WO2022043105A1

WO2022043105A1 - Method and apparatus for controlling an electronic converter by means of machine learning methods

Info

Publication number: WO2022043105A1
Application number: PCT/EP2021/072646
Authority: WO
Inventors: Maja Rita Rudolph; Michael JIPTNER; Dennis Bura; Samuel Vasconcelos Araujo
Original assignee: Robert Bosch Gmbh
Priority date: 2020-08-27
Filing date: 2021-08-13
Publication date: 2022-03-03
Also published as: DE102020210825A1

Abstract

The invention relates to a method for operating an electronic system (1) having an electronic converter (3) controlled by at least one original control signal (S) for an electromechanical device or a power converter, said method comprising the following steps: - providing (S1) a profile of the original control signal (S) for controlling the electronic converter (3), - modifying (S2) the original profile of the at least one control signal according to a trainable data-based control signal model in order to obtain a profile of at least one modified control signal (S'), said control signal model being designed to determine the profile of the at least one modified control signal (S') depending on the profile of the at least one original control signal (S) and in particular on operating conditions of the electronic converter (3) and/or on operating properties and/or system properties; - controlling (S3) the electronic converter (3) with the at least one modified control signal (S').

Description

description

title

Method and device for controlling an electronic converter using machine learning methods

technical field

The invention relates to controlling electronic converters using control signals. In particular, the present invention relates to measures for shaping the control signals in order to achieve improved operating characteristics of the electronic converter.

Technical background

Electronic converters are controlled using one or more control signals. In many applications, control signals follow predetermined courses for different events. In power electronic converters, the commutation of transistors is often controlled via voltage or current signals via their gate inputs.

The control signals are generated by a control unit or driver unit. These signals often have steps or ramps and/or periodic profiles in order to achieve the desired behavior of the electronic system.

Disclosure of Invention

According to the invention are a method for operating an electronic converter in an electronic system according to claim 1 and a device and an electronic system according to the independent claims.

Further developments are specified in the dependent claims.

According to a first aspect, a method for operating an electronic system with an electronic converter controlled by at least one original control signal for an electromechanical device or a power converter is provided, with the following steps:

Providing a course of the original control signal for driving the electronic converter,

Modifying the original course of the at least one control signal according to a trainable, data-based control signal model in order to obtain a course of at least one modified control signal, the control signal model being designed to change the course of the at least one modified control signal as a function of the course of the at least one original control signal and in particular of operating conditions of the electronic converter and/or operating characteristics and/or system characteristics;

- Controlling the electronic converter with the one or more modified control signals.

Electronic systems often include an electronic transducer that is part of or drives electronic circuitry. The electronic converter usually has active electronic components such. B. transistors, a MOSFET or the like and is driven by one or more control signals that are provided by a control unit or control circuit. For its part, the converter often controls electromechanical actuators, electrical machines or the like.

The one or more control signals can be determined by pulse signals, their pulse shapes, periodicities, in particular their cycle frequencies, pulse duty factors and the like, which bring about a corresponding reaction in the converter. The behavior of the one or more control signals can decisively determine the behavior of the electronic system. For example, losses and faults in the electronic system, the service life of the System and loads on the converter can be largely determined by the time course of the state transition or state course of the control signal.

According to the above method, provision is made for providing a data-based control signal model in order to optimize the form (the time profile) of the control signal for driving the electronic converter. The control signal model is trained to provide an optimized/modified course of the modified control signal depending on the time course of the control signal or the course of the specified state transition or the given state course of the control signal, operating parameters of the electronic converter and its component properties and/or system properties to generate. The course of the optimized/modified control signal is optimized with regard to a behavior of the electronic converter with regard to one or more criteria.

To train the data-based control signal model, the behavior of the electronic converter can be simulated using a circuit simulation tool such as SPICE, or a circuit simulation, or with the (differential) equations that map the system behavior and those in the electronic circuit resulting response signals or the effect of applying the one or more control signals to the electronic converter are evaluated according to the criteria. So, for example, for a given control signal losses that z. B. correspond to the required switching energy, disturbances such. B. Vibrations as a step response to the state transition of the control signal, as well as loads on the electronic converter or the entire system that can affect the service life, such. B. temporary overvoltages or overcurrents, as well as high temperatures due to strong heat development.

A performance measure can be provided to optimize performance. One or more of the criteria for the behavior of the electronic system can be evaluated and, in particular, mapped to the behavior measure via a (differentiable) cost function. The behavior measure is used to train the data-based control signal model. In doing so, a optimized/modified behavior measure of the modified control signal (provided by the control signal model) is used to train the data-based control signal model by minimizing this behavior measure.

In order to train the data-based control signal model based on a behavior measure based on a circuit simulation, the cost function for determining the behavior measure and the model equations of the circuit simulation must be automatically differentiable. This allows the calculation of the behavior measure based on the original and modified control signals to be combined directly with other components of the control signal model. Since the model equations of the circuit simulation (e.g. simulation tool, especially SPICE, or differential equations that describe the system behavior) can be differentiated for the training, the model parameters of the data-based control signal model can be trained directly with gradient-based methods (e.g. backpropagation). .

In particular, the control signal model can be designed as a trainable data-based model, in particular as an artificial neural network, such as. B. a multi-layer perceptron or a recursive neural network.

According to one embodiment, the curve of the original control signal for controlling the electronic converter can be parameterized or defined by control signal parameters and/or the curve of the modified control signal can be parameterized or defined by corresponding control signal parameters.

In particular, the profile of the original control signal and/or the modified control signal can be parameterized by time segments and the values of an electrical variable, in particular a voltage or a current, assigned to the time segments. Further control signal parameters can be a cycle frequency of a periodic control, in particular a frequency and/or pulse width modulation, a degree of modulation, a duty cycle, a pulse duration and/or a pulse shape.

Furthermore, the operating conditions of the electronic converter to be controlled can relate to the operation of the electronic converter and in particular can include one or more of the following variables: one on the electronic Converter applied voltage, a current flowing through the electronic converter current at the current time and a current temperature of the electronic converter.

According to one embodiment, the operating characteristics may relate to general characteristics of the type of electronic converter, in particular a transistor, and in particular may include one or more of the following parameters: a threshold voltage, a leakage current at the gate terminal, an on-state resistance and their variances of these quantities .

Provision can be made for the system properties to relate to other components of the electronic circuit that affect the operation of the electronic converter, and in particular to include one or more of the following parameters: a thermal resistance of the overall structure, the capacitance of a back-up capacitor coupled to the electronic converter is, and variances of these quantities.

Furthermore, the control signal model can be trained by optimizing a behavior measure that results from the control signal model and a downstream circuit simulation in order to adapt model parameters of the control signal model, the circuit simulation, in particular with model equations, being designed to convert a modified control signal into one or more behavior variables to be determined, a behavior measure being able to be determined as a function of the one or more behavior variables, which indicate the behavior of the electronic system in response to modified control signals, as a function of a predefined cost function.

In particular, the one or more behavioral variables can indicate a performance of the electronic system when controlled by the modified control signal, the behavioral variable being a power loss, a degree of interference, in particular in the form of oscillations or overshoots, a thermal load in particular on the electronic system and/or a indicates an expected service life of the electronic converter or the electronic circuit influencing the load level ?? Sentence.

It can be provided that the optimization is carried out taking into account a side condition that the achievement of the control signal in the electronic system function ensured by the modified control signal and / or wherein the cost function for calculating the measure of behavior takes into account a parameter that evaluates how the function caused by the control signal is achieved by the modified control signal.

According to a further aspect, a device, in particular a control signal model block, is provided for operating an electronic system with an electronic converter controlled by at least one original control signal for an electromechanical device or a power converter, the device being designed for:

- Controlling the electronic converter with the at least one modified control signal.

According to a further aspect, an electronic system with a converter, with an electronic circuit and the above device is provided.

Brief description of the drawings

Embodiments are explained in more detail below with reference to the accompanying drawings. Show it:

FIG. 1 shows a schematic representation of an electronic system with an electronic converter which is to be controlled in an optimized manner with one or more optimized control signals; FIGS. 2a-2c diagrams to illustrate signal responses of the electronic circuit;

FIG. 3 shows a possibility for parameterizing the control signal; and

Figure 4 is a flowchart to illustrate a method for operating the electronic converter in the electronic system of Figure 1.

Description of Embodiments

Figure 1 shows a schematic representation of an electronic system 1 with an electromechanical device 2, such as. B. an electrical machine, an electromechanical actuator, or a power converter, which is coupled in any way with an electronic converter 3. The electronic converter 3 can be, for example, an inverter, an H-bridge circuit, a B6 bridge circuit, any other electronic power circuit with transistors, such as. B. MOSFETS o. Ä., Or the like.

The electronic converter 3 is activated as a function of one or more control signals S in order to carry out a function in the electromechanical device 2, it being possible for the electronic converter 3 to be part of the electromechanical device 3 or separate therefrom.

The one or more control signals S are provided by a control unit 4 . The one or more control signals S can be a current or voltage signal and can be applied to a control input of the electronic converter 3 in order to implement a desired function. The one or more control signals S are presented as a time course of an electrical condition, such as e.g. a voltage or a current, and may include state transitions or state histories. Using a control signal model block 5, the one or more original control signals S provided by the control unit 4 are modified with a control signal model and provided as corresponding modified control signals S′.

For example, FIGS. 2b and 2c show signal curves of a voltage or a current of a resulting signal in the electromechanical device 2 based on a curve of a provided control signal S shown in FIG Flank and / or have a current peak.

Furthermore, the actuation with the one or more control signals S in the electromechanical device 2 can generate power losses due to the switching energy consumed, which depend on the course of the control signals S over time. The electronic converter 3 or the electronic system 1 can also be subject to loads depending on the shape of the curve of the control signal S, which can impair the service life of the electronic system 1 . This can be caused, for example, by temporary overvoltages or overcurrents and high temperatures. The behavior of the electronic system 1 can be evaluated according to various criteria and mapped to a behavior variable. By varying the course and/or form of the control signals S, the various behavior variables can be varied in order to improve the behavior of the electronic system 1 .

The one or more behavior variables can be determined by a circuit simulation, such as in the SPICE programming language, with which the reaction of the electromechanical device 2 to the course of any control signals can be determined. For this purpose, response signals to the control signals S are modeled and evaluated using one or more behavior variables according to one of the criteria mentioned above (power loss, switching losses, overshoot, load, etc.). Furthermore, the one or more behavior variables can be combined using a (differentiable) cost function in order to determine a behavior measure that characterizes a behavior of the electronic system 1 . For example, the measure of behavior can be calculated by a sum of the weighted measures of behavior. The weightings can be specified according to an optimization criterion. Limit values can be used as criteria, e.g. B. be defined for temperatures, voltages or the like, so that the behavior measure is determined depending on a distance of the respective behavior variable from the respective limit value.

In order to process the control signal S in the control signal model block 5, it must be parameterized in a suitable manner. This can take place in the control unit 4 or on the input side in the control signal model block 5 . In the latter case, a provided analog control signal can be sampled and parameterized in a suitable manner.

The control unit 4 can specify the one or more original control signals by time segments and the values of an electrical variable, in particular a voltage or a current, assigned to the time segments.

Thus, the control signal S can be defined in several time intervals/time steps with different durations t1, t2, . . . , tn and with respective amplitudes A1, A2, . The modified or optimized control signal S' to be determined by the control signal model can be parameterized in the same way or in a different way.

Alternatively, the control unit 4 can also indicate the one or more control signals by one or more parameters of a periodic activation, such as e.g. B. a cycle frequency of a periodic control, in particular a frequency and / or pulse width modulation, a degree of modulation, a duty cycle, a pulse duration and / or a pulse shape.

The control signal model can be a data-based, trainable model, in particular an artificial neural network, or a regression model. In the present case, an artificial neural network is assumed as a model, since this can be trained in a simple manner by differentiation. In principle, the control signal model can be implemented in different variants. On the one hand, the control signal model can be implemented as hardware or software in the control unit 4 as part of the electronic system 1 or separately thereto. Alternatively, in one variant, the control signal model can be implemented in the control unit 4 with an adaptation function, in which the control signal model is retrained accordingly at predetermined points in time or regularly in order to correct aging effects in the electronic system 1 . An implementation in the latter variant is described in more detail below with reference to a flowchart in FIG.

The method presented there can be implemented in a control unit 4 using hardware and/or software.

For this purpose, in step S1, the control unit 4 of the electronic system provides parameters for a suitable parameterization of the time profile of the one or more control signals S that are to be applied to the electronic converter 3 in sequence. In this case, the original control signal S is provided in parameterized form by the control unit 4 or provided as a progression of the state of an electrical variable (current or voltage). In the latter case, the parameterization can be carried out in the control signal model block 5 .

As described above, the parameterization can provide, for example, for the control signal to be divided into a number of time steps and for a corresponding amplitude value to be assigned. In this way, an amplitude can be specified for a number of upcoming magazines.

In step S2, the one or more parameterized, original control signals S are transmitted to the control signal model together with the operating conditions of the electronic converter 3 to be controlled and operating properties and/or system properties.

The operating conditions relate to the operation of the electronic converter 3 and can include one or more of the following variables: a voltage present at the electronic converter 3 and a current flowing through the electronic converter 3 at the current time, and a current Temperature of the electronic converter 3. The operating characteristics can relate to general characteristics of the type of electronic converter 3 and concern one or more of the following parameters: threshold voltage, leakage current at the gate terminal, on-state resistance as well as their variances, which are caused either by scattering due to Manufacturing tolerances or due to aging effects.

Furthermore, the system properties can relate to other components of the electronic circuit that can affect the operation of the electronic converter 3 . For example, the system properties can include one or more of the following parameters of other system components of the electronic circuit: a thermal resistance of the overall structure, which is decisive for the temperature of the electronic converter 3 to be controlled, the capacitance of a back-up capacitor, which is coupled to the electronic converter 3 , and the same. Correspondingly, variances in these variables, which result either from manufacturing tolerances or from aging, can also be taken into account as such parameters.

From the corresponding input variables, the control signal model determines a modified control signal that can be used to control the electromechanical device 2 .

The result is a modified course of a modified control signal S'. This is provided in parameterized form.

In a subsequent step S3, the electronic converter 3 is controlled in accordance with the profile specified by the modified control signal parameters of the modified control signal S'. The conversion of the modified control signal parameters, which define the modified control signal S', into the analog control signal S' can preferably take place in the control signal model block 5 or in a separate device.

An adaptation criterion can be checked in step S4. If the adaptation criterion provides for a further adaptation or update of the control signal model (alternative: yes), the method continues with step S5, otherwise (alternative: no) the process jumps back to step S1. Adaptation criteria can depend, for example, on a predetermined period of time since the last adaptation, on the course of the original control signal, or on an external adaptation signal. Furthermore, an update of the control signal model can be triggered if system properties such. B. a temperature, voltage values, measurement signal and the like, deviate from corresponding predetermined reference values by more than a predetermined amount of deviation. In particular, the adaptation criteria should be used to check whether the control signal model needs to be retrained due to component aging, wear or other systematic changes in the operating conditions.

In step S5, the control signal model is retrained.

The retraining of the control signal model serves to correct inaccuracies in the circuit simulation and the component models on which it is based in the control signal model, e.g. B. due to component aging, wear or other systematic changes in the operating conditions, the learned control signal (adjustments) are no longer optimal.

The retraining of the control signal model aims to generate modified control signals from given control signals, for which the behavior measure is optimized. In particular, the cost function of the behavior measure and the system of equations of the circuit simulation can be used in connection with a gradient descent method in order to further train the model parameters of the control signal model.

The control signal model is initially created using a training data set made up of control signals and associated modified control signals. The modified control signals are each determined based on a specified control signal and depending on a behavior measure that is determined from one or more behavior variables that are determined for the modified control signal using a circuit simulation and the specified cost function. To build up the training data, control signals are specified with as many variants as possible within the scope of the possible control signals for driving the transistor 3 .

The circuit simulations can determine the one or more corresponding behavior variables or the resulting behavior measure (depending on a predetermined cost function) based on the parameterization of the modified control signal S'.

Using an optimization method such as B. stochastic gradient descent, a training data record for the control signal model can be performed by optimizing (minimizing) the behavior measure with respect to a relevant control signal, in particular by backpropagation. This is possible because, as a rule, the functions underlying a circuit simulation can be differentiated, so that the optimized performance measure can be determined by differentiating the cost function and the function of the circuit simulation for updating the model parameters of the control signal model.

So that the function brought about by the control signal in the electronic system can also be achieved by the modified control signal, the optimization can be carried out with suitable secondary conditions. As an alternative or in addition, a parameter can also be taken into account in the cost function for calculating the behavior measure, which parameter evaluates how the function brought about by the control signal is achieved by the modified control signal.

In particular, the control signal model in the form of a neural network, such as. B. a recurrent neural network (LSTM, GRU), a multi-layer perceptron or the like. Thus, the control signal model can be trained using a back-propagation method based on minimizing the behavior measure.

In order to relieve the control unit 4 of the electronic system 1 , the control signal model can also be retrained externally to the electronic system 1 . For this purpose, the parameterized original and modified control signals S, S' as well as the operating conditions and the component parameters and system parameters are transmitted to an external computing unit, which carries out the circuit simulations in the knowledge of the electromechanical device 3 in order to calculate the behavioral variables or the behavioral measure. The model parameters of the neural network can thus be retrained externally and the model parameters can be transmitted back to the electronic system 1 so that they can be used in sequence.

In a further embodiment, the control signal model can also be implemented as a lookup table in the electronic system, so that the control signal parameters of the original control signal are assigned according to the lookup table to modified control signal parameters which represent the modified control signal S'. The lookup table is created based on a control signal model that can be implemented externally to the electronic system 1 . In this way, the computing effort in the electronic system 1 can be significantly reduced.

Claims

Expectations

1. A method for operating an electronic system (1) with an electronic converter (3) controlled by at least one original control signal (S) for an electromechanical device or a power converter, with the following steps:

Providing (S1) a profile of the original control signal (S) for driving the electronic converter (3),

Modifying (S2) the original course of the at least one control signal according to a trainable, data-based control signal model in order to obtain a course of at least one modified control signal (S'), the control signal model being designed to determine the course of the at least one modified control signal (S') as a function of to determine the course of the at least one original control signal (S) and in particular operating conditions of the electronic converter (3) and/or operating properties and/or system properties;

- Controlling (S3) the electronic converter (3) with the at least one modified control signal (S').

2. The method according to claim 1, wherein the control signal model is designed as a trainable data-based model, in particular as an artificial neural network, such as. B. Multi-Layer Perceptron, or recursive neural network.

3. The method according to claim 1 or 2, wherein the profile of the at least one original control signal (S) for driving the electronic converter (3) is parameterized or defined by control signal parameters and/or wherein the profile of the at least one modified control signal (S') is corresponding modified control signal parameters is parameterized or defined.

4. The method according to claim 3, wherein the parameterization of the course of the at least one original control signal (S) and/or the at least one mo- modified control signal by time intervals and the values of an electrical variable assigned to the time intervals, in particular a voltage or a current, a cycle frequency of a periodic control, in particular a frequency and/or pulse width modulation, a degree of modulation, a duty cycle, a pulse duration and/or a pulse shape will. Method according to one of Claims 1 to 4, in which the operating conditions of the electronic converter (3) to be controlled relate to the operation of the electronic converter (3) and in particular comprise one or more of the following variables: a voltage present at the electronic converter (3), a current flowing through the electronic converter at the current time and a current temperature of the electronic converter (3). Method according to one of Claims 1 to 5, in which the operating characteristics relate to general characteristics of the type of electronic converter (3) and in particular comprise one or more of the following parameters: a threshold voltage, a leakage current at the gate terminal, an on-state resistance and the variances of these quantities. The method according to any one of claims 1 to 6, wherein the system properties relate to other components of the electromechanical device or the power converter that affect the operation of the electronic converter (3), and in particular one or more of the following parameters include: a thermal resistance of the overall structure, a Capacitance of a back-up capacitor coupled to the electronic converter (3) and variances of these quantities. Method according to one of Claims 1 to 7, in which the control signal model is trained by optimizing a behavior measure which results from the control signal model and a downstream circuit simulation in order to adapt model parameters of the control signal model, the circuit simulation being designed to generate from a modified control signal ( S') to determine one or more behavioral variables, a behavioral measure being able to be determined as a function of a predetermined cost function, depending on the one or more behavioral variables which indicate the behavior of the electronic circuit (2) in response to modified control signals,. 17 The method according to claim 8, wherein the one or more behavioral variables indicate a performance of the electronic system when controlled by the modified control signal, the behavioral variable being a power loss, a degree of interference, in particular in the form of a strength of oscillations or an overshoot, a particular indicates thermal stress on the electronic system and / or a degree of stress that is influenced in particular by an expected service life of the electronic converter (3) or the electromechanical device or the power converter. The method according to claim 8 or 9, wherein the optimization is carried out taking into account a constraint that ensures that the modified control signal achieves the function caused by the control signal in the electronic system and/or wherein the cost function for calculating the behavior measure takes into account a parameter that evaluates how the function caused by the control signal is achieved by the modified control signal. Device, in particular control signal model block (5), for operating an electronic system (1) with an electronic converter (3) controlled by at least one original control signal (S) for an electromechanical device or a power converter, the device being designed for:

Providing a course of the original control signal (S) for driving the electronic converter (3),

Modifying the original course of the at least one control signal according to a trainable, data-based control signal model in order to obtain a course of at least one modified control signal (S'), the control signal model being designed to change the course of the at least one modified control signal (S') depending on the course of the to determine at least one original control signal (S) and in particular operating conditions of the electronic converter (3) and/or operating properties and/or system properties;

- Controlling the electronic converter (3) with the at least one modified control signal (S'). Electronic system with a converter (3), with an electronic circuit (2) and a device according to claim 11. 18. A computer program product comprising instructions which, when the program is executed by a computer, cause it to carry out the steps of the method according to any one of claims 1 to 10. A machine-readable storage medium comprising instructions which, when executed by a computer, cause it to perform the steps of the method of any one of claims 1 to 10.