DE102022108392A1 - Computer-implemented method for calibrating a control unit of an electric powertrain of a vehicle - Google Patents

Abstract

The present invention relates to a computer-implemented method for calibrating a control unit (1) of an electric drive train (2) of a vehicle, wherein a model of the drive train (2) required for a learning reinforcement agent (3) is described by a model formula which is represented by a symbolic regression model ( 4) is determined.

Description

The present invention relates to a computer-implemented method for calibrating a control unit of an electric drive train of a vehicle.

Definition: In the context of the present application, the term “vehicle” is intended to mean land vehicles, in particular motor vehicles, watercraft, in particular boats, and aircraft.

An electric drive train that includes an electric drive unit can be characterized by various parameters, such as the output voltage, which in an electric vehicle is usually an alternating voltage, and the internal torque that is generated by the electric drive unit. Furthermore, the electrical power loss of the electric drive unit is of interest, which arises in particular from the materials used, such as copper and iron in the rotor and stator, since these materials give off heat, especially through friction. With a high power loss, the electric drive unit therefore requires more energy to generate the desired torque. Therefore, knowledge of the power loss of the electric drive unit is of great importance for the operation of a vehicle.

Some approaches for calibrating an electric drive train of a vehicle are already known from the prior art, which use methods of artificial intelligence, in particular methods of reinforcement learning. A model of the electric powertrain is required to perform the calibration.

The calibration of electric powertrains using model-based reinforcement learning is already state of the art. The models used are so-called simulators. Such simulators attempt, for example, to simulate the dynamics of the environment around the electric drive train in detail. This can be done on a physical basis (for example through physical-mathematical equations) or on a data-driven basis (for example with neural networks). Another approach is to use a Markov decision process as a model.

Data-driven models, such as artificial neural networks or genetic algorithms, work well in practice, but are often not fully understood by users. Furthermore, it is difficult to prove that they work or can extrapolate beyond the limits of the training data set. For example, if there is a systematic problem in the machine learning model, it is very difficult to solve this problem using a physics approach.

Physical models of the electric powertrain are better understood by users, but are often difficult to implement. However, in practice, creating an environment model proves to be very time-consuming and resource-consuming.

The DE 10 2019 126 244 A1 discloses a method and a system for calibrating a control unit of an electric motor. The electric motor is modeled using a combination of reinforcement learning and regression analysis.

The present invention aims to provide a computer-implemented method for calibrating an electric powertrain of a vehicle, which can simplify the required creation of the environment model for a learning reinforcement agent.

The solution to this problem is provided by a computer-implemented method for calibrating a control unit of an electric drive train of a vehicle with the features of claim 1. The subclaims relate to advantageous developments of the invention.

• - Initialisieren eines symbolischen Regressionsmodells, welches dazu ausgebildet ist, Daten und Messwerte zumindest eines Betriebsparameters des elektrischen Antriebstrangs zu empfangen und zumindest eine mathematische Modellformel, welche die Kenngröße charakterisiert, zu generieren und an einen Lernverstärkungsagenten auszugeben (Schritt 100),
• - Bestimmen eines Zustands des elektrischen Antriebsstrangs, wobei dieser Zustand durch die Daten und Messwerte des zumindest einen Betriebsparameters des elektrischen Antriebsstrangs definiert wird und der Wert der zumindest einen Kenngröße zumindest teilweise von diesem Betriebsparameter abhängt und wobei der Lernverstärkungsagent diesen Zustand sowie die Modellformel des symbolischen Regressionsmodells als Eingangsgrößen erhält und dadurch die Umgebung beobachtet (Schritt 200),
• - Ausführen einer Lernverstärkungsfunktion, bei der der Lernverstärkungsagent mit der Steuereinheit interagiert und zumindest eine Aktion ausführt, welche die Steuereinheit dazu veranlasst, den Zustand des elektrischen Antriebsstrangs zu verändern, wobei auch der Zustand des Lernverstärkungsagenten verändert wird und der Lernverstärkungsagent eine positive oder negative Belohnung für seine Aktion erhält (Schritt 300),
• - erneutes Bestimmen des Zustands des elektrischen Antriebsstrangs, wobei die Daten und Messwerte des zumindest einen Betriebsparameters dem symbolischen Regressionsmodell zugeführt werden (Schritt 400),
• - Ausführen des symbolischen Regressionsmodells mit dem zumindest einen, im vorhergehenden Verfahrensschritt bestimmten Betriebsparameter, wobei die zumindest eine Modellformel für die Modellierung der zumindest einen Kenngröße modifiziert wird (Schritt 500),
• - wiederholtes Ausführen der Schritte 200 bis 500 - also der zweiten bis fünften Verfahrensschritte - mit der im vorhergehenden Verfahrensschritt modifizierten Modellformel, bis die Kalibrierungsaufgabe gelöst ist.

A computer-implemented method according to the invention for calibrating a control unit of an electric drive train of a vehicle, in which at least one parameter is determined, comprises the steps:

• - Initializing a symbolic regression model, which is designed to receive data and measured values of at least one operating parameter of the electric drive train and to generate at least one mathematical model formula that characterizes the parameter and output it to a learning reinforcement agent (step 100),
• - Determining a state of the electric drive train, this state being defined by the data and measured values of the at least one operating parameter of the electric drive train and the value of the at least one parameter depending at least partially on this operating parameter and the learning reinforcement agent determining this state and the model formula of the symbolic regression model as input variables receives and thereby observes the environment (step 200),
• - Executing a learning reinforcement function in which the learning reinforcement agent interacts with the control unit and performs at least one action which causes the control unit to change the state of the electric powertrain, wherein the state of the learning reinforcement agent is also changed and the learning reinforcement agent provides a positive or negative reward for receives its action (step 300),
• - re-determining the state of the electric drive train, with the data and measured values of the at least one operating parameter being fed to the symbolic regression model (step 400),
• - Executing the symbolic regression model with the at least one operating parameter determined in the previous method step, wherein the at least one model formula for modeling the at least one parameter is modified (step 500),
• - Repeatedly carrying out steps 200 to 500 - i.e. the second to fifth process steps - with the model formula modified in the previous process step until the calibration task is solved.

An essential basic idea of the invention is that the model of the electric drive train required for the learning reinforcement algorithm (reinforcement learning algorithm) carried out by the learning reinforcement agent is described by at least one mathematical model formula, which in particular characterizes the at least one parameter and by one that carries out a symbolic regression Algorithm is determined. In other words, a model-based approach of reinforcement learning for calibrating the control unit of the electric drive train is combined with a symbolic regression algorithm in order to automatically model the electric drive train using at least one model formula that is as accurate as possible. The approach proposed here combines two individual algorithms with each other. A first algorithm carries out a symbolic regression with the aim of determining the mathematical model formula that best describes the electric drive train, in particular the at least one parameter. A second algorithm is the model-based learning reinforcement algorithm, which uses the model formula (model prediction) derived by the first single algorithm as a model. The aim of this model-based learning reinforcement algorithm is to calibrate the electric drive train by minimizing or maximizing the deviation between a target value and a measured actual value of the parameter - depending on the type of parameter. This parameter can in particular be an electrical power loss that is intended to be minimized by the calibration.

• - symbolische und unabhängige Variablen, wie zum Beispiel U (Spannung), I (Strom), P (Leistung), T (Drehmoment), usw.;
• - physikalische Einheiten, wie zum Beispiel V; A, 1/min, Nm, etc.;
• - symbolische Operatoren, wie zum Beispiel +, -, *, /, trigonometrische Funktionen, usw.

The input variables for the first algorithm that performs the symbolic regression are data and values of operating parameters of the electric powertrain that are required for creating the model of the electric powertrain by the symbolic regression, these operating parameters being, for example, the speed, the torque, The temperature, the electrical current, the electrical voltage, the electrical power of an electric drive unit of the electric drive train or the state of charge of a traction battery can be. Other input variables can in particular be:

• - symbolic and independent variables, such as U (voltage), I (current), P (power), T (torque), etc.;
• - physical units, such as V; A, 1/min, Nm, etc.;
• - symbolic operators, such as +, -, *, /, trigonometric functions, etc.

• - die Daten und Messwerte der Betriebsparameter, die den Antriebsstrang charakterisieren, wie zum Beispiel Drehzahl, Drehmoment, Temperatur, Strom, Spannung, Leistung, Ladezustand der Traktionsbatterie, etc.
• - Eine Verlustgröße könnte die Abweichung zwischen gemessenen Ist-Werten und den Ziel-Werten sein, z. B. die Differenz zwischen einer Ziel-Stromtrajektorie und der tatsächlichen Ist-Stromtrajektorie;
• - Die Modellformel, die mittels der symbolischen Regression erhalten wird.

The input variables of the model-based learning reinforcement algorithm are states of the machine learning agent, such as:

• - the data and measured values of the operating parameters that characterize the drive train, such as speed, torque, temperature, current, voltage, power, state of charge of the traction battery, etc.
• - A loss size could be the deviation between measured actual values and the target values, e.g. B. the difference between a target current trajectory and the actual actual current trajectory;
• - The model formula obtained using symbolic regression.

In one embodiment, it is proposed that an electrical power loss of the electric drive train and/or an output voltage and/or an internal torque and/or an output temperature of a coolant and/or a flow velocity of the coolant is used as a parameter. Further operating parameters can be a current, in particular an alternating current, or a torque angle or an angular frequency or a coil winding temperature or a magnet temperature or a power loss or an input temperature of a coolant or a flow rate of a coolant include. The use of other parameters that characterize the electric drive train is also possible.

In one embodiment, it can be provided that the data and measured values of the at least one operating parameter include a speed of the drive unit, a torque of the drive unit, a temperature, an electrical current, an electrical voltage, an electrical power or a state of charge of a traction battery of the electric drive train. In principle, it is also possible to use data and measured values from other operating parameters.

In one embodiment, it is proposed that a DYNA algorithm be used as a model-based learning reinforcement algorithm. The use of other model-based learning reinforcement algorithms is also possible.

In one embodiment of the invention, the determination of the model formula and the change of the model formula can be carried out by means of symbolic regression by an artificial neural network. The so-called “AI Feynman algorithm” (Udrescu and Tegmark in “Science Advances”, Vol. 6, Issue 16, April 15, 2020) is particularly suitable for this.

In an alternative embodiment, it is possible that the determination of the model formula and the change of the model formula are carried out using symbolic regression using a genetic algorithm.

In one embodiment, it is proposed that the actions of the learning reinforcement agent consist in increasing, reducing or leaving the value of at least one operating parameter of the electric drive train, in particular within the framework of two-dimensional or three-dimensional maps. The learning reinforcement agent receives a positive reward if, for example, the action of the learning agent causes the difference between a target value of the parameter resulting from the symbolic regression model and the actual value of the parameter resulting from a measurement or calculation (for example the electrical power loss) reduced.

According to a further aspect, the present invention relates to a system for calibrating a control unit of an electric powertrain of a vehicle.

A system according to the invention comprises, according to claim 10, a digital processing unit and a digital storage unit. Instructions that can be read and executed by the processing unit are stored in the digital storage unit. The instructions are designed to cause the digital processing unit to carry out a method according to an embodiment of the invention when reading the instructions.

• 1 eine schematisch stark vereinfachte Darstellung, die Einzelheiten eines computerimplementierten Verfahrens zur Kalibrierung einer Steuereinheit eines elektrischen Antriebsstrangs eines Fahrzeugs veranschaulicht,
• 2 eine schematische Ansicht, welche den Ablauf der einzelnen Verfahrensschritte zeigt.

Further features and advantages of the present invention will become clear from the following description of a preferred exemplary embodiment with reference to the accompanying figures. Show:

• 1 a schematically highly simplified representation illustrating details of a computer-implemented method for calibrating a control unit of an electric drive train of a vehicle,
• 2 a schematic view showing the sequence of the individual process steps.

With reference to 1 a system 10 for carrying out a method for calibrating a control unit 1 of an electric drive train 2 of a vehicle includes a digital data memory (not explicitly shown here) and a digital processing unit (also not shown), instructions being stored in the digital data memory, the processing unit being designed to do so Read out and execute instructions, the processing unit being designed to carry out the method described below when it executes the instructions. Furthermore, a learning reinforcement agent 3 and an artificial neural network 5 are provided, which are executed by means of the digital processing unit.

The electric drive train 2 is connected to a control unit 1, which is to be calibrated using the method. At least one parameter of the electric drive train 2 is determined. Input signals of the electric drive train 2, with which an electric drive unit 20 of the electric drive train 2 is operated, are, for example, an electrical voltage or an electrical current.

The learning reinforcement agent 3 is configured to execute a reinforcement learning algorithm. A model of the electric drive train 2 is required for this. The artificial neural network 5 is designed to execute a symbolic regression model 4. The aim of the symbolic regression model 4 is to describe the electric drive train 2 as precisely as possible using at least one mathematical model formula, which in particular characterizes the at least one parameter describe. This model formula is determined by an algorithm that carries out symbolic regression.

With further reference to 2 The basic process of the computer-implemented method for calibrating the control unit 1 of the electric drive train 2 will be illustrated below.

In a first step 100, the symbolic regression model 4, which is executed using the artificial neural network 5, is initialized. The symbolic regression model 4 is designed to receive and process data and measured values of at least one operating parameter of the electric drive train 2. Typically, data and measured values from several operating parameters are received and processed. These operating parameters can include, for example, a power provided by the electric drive unit 20 of the electric drive train 2, a speed of the drive unit 20, a torque provided by the drive unit 20 and/or a temperature of the drive unit 20.

Furthermore, the symbolic regression model 4 is designed to generate at least one mathematical model formula that characterizes the at least one parameter. In addition, the symbolic regression model 4 is designed to output this mathematical model formula to the learning reinforcement agent 3. A state of the electric drive unit 2 is defined by the selection of specific values and data of operating parameters. A voltage, an alternating current, a torque angle, an angular frequency, a coil winding temperature, a magnet temperature, an electrical power loss, an input temperature of a coolant, a flow rate of a coolant or a power loss can each represent an operating parameter.

In a second step 200, a current state of the electric drive train 2 is determined, this state being defined by the data and measured values of the at least one operating parameter of the electric drive train 2 and the value of the at least one parameter depending at least partially on this operating parameter. The learning reinforcement agent 3 receives this state and the model formula of the symbolic regression model 4 as input variables and thereby observes the environment. The model formula forms a model prediction of the symbolic regression model 4.

In a third step 300, a learning reinforcement function is executed based on the current state and the model formula, in which the learning reinforcement agent 3 interacts with the control unit 1 and carries out at least one action A, which causes the control unit 1 to change the state of the electric powertrain 2 , whereby the state of the learning reinforcement agent 3 is also changed and the learning reinforcement agent 3 receives a positive or negative reward R for its action A. Possible actions A are: The value of at least one operating parameter of the electric drive train 2 is increased, reduced or not changed by means of the control unit 1. The actions A of the learning reinforcement agent 3 can consist in particular in that the values of the operating parameters of the electric drive train 2 are increased, reduced or left unchanged within the framework of two-dimensional or three-dimensional maps.

Preferably, in a state module not explicitly shown here, a calculated model value K _M of the parameter is now compared with the actually determined actual value K _R of the parameter. The calculated value K _M of the parameter is based on the changed values for the selected operating parameters by the learning reinforcement agent 3. In a reward module, which is also not explicitly shown here, the degree of deviation between the calculated model value K _M and the measured value K _R of the parameter is compared and a reward B is assigned to the degree of deviation. Since the degree of deviation depends on the selection of the respective action A, a reward is preferably assigned to the respective selected action A in a matrix or function or other database. The reward B preferably has the values +1 and -1, with a small deviation between the calculated value K _M of the parameter and the measured value K _R of the parameter being rewarded with +1 and thus reinforced, whereas a significant deviation with - 1 is rewarded and is therefore evaluated negatively. However, it is also conceivable that values >1 and values <1 are used.

The method presented here therefore uses a first algorithm, which is formed by the symbolic regression model 4 of the electric drive train 2, and a second algorithm, which is formed by the learning reinforcement agent 3. An essential basic idea is that the model of the electric drive train 2 required for the learning reinforcement algorithm (reinforcement learning algorithm) executed by the learning reinforcement agent 3 is described by at least one mathematical model formula, which is represented by a The algorithm executing the symbolic regression model 4 is determined and which is subsequently made available to the learning reinforcement agent 3.

In a fourth step 400, the current state of the electric drive train 2 is determined again. As already mentioned above, the current state is defined by the data and measured values of the at least one operating parameter of the electric drive train 2.

In a fifth step 500, the symbolic regression model 4 is executed again with the at least one operating parameter determined in the previous method step 400, with the at least one model formula for modeling the electric drive train being modified.

Subsequently, in a step 600, steps 200 to 500 are carried out repeatedly with the model formula modified in the previous method step 500 until the calibration task is solved and, for example, the electrical power loss is minimized. The learning reinforcement agent 3 thus plans its next action based on the modified model prediction of the symbolic regression model 4, from which the modified model formula results, and its current state.

QUOTES INCLUDED IN THE DESCRIPTION

This list of documents listed by the applicant was generated automatically and is included solely for the better information of the reader. The 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

• DE 102019126244 A1 [0008]

Claims (8)

Computer-implemented method for calibrating a control unit (1) of an electric drive train (2) of a vehicle, in which at least one parameter is determined, comprising the steps: - Initializing (100) a symbolic regression model (4), which is designed to receive data and measured values of at least one operating parameter of the electric drive train (2) and to generate at least one mathematical model formula which describes the parameter and to send it to a learning reinforcement agent (3 ) to spend, - Determining (200) a state of the electric drive train (2), this state being defined by data and measured values of the at least one operating parameter of the electric drive train (2) and the value of the at least one parameter depending at least partially on this operating parameter and the learning reinforcement agent (3) receives this state and the model formula of the symbolic regression model (4) as input variables and thereby observes the environment, - Executing (300) a learning reinforcement function, in which the learning reinforcement agent (3) interacts with the control unit (2) and carries out at least one action which causes the control unit (3) to change the state of the electric drive train (2), whereby also the state of the learning reinforcement agent (3) is changed and the learning reinforcement agent (3) receives a positive or negative reward for its action, - re-determining (400) the state of the electric drive train (2), the data and measured values of the at least one operating parameter being fed to the symbolic regression model (4), -Executing (500) the symbolic regression model (4) with the at least one operating parameter determined in the previous method step (400), wherein the at least one model formula for modeling the at least one parameter (K) is modified, - repeatedly executing (600) steps (200) to (500) with the model formula modified in the previous method step (500) until the calibration task is solved. Computer-implemented method Claim 1 , characterized in that an electrical power loss of the electric drive train (2) and/or an output voltage and/or an internal torque and/or an output temperature of a coolant and/or a flow velocity of the coolant is used as a parameter. Computer-implemented method according to one of the Claims 1 or 2 , characterized in that the data and measured values of the at least one operating parameter are a speed of the drive unit (20), a torque of the drive unit (20), a temperature, an electrical current, an electrical voltage, an electrical power or a state of charge of a traction battery of the electric Drive train (2) include. Computer-implemented method according to one of the Claims 1 until 3 , characterized in that the determination of the model formula and the change of the model formula are carried out by means of the symbolic regression model (4) by an artificial neural network (5). Computer-implemented method according to one of the Claims 1 until 4 , characterized in that the determination of the model formula and the change of the model formula are carried out using the symbolic regression model (4) by a genetic algorithm. Computer-implemented method according to one of the Claims 1 until 5 , characterized in that a DYNA algorithm is used as the model-based learning reinforcement algorithm that is executed by the learning reinforcement agent (3). Computer-implemented method according to one of the Claims 1 until 6 , characterized in that the value of at least one operating parameter of the electric drive train, in particular in the context of two-dimensional or three-dimensional maps, is increased, reduced or left unchanged. System for calibrating an electric drive train of a vehicle, comprising a digital data memory and a digital processing unit, instructions being stored in the digital data memory, the processing unit being designed to read out and execute the instructions, the processing unit being designed to carry out a method according to one of the previous claims when it executes the instructions.

Images