error between the output of the neural training network (10) and the at least one reference model (18) for training the neural control network (10) and to feed this control error to the neural control network (10) . System (100) according to one of the Claims 1 to 9 , wherein the reference model (18) has a transfer function that maps a reference system behavior. A method for training at least one neural control network (10) for an adaptive speed control system of a vehicle, the neural control network (10) serving to determine a control parameter for the adaptive speed control system that influences the driving behavior of the vehicle, the method comprising the following steps: training a neural Training network (12), the at least one neural training network (12) being supplied as input information (x _veh, states) information that can be used at least for speed control of a vehicle, the at least one neural training network (12) being a training model of the vehicle behavior from the input information (x _veh, states), determining at least one reference parameter as the output of at least one reference model (18), and training the at least one neural control network (10) at least on the basis of the output (a _{veh, des} ) of the reference model (18) and the Output (a _estmd , x _{veh_sts} , _estmd ) of the neural training network (12) in order to output the at least one control parameter that influences the driving behavior of the vehicle. Procedure according to Claim 11 , wherein a model error of the training model (12) is determined and fed to the neural training network (12). Procedure according to Claim 12 , the model error of the training model being reduced, the determined model error being fed to the neural training network (12) via a predetermined learning method. Procedure according to Claim 13 , the training of the at least one neural training network (12) being terminated when the model error falls below at least a predetermined limit value or the model error no longer decreases. Method according to one of the Claims 11 to 14th , wherein after the training of the at least one neural training network (12) has been completed, the neural control network (10) is trained to determine a parameter that influences the driving behavior of the vehicle. Procedure according to Claim 15 , the neural training network (12) providing input information (a _estmd , x _{veh_sts} , _estmd ) with which the neural control network (10) is trained. Procedure according to Claim 15 or 16 , wherein for training the neural control network (10) a control error between the output of the neural training network (12) and the at least one reference model (18) is determined and this control error is fed to the neural control network (10). Neural control network (10) for an adaptive speed control system of a vehicle, the neural control network (10) having a system (100) according to one of the Claims 1 to 10 and / or with a method according to one of the Claims 11 to 17th was trained. Control (200) for an adaptive cruise control system of a vehicle, the control (200) according to at least one neural control network (10) Claim 18 or with at least one regulator network Claim 18 is connectable. Vehicle with a control (200) according to Claim 19 for an adaptive cruise control system.

Images