Computer-implemented method for controlling actuators in a motor vehicle

Abstract

The invention relates to a computer-implemented method for controlling actuators (9) in a motor vehicle (10), comprising the following steps: - Reception of sensor data (7); - Control of the actuators (9) using the sensor data (7), characterized in that energy functions (8) of an artificial neural network are used in the control.

Description

[0001] The present invention relates to a computer-implemented method for controlling actuators in a motor vehicle according to the preamble of claim 1.

[0002] Methods are known from the prior art in which an artificial neural network is used to control actuators of a motor vehicle depending on sensor data.

[0003] EP 4 300 443 A1 discloses a system that uses sensors in the vehicle to record torque, engine and wheel speed, battery charge level, air conditioning settings, interior and exterior temperature, seat position, or the state of the entertainment system. The recorded data is then classified using an unsupervised clustering method and converted into a sequence of cluster IDs. These cluster IDs are tokenized, and a vocabulary is created. Each vocabulary corresponds to a coded sequence. Finally, an artificial neural network is trained for feature extraction through self-supervised learning of the coded sequences.

[0004] In contrast, the present invention is based on the object of enabling a control system that is better adapted to the real world.

[0005] This object is achieved by a method according to claim 1. Embodiments of the invention are specified in the dependent claims.

[0006] Sensor data is received. The sensor data can be measured or modeled. The actuators are controlled using the sensor data. According to the invention, energy functions of an artificial neural network are used for the control. In the context of this description, an energy function is understood in particular to be a function of the state variable x of an artificial neural network that decreases or remains constant during the temporal development of the artificial neural network. The use of energy functions is particularly advantageous in order to be able to better consider or process uncertainties existing in the real world. The energy functions can, for example, be added to one another, subtracted from one another, or multiplied with one another.

[0007] The energy functions depend on the sensor data and a latent variable. The latent variable is part of a graphic. The graphic can, in particular, be a pixel graphic. For example, the latent variable can be defined by the colors of the pixels.

[0008] According to one embodiment of the invention, various control functions of the motor vehicle can be implemented by controlling the actuators. Each of the control functions can be assigned one of the energy functions. In particular, it is possible for each of the control functions to be assigned exactly one of the energy functions. The control functions can, for example, be functions of the motor vehicle that are implemented through the interaction of the actuators.

[0009] According to one embodiment of the invention, the graphic can represent the latent variable as a function of the sensor data. The graphic can, for example, be created by a user. The user can, in particular, be a specialist, such as an engineer. In this way, the control system for subsequent operation of the motor vehicle can be configured when creating the graphic.

[0010] According to one embodiment of the invention, the latent variable can be a permissible deviation between a target value and a true value. The true value can be, for example, a measured value. The true value can, for example, be part of the sensor data. By defining the permissible deviation, a target for controlling the actuators is established. The actuators are thus controlled to keep the deviation between the target value and the true value less than or equal to the permissible deviation.

[0011] The permissible deviation can be specified by a color of at least one area of ​​the graphic. The area can consist of one or more pixels, for example. The true value can be influenced by controlling at least one of the actuators. The true value can thus be influenced by the energy functions or by an energy function.

[0012] According to one embodiment of the invention, the energy functions can be trained unsupervised before controlling the actuators. The training can be carried out, in particular, with the goal of implementing the control functions by the energy functions.

[0013] According to one embodiment of the invention, the energy functions can have been trained using training data sets. The training data sets each comprise sensor data, deviations between target values ​​and true values, a graph with the latent variable, which can, for example, define a permissible deviation, and a control signal for at least one of the actuators as a component. The training can be carried out, for example, using a "Joint Embedding Predictive Architecture" (JEPA), an "Image Joint Embedding Predictive Architecture" (I-JEPA), or a "latent-variable energy-based model" (LVEBM).

[0014] According to one embodiment of the invention, the control functions can include performance-optimized control of one of the actuators, energy-saving control of one of the actuators, and / or idle control of the actuator or actuators. With performance-optimized control, for example, the highest possible power of a motor vehicle's drive can be achieved. With energy-saving control, the motor vehicle's drive can, for example, be operated in a particularly energy-efficient manner. Idle control can, for example, involve controlling the actuator while the drive is not being used to drive the motor vehicle.

[0015] According to one embodiment of the invention, the actuators may comprise a relay of a traction battery, an electric drive, a throttle valve of an internal combustion engine, a valve, and / or a spark plug. The drive may, for example, be designed to drive the entire motor vehicle.

[0016] According to one embodiment of the invention, the sensor data can include a speed of the electric drive or the combustion engine, a torque request, a pedal position, pre-ignition, a boost pressure and / or energy consumption. The torque request is understood to mean, in particular, the request for a specific torque from the electric drive or the combustion engine. The request can be triggered automatically or by a user. The pedal position can, for example, be the position of an accelerator pedal. In the case of a combustion engine, pre-ignition can, for example, be the ignition of the gas mixture by an ignition spark before the piston reaches a position, preferably top dead center. The boost pressure can, for example, be the air pressure in an intake line of the combustion engine. In the case of an electric drive, energy consumption can be measured in kWh, for example.With combustion engines, energy consumption can be proportional to fuel consumption.

[0017] Further features and advantages of the present invention will become clear from the following description of preferred embodiments with reference to the accompanying drawings. The same reference numerals are used for identical or similar components, features, or elements and for components, features, or elements with identical or similar functions. Fig. 1A-F schematic views of graphs defining a latent variable; Fig. 2 is a schematic view of a method according to an embodiment of the invention; Fig. 3 a schematic view of the training of the energy functions; and Fig. 4 a schematic representation of the training loss when reconstructing the graph from the training data.

[0018] The Fig. Figure 1A shows several different areas that are shown in the Fig. 1B to 1F. Graph 1 can be created, for example, by a specialist such as an engineer. Graph 1 represents a latent variable as a function of parameters. The parameters can be, for example, a drive speed or a drive torque. The latent variable is a permissible deviation between a target value and a true value. The permissible deviation is determined by the color value of a pixel or a pixel range.

[0019] The true value can, for example, be a measured value. The true value can, for example, be part of the sensor data. By defining the permissible deviation, a target for controlling the actuators is established. The actuators are thus controlled to keep the deviation between the target value and the true value less than or equal to the permissible deviation. The true value can be influenced by controlling at least one of the actuators. When creating Figure 1, the expert can therefore specify the permissible deviation depending on one or more parameters. The parameter(s) can be part of the sensor data.

[0020] In Fig. 1B shows an example of a first section 2 of Figure 1, which may have been defined by a person skilled in the art for the purpose of a particularly energy-saving driving style. Fig. 1C shows a second section 3 of Figure 1 as an example, which may have been defined by a person skilled in the art for the purpose of controlling the drive in idle mode. Fig. 1D shows a third section 4 of Figure 1 as an example, which may have been defined by a person skilled in the art for the purpose of exhaust gas control. Fig. 1E shows, as an example, a fourth section 5 of the graph 1, which may have been defined by the person skilled in the art for the purpose of a particularly small permissible deviation (e.g. less than 2%), while the fifth section 6 of the graph 1 in Fig. 1F, for example, can define a slightly lower permissible deviation (e.g. between 2 and 4%).

[0021] During vehicle operation, the permissible deviation defined in Figure 1 can be used as a latent variable by an energy function. The energy function then controls the vehicle's actuators so that the actually measured deviation between the true value and the target value is less than or equal to the permissible deviation.

[0022] In the procedure according to Fig. 2, sensor data 7 are fed to a control unit of the motor vehicle 10. The sensor data 7 can be actually measured by sensors of the motor vehicle 10 or simulated. The sensor data 7 can include a speed of the electric drive or the combustion engine, a torque request, a pedal position, an ignition pre-ignition, a boost pressure, and / or an energy consumption. The sensor data 7 are fed to several previously trained energy functions 8. Each of the energy functions 8 is assigned to a control function of the motor vehicle 10. The energy functions can be added to one another, subtracted from one another, or multiplied with one another. The energy functions 8 can also be represented as E1(x, z1) ... E n (x, z n). The index n corresponds to the number of energy functions 8 and x to at least a part of the sensor data. The energy functions 8 depend on the sensor data 7 and the latent variable z. For example, a first energy function E1(x, z1) can depend on the latent variable z1, which is designed for energy-efficient operation of the motor vehicle, while an n-th energy function E n (x, z n ) from the latent variable z n which is designed for operation of the drive at idle.

[0023] In Fig. 2 shows that the two energy functions E1(x, z1) and E n (x, z n ) are added together and, as the added energy function E(x), control the actuators 9. The actuators 9 influence the operation of the motor vehicle 10.

[0024] In Fig. Figure 3 schematically illustrates the training of the energy functions 8. The training is performed using a training architecture 11 such as a "Joint Embedding Predictive Architecture" (JEPA), an "Image Joint Embedding Predictive Architecture" (I-JEPA), or a "latent-variable energy-based model" (LVEBM). The sensor data 7 and an original graphic 12 are fed to the training architecture 11 as training data. The original graphic 12 contains little or no information. It is modified during training using the sensor data 7 and a measured deviation between a target value and a true value.

[0025] In addition to the sensor data 7 and the original graph 12, the training data also includes graph 13, which contains the latent variable z1. During training, the training architecture outputs a control signal for the actuators 9 of the motor vehicle 10.

[0026] The training loss during training is in Fig. 4 as the difference between the original graphic 12 and a conditioned graphic 14, wherein the conditioned graphic 14 represents the control concept to be learned by the training architecture 11.

[0027] Training is performed unsupervised. The control concept is learned independently. For this purpose, the energy function is conditioned on an image containing the latent variable and illustrating the control concept to be learned. By minimizing the value of the energy function, the loss between the original graph 12 and the conditioned graph 14 is calculated. The energy functions are trained using auto-encoding, so that energy minimization on the input energy function reconstructs the conditioned graph 14. Additional input data for the energy function is the sensor data 7.

Claims

[1] Computer-implemented method for controlling actuators (9) in a motor vehicle (10), comprising the following steps: - Reception of sensor data (7); - controlling the actuators (9) using the sensor data (7), wherein energy functions (8) of an artificial neural network are used in the control, characterized by that the energy functions (8) depend on the sensor data (7) and a latent variable (z), wherein the latent variable (z) is part of a graph (1). [2] Method according to claim 1, characterized by that various control functions of the motor vehicle (10) are implemented by controlling the actuators (9), each of the control functions being assigned one of the energy functions (8). [3] Method according to one of the preceding claims, characterized by that the graph (1) represents the latent variable (z) as a function of the sensor data (7). [4] Method according to one of the preceding claims, characterized by that the latent variable (z) is a permissible deviation between a target value and a true value, wherein the permissible deviation is predetermined by a color of at least one area of ​​the graphic (1), wherein the true value can be influenced by the control of at least one of the actuators (9). [5] Method according to one of the preceding claims, characterized by that the energy functions (8) have been trained unsupervised before controlling the actuators (9). [6] Method according to the preceding claim, characterized by in that the energy functions (8) have been trained using training data sets, wherein the training data sets each comprise sensor data (7), deviations between target values ​​and true values, a graphic (12; 13) with the latent variable (z) and a control signal for at least one of the actuators (9) as a component. [7] Method according to one of the preceding five claims, characterized by that the control functions comprise a performance-optimized control of one or more of the actuators (9), an energy-saving control of the actuator (9) or the actuators (9) and / or an idle control of the actuator (9) or the actuators (9). [8] Method according to one of the preceding claims, characterized by that the actuators (9) comprise a relay of a traction battery, an electric drive, a throttle valve of an internal combustion engine, a valve and / or a spark plug. [9] Method according to the preceding claim, characterized by that the sensor data (7) include a speed of the electric drive or the combustion drive, a torque request, a pedal position, an ignition pre-ignition, a boost pressure and / or an energy consumption.

Images