Method and apparatus for controlling electronic switches in a technical system by means of artificial intelligence

Abstract

The invention relates to a method for operating a technical device having an electronic converter controlled by at least one control signal, having the following steps: providing a provided control signal time sequence consisting of one or more successive original control signals, with which the electronic converter should be operated; predicting a predicted control signal time sequence consisting of one or more future control signals from the first control signal time sequence; modifying the provided control signal time sequence on the basis of the provided control signal time sequence and the predicted control signal time sequence in accordance with a trainable data-based control signal model in order to obtain a modified control signal time sequence consisting of one or more successive control signals, wherein the control signal model is trained to determine the modified control signal time sequence from the provided control signal time sequence and the predicted control signal time sequence; operating the electronic converter in accordance with the waveform of the modified control signal time sequence.

Description

Technical Field

[0001] This invention relates to the control of electronic converters in a technical system by means of control signals. More particularly, this invention relates to measures for adapting control signals to achieve improved operating characteristics of the technical system. Background Technology

[0002] An electronic converter is an electronic circuit having one or more converter devices, such as transistors, MOSFETs, IGBTs, etc., to provide defined electrical parameters to control a technical system, such as a motor.

[0003] Electronic converters are operated by means of control signals, which are applied to the base or gate connections of one or more active converter devices. In many applications, the control signals follow fixed, pre-defined waveforms for different events and are often used to implement the transition from a cutoff state to a conduction state in a defined manner.

[0004] Control signals are generated by the control unit or other drive unit. These signals often have stepped or ramped waveforms to achieve the desired switching or transmission characteristics of the converter device.

[0005] Such converters, such as those used in the electric traction systems of motor vehicles, affect the operating status of the traction system. Thus, operating characteristics may include the vehicle's range and the aging condition of components such as the traction battery pack, electronic converter, and traction motor. Summary of the Invention

[0006] According to the present invention, a method for operating a technical apparatus having an electronic converter controlled by at least one control signal is proposed, as well as a corresponding device and a corresponding technical system.

[0007] Other design schemes are described in different implementation methods.

[0008] According to the first aspect, a method is provided for operating a technical device having an electronic converter controlled by a control signal, the method comprising the following steps:

[0009] - Provides the control signal waveforms that should be used to operate the electronic converter;

[0010] - Predict the predicted control signal waveform based on the provided control signal waveform, as the predicted future waveform of the control signal;

[0011] - Modify the provided control signal waveform according to the provided control signal waveform and the predicted control signal waveform using a trainable data-based control signal model to obtain a modified control signal waveform.

[0012] - The control signal model is trained to determine the modified control signal waveform based on the provided control signal waveform and the predicted control signal waveform;

[0013] - Operate the electronic converter according to the modified control signal waveform.

[0014] Electronic systems typically include transducers, which are part of or control electronic circuitry. Such transducers, especially transducer devices, or active electronic devices like transducers, are controlled by a sequence of control signal values ​​provided by a control unit. In this document, the sequence of control signal values ​​is referred to as the control signal time series. For the sake of simpler signal processing, the control signal time series is assumed to be time-discrete. Furthermore, a transducer can also be controlled by more than one control signal waveform in the same or different ways.

[0015] The control signal waveform or time series used to operate the converter may have state transitions or state waveforms, particularly in the form of potential or current changes along or ramping, which cause corresponding responses in downstream electronic circuitry. In the case of an electronic converter, the control signal time series may have state transitions or state waveforms, particularly in the form of potential or current changes along or ramping, which follow a periodicity and cause corresponding responses in the associated technical device.

[0016] The waveform of the control signal, especially the form of state transitions (voltage steps, voltage ramps, etc.), can decisively determine the behavior of the converter and the technical device that operates therethrough. For example, the power loss and interference of the technical device, the lifespan of the system, and the load on the converter device or electronic converter can be determined decisively by the state waveform, state transition, or state waveform of the control signal waveform or the time sequence of the control signal over time.

[0017] According to the above method, a data-based control signal model is established to optimize the shape (waveform over time) of the control signal waveform for the operation of the electronic converter. The control signal model is trained to generate a modified control signal waveform to operate the converter device and ultimately the technical apparatus, based on the provided control signal waveform (i.e., the control signal time series) and the predicted control signal waveform (i.e., the predicted control signal time series that updates the provided control signal waveform). Here, the optimized / modified control signal waveform is optimized in terms of the predicted long-term behavior of the technical apparatus.

[0018] The control signal model can also be constructed to determine the modified control signal waveform based on one or more of the following parameters: one or more operating parameters of the electronic converter and / or the technical device to be controlled; one or more operating characteristics of the electronic converter and / or the technical device to be controlled; one or more system characteristics of the electronic converter and / or the technical device to be controlled; and one or more system parameters of the technical device.

[0019] Here, operating parameters can affect the operation of the technical device and include, in particular, one or more of the following parameters: the thermal resistance of the overall structure; the capacitance of the standby capacitor (Stützkondensator) coupled to the electronic components; and the variance of these parameters.

[0020] It can be specified that the control signal model is constructed as a trainable, data-based model, especially as an artificial neural network, such as a multi-layer perceptron or a recurrent neural network.

[0021] Furthermore, the control signal waveforms provided and predicted for manipulating the electronic converter can be parameterized or limited by control signal parameters, and / or the modified control signal waveforms can be parameterized or limited by corresponding modified control signal parameters.

[0022] In particular, parameterization of the provided control signal waveform and / or modified control signal waveform can be performed by time periods and electrical parameters, especially voltage or current, assigned to those time periods.

[0023] Alternatively, parameterization of the provided control signal waveform and / or the modified control signal waveform can be performed by periodically manipulating one or more parameters, particularly the periodic frequency, and in particular frequency and / or pulse width modulation, modulation rate, duty cycle, pulse duration and / or pulse shape.

[0024] In one implementation, the prediction of the predicted control signal waveform can be performed using a data-based prediction model trained to determine the predicted control signal waveform based on the provided control signal waveform. This data-based prediction model includes, in particular, recurrent neural networks, state-space models, sequence-to-sequence models, or NARXGP models.

[0025] Modifications to the provided control signal time series can be performed according to a trainable, data-based control signal model based on the waveform of at least one original control signal and the future waveform of at least one original control signal, as well as based on one or more operating parameters that describe the state of the technical device to be controlled.

[0026] Modifications to the provided control signal waveform can also be performed according to a trainable, data-based control signal model based on the provided control signal waveform and the predicted control signal waveform, as well as based on one or more operating parameters of the technical device, which in particular describe the state of the technical device to be controlled based on the predicted control signal waveform.

[0027] According to another aspect, a method for training a control signal model, particularly for use in the above-described method, is specified, wherein the control signal model is constructed to determine a modified control signal waveform based on a provided control signal waveform and a predicted control signal waveform based on the provided control signal waveform. This method comprises the following steps:

[0028] - Provide a training dataset, which includes the provided control signal waveform that should be used to manipulate the electronic converter (2) and the predicted control signal waveform;

[0029] - The control signal model is trained based on a loss function such that the provided control signal waveform and the predicted control signal waveform are mapped to a modified control signal waveform, wherein the loss function depends on one or more behavioral metrics of the technical device when the predicted control signal waveform is used to manipulate the converter.

[0030] Furthermore, the one or more behavioral metrics can each characterize the properties of the technical device depending on one or more behavioral parameters, which describe the behavior of the electronic circuitry depending on the modified control signal waveform, wherein in particular the corresponding behavioral metrics are determined according to a pre-given cost function for evaluating the behavioral parameters.

[0031] It can be specified that behavioral parameters are determined by means of measurement and / or circuit simulation.

[0032] Control signal models can be constructed in particular as trainable, data-based models, especially as artificial neural networks, such as multilayer perceptrons or recurrent neural networks.

[0033] The loss function can also depend on a weighted average of behavioral metrics.

[0034] The cost function maps multiple behavioral parameters to behavioral metrics and is, in particular, differentiable, allowing it to be applied in conjunction with gradient-based training methods for training control signal models. This loss function can, in particular, depend on the overall behavioral metric, which considers the behavioral metrics at each time step of the modified control signal waveform.

[0035] It can be specified that the one or more behavioral parameters describe the functional capability of the technical device when operated by a modified control signal waveform, wherein in particular the one or more behavioral parameters describe power loss, interference measures, especially measures concerning the occurrence of oscillations or overshoots, thermal load on the technical device, and / or load measures affecting the converter device or the expected service life of the technical device.

[0036] To train a data-based control signal model, the behavior of the device can be simulated using circuit simulation tools such as SPICE or by employing (differential) equations that map system behavior. The response signal obtained in the device or the effect or action of applying a control signal time series to the converter device can be evaluated using behavioral metrics according to an evaluation criterion (cost function). Thus, for example, for a pre-given modified control signal time series, one can evaluate, for example, power loss corresponding to the required switching energy, interference, such as oscillations as a step response to a state transition of the control signal, and loads that may impair the lifespan of the converter device or the entire system, such as temporary overvoltages or overcurrents, and high temperatures due to intense heating.

[0037] To optimize operational behavior, behavioral metrics can be provided using a cost function. Here, one or more criteria for the behavior of the electronic system can be evaluated and, in particular, mapped to this behavioral metric via a (differentiable) cost function. To train a data-based control signal model, behavioral metrics are used during the time steps of a modified control signal time series. Here, optimized / modified behavioral metrics of the modified control signal time series (provided by the control signal model) are used to train the data-based control signal model in such a way that the resulting loss function is considered or used, which maps the overall behavioral metric to a function of the modified control signal time series during these time steps.

[0038] To train a data-based control signal model based on behavioral metrics derived from circuit simulation, the cost function used to determine the behavioral metrics, the loss function that combines these metrics, and the model equations from the circuit simulation must be automatically differentiable. Thus, the computation of behavioral metrics based on the original and modified control signals can be directly integrated with the other components of the control signal model. Since the model equations from the circuit simulation (e.g., simulation tools, especially SPICE, or differential equations describing system behavior), the cost function, and the loss function are all differentiable for this training, the model parameters of the data-based control signal model can be directly trained using gradient-based methods such as backpropagation.

[0039] According to another aspect, an apparatus is provided for operating a technical device having an electronic converter controlled by at least one control signal, wherein the apparatus is configured to:

[0040] - Provides the control signal waveforms that should be used to operate the electronic converter;

[0041] - Predict the predicted control signal waveform based on the provided control signal waveform, as the predicted future waveform of the control signal;

[0042] - Modify the provided control signal waveform according to the provided control signal waveform and the predicted control signal waveform according to the trainable data-based control signal model in order to obtain the modified control signal waveform;

[0043] The control signal model is trained to determine the modified control signal waveform based on the provided control signal waveform and the predicted control signal waveform.

[0044] - Operate the electronic converter according to the modified control signal waveform.

[0045] According to another aspect, a device for training a control signal model, particularly for use in the above-described method, is specified, wherein the control signal model is configured to determine a modified control signal waveform based on a provided control signal waveform and a predicted control signal waveform based on the provided control signal waveform, wherein the device is configured to:

[0046] - Provide a training dataset, which includes the provided control signal waveform that should be used to manipulate the electronic converter and the predicted control signal waveform;

[0047] - The control signal model is trained based on a loss function such that the provided control signal waveform and the predicted control signal waveform are mapped to a modified control signal waveform, wherein the loss function depends on one or more behavioral metrics of the technical device when the modified control signal waveform is used to manipulate the converter.

[0048] According to another aspect, an electronic system is provided, which has technical means and the aforementioned equipment, the technical means including a converter device for an electronic converter. Attached Figure Description

[0049] The implementation methods will then be described in more detail with reference to the accompanying drawings. Wherein:

[0050] Figure 1 A schematic diagram of an electronic system is shown, which has a technical device controlled by an electronic converter for manipulation using a modified control signal waveform;

[0051] Figure 2 a-2c shows a graph used to illustrate the signal response of the electronic converter;

[0052] Figure 3 This demonstrates the possibilities for parameterizing control signals;

[0053] Figure 4 This illustrates the method used to explain the operation in Figure 1 Flowcharts of methods for technical devices in technical systems; and

[0054] Figure 5 A block diagram illustrating the training of the control signal model is shown. Detailed Implementation

[0055] Figure 1 A schematic diagram of a technical system 1 with an electronic converter 2 is shown. This electronic converter includes a converter device 3 in any manner, but is exemplarily represented here as a transistor 3. The converter device 3 can be a bipolar transistor, a field-effect transistor, an IGBT, a MOSFET, etc. The electronic converter 2 can be configured, for example, as an inverter, particularly in the form of an H-bridge, a B6 circuit, etc.

[0056] The converter device 3 is operated according to the control signal S to implement the functions in the converter 2. Thus, the technical device 6, including the converter 2, can be controlled in a desired manner through the converter. For example, the converter 2 can be a power converter for controlling a power consumer (technical device), such as a traction drive device 6b of an electric vehicle controlled by a vehicle battery pack 6a.

[0057] The control signal S is provided by the control unit 4 as a signal waveform. The control signal S can be a current or voltage signal and is applied to the control input (base, gate) of the converter device 3 to achieve the desired function. The control signal S is preferably provided discretely over time as a time sequence of control signal waveforms of electrical states, such as voltage or current, and may include state transitions or state waveforms.

[0058] Furthermore, based on the control signal time series S consisting of multiple control signal values ​​at consecutive time steps, the predicted control signal time series S consisting of control signal values ​​at one or more time steps t up to the prediction range T can be estimated in a trainable data-based prediction block 7. t+1 ...S t+T The predicted control signal time series S t+1 ...S t+T This corresponds to the predicted control signal waveform.

[0059] Prediction block 7 includes a prediction model, which can be constructed as a recurrent neural network (LSTM, GRU), a state-space model, a Sequence2Sequence model, or a NARXGP model.

[0060] The prediction model can be trained by predicting one or more subsequent values ​​(values ​​at the next time step) of the control signal at a given time series (either the provided control signal time series or the provided control signal waveform) as the predicted control signal waveform. This training is based on conventional training methods (such as backpropagation or other training methods) using the current control signal time series, where the loss function used to train the prediction model describes the difference between the predicted value and the known value at the corresponding next time step of the provided control signal time series.

[0061] With the help of control signal model block 5, the original control signal waveform S provided by control unit 4 is based on the predicted control signal waveform S t+1 ...S t+T The control signal model is used to modify and is provided as the modified control signal waveform S'.

[0062] The provided control signal waveform, the predicted control signal waveform, and the modified control signal waveform can each include the control signal waveform within a pre-given time window. Thus, the predicted control signal waveform can relate to the control signal waveform over a predetermined time interval from the current evaluation time point to a future time interval. Correspondingly, the modified control signal waveform S' can relate to another predetermined time interval from the current evaluation time point to a future time interval, which can be shorter, longer, or equal to the predicted control signal time interval. The provided control signal waveform can relate to control signal waveforms occurring in the past up to the current evaluation time point.

[0063] The control signal model can also obtain the instantaneous operating parameters of the technical device 6 as other input parameters, such as component temperature, cooling circuit temperature, power consumption, current or voltage amplitude, and operating time to date.

[0064] For example, Figure 2 b and 2c show the electronic converter 2 based on... Figure 2 Figure a shows the voltage or current signal waveform of the signal obtained from the control signal S provided by control unit 4. Different step responses (to a step of the control signal) can be observed, which have, on the one hand, damped oscillations, overshoot, excessively flat edges, and / or current peaks.

[0065] The control signal model can be trained to optimize a behavior metric that evaluates the future behavior of the technical device 6. Therefore, the control signal model generates a modified control signal time series S', which is determined according to the target specification.

[0066] For example, manipulation using a modified control signal time series S' may result in power loss in the electronic converter 2 due to the switching energy consumed, the power loss depending on the dynamics of the control signal time series S. The shape of the waveform of the control signal time series S may also create loads on the electronic converter 2 or the electrical device 6 thereby controlled, which may impair the service life of the device 6. This may be caused, for example, by temporary overvoltages or overcurrents and high temperatures, which may accelerate the aging of the converter device 3, converter 2, and / or device 6.

[0067] For example, the technology device 6 may include an electric drive unit operated by the vehicle battery pack, wherein the manipulation may affect the aging of the vehicle battery pack and the electric drive unit, as well as affect energy consumption or efficiency.

[0068] The behavior of the technical device 6 can be evaluated according to different criteria and mapped to a behavior metric at each time step, which is determined based on one or more behavior parameters according to a pre-given cost function. By changing the shape of the control signal time series, the one or more behavior parameters and the assigned behavior metric can be changed for each time step. In particular, the control signal time series can be modified for the purpose of improving the behavior of the electronic converter 2 and the technical device 6 regarding long-term effects, such as the aging of components of the technical system 1.

[0069] The one or more behavioral parameters may include electrical parameters such as current, voltage, power loss, efficiency, etc., and are determined through circuit simulation, such as circuit simulation using a programming language like SPICE. This circuit simulation allows determination of the response of the electronic converter 2 and the technical device 6 to the waveform of any control signal S. For this purpose, the response signal to the control signal S is modeled / simulated / or measured in an experimental setup and evaluated using one or more behavioral parameters according to one of the aforementioned criteria (power loss, switching loss, overshoot metric, load, etc.). Thus, for example, temperature development in the converter 2 can be monitored. Since higher temperatures accelerate component aging, the impact on the aging of the converter 2 can be obtained from the waveform of temperature over time.

[0070] The determination of behavioral parameters based on the applied control signals enables the determination of a behavioral metric, which describes the quality of the operational behavior of the technical device 6 with respect to a long-term standard for each considered time step. In this way, a control signal model can be trained to provide modified control signal waveforms with respect to the desired behavioral aspects of the technical device. This desired behavior can be described by an appropriate cost function based on the behavioral parameters in the behavioral metric.

[0071] Furthermore, the cost function used to combine the one or more behavioral parameters can be differentiable in order to determine the behavioral metric. The behavioral metric characterizes the behavior of the technical device 6 or the entire technical system 1 over a time step with respect to pre-given standards, such as aging, energy efficiency, etc. For example, the behavioral metric can be calculated by a weighted sum of behavioral parameters. The weights can be pre-given according to optimization criteria. For example, limits for temperature, voltage, etc., can be defined as standards, such that the behavioral metric is determined based on the distance between the corresponding behavioral parameter and the corresponding limit.

[0072] The control signal model can be: a data-based trainable model, especially an artificial neural network; or a regression model. In the current case, we assume an artificial neural network as the model because it can be easily trained by differentiation.

[0073] In order to process the provided control signal time series S in control signal model block 5 t The control signal time series must be parameterized in an appropriate manner. This can be implemented in control unit 4 or at the input side of control signal model block 5. In the latter case, the provided analog control signal can be sampled to provide it as a time-discrete control signal time series, and the provided analog control signal can be parameterized in an appropriate manner.

[0074] Therefore, the control unit 4 can pre-define the time sequence of the control signal by the values ​​sampled at consecutive time points or time steps.

[0075] Alternatively, the control unit 4 can pre-define one or more original control signals S by assigning values ​​of time periods and electrical parameters, especially voltage or current, to these time periods.

[0076] Thus, as exemplarily in Figure 3 As shown, the control signal S can be defined in multiple time intervals / time steps using different durations t1, t2, ..., tn and corresponding amplitudes A1, A2, ..., An. The modified or optimized control signal waveform S' to be determined by the control signal model can be parameterized in the same way or in a different way.

[0077] Alternatively, the control unit 4 may also define one or more control signals by periodically manipulated parameters, such as the periodic frequency of the periodic manipulation, especially frequency and / or pulse width modulation, modulation rate, duty cycle, pulse duration and / or pulse shape.

[0078] In principle, the control signal model can be implemented in different variations. On one hand, the control signal model can be implemented as hardware or software in the control unit 4, which is part of the electronic system 1, or implemented separately from the control unit. Alternatively, in one variation, the control signal model can be implemented in the control unit 4 using an adaptation function, wherein the control signal model is retrained correspondingly at predetermined time points or periodically to correct for aging effects in the electronic system 1. Finally, the implementation schemes in the variations are discussed. Figure 4 The flowchart is described in more detail.

[0079] The method shown there can be implemented in control device 4 by means of hardware and / or software.

[0080] Therefore, in step S1, the control unit 4 of the electronic system 1 provides a control signal time sequence S as a control signal waveform, which should be applied sequentially to the converter device 3. Here, the provided control signal time sequence S is parameterized by the control unit 4 or provided as a state waveform of an electrical parameter (current or voltage). In the latter case, the parameterization can also be performed in the control signal model block 5.

[0081] The provided control signal time series S may include various current or voltage steady segments or control signal sequences consisting of changing current or voltage.

[0082] In step S2, the provided control signal time series S is fed to the prediction model so as to determine the predicted control signal time series S based on the control signal time series S detected up to the current time point. t+1 ...S t+T .

[0083] In step S3, the operating conditions of the electronic converter 2 and / or the technical device 6 to be controlled, as well as the operating characteristics of the converter 2 and / or the technical device 6 and / or the system characteristics of the converter 2 and / or the technical device 6, are detected.

[0084] In step S4, the parameterized original control signal time series S, especially the parameterized original control signal time series caused by continuously transmitting the current value of the control signal (for storage in control signal model block 5), and the parameterized predicted control signal time series S t+1 ...S t+T In particular, it is transmitted to the data-based control signal model along with operating conditions, operating characteristics, and / or system characteristics.

[0085] Operating parameters relate to the operation of converter 2 or technical device 6 and may include one or more of the following parameters: the voltage applied to converter device 3, particularly to the transistor of converter device 3; the current flowing through converter device 3, particularly through the transistor of converter device 3, or through technical device 6 at the current time; and the current temperature of converter device 3 or technical device 6. Operating characteristics may generally relate to the characteristics of the type of converter device 3 or the characteristics of the type of transistor of converter device 3 (transistor characteristics), the characteristics of converter 2 or technical device 6, and to one or more of the following parameters: threshold voltage, leakage current at the gate connection, resistance in the on-state, and its variance, which arises either from deviations due to manufacturing tolerances or from aging effects.

[0086] Furthermore, system characteristics may involve other components of the electronic circuitry that may affect the operation of the electronic converter 2. For example, system characteristics may include one or more parameters of other system components of the electronic circuitry, such as: the thermal resistance of system 1, which is crucial for controlling the temperature of the electronic converter 2; the capacitance of the backup capacitor coupled to the electronic converter 2; and similar parameters. Correspondingly, the variances of operating parameters, operating characteristics, and system characteristics may also be considered, which may arise from manufacturing tolerances or aging.

[0087] The control signal model is determined based on the corresponding input parameters: a modified control signal waveform S', which can be used to control electronic circuit 2; or a corresponding modified control signal time sequence S. t+1 ...S t+T The modified control signal time series is provided in parameterized form.

[0088] In the subsequent step S5, the electronic converter 2 is correspondingly controlled by the modified control signal waveform S'. Preferably, the modified control signal parameters defining the modified control signal waveform S' can be converted into an analog control signal S' in the control signal model block 5 or in a separate device.

[0089] In step S6, an adaptation criterion can be checked. If the adaptation criterion specifies further adaptation or updating of the control signal model (either option: yes), the method continues to step S7; otherwise (either option: no), it jumps back to step S1.

[0090] Adaptation criteria may depend, for example, on a predetermined duration since the last adaptation, the waveform of the original control signal, or an external adaptation signal. Furthermore, if system characteristics, such as temperature, voltage values, measured signals, etc., deviate from corresponding predetermined reference values ​​by a predetermined deviation amount, an update to the control signal model can be triggered. In particular, adaptation criteria should be used to check whether the control signal model needs retraining due to component aging, wear, or other systematic changes in operating conditions.

[0091] In step S7, the control signal model is adapted by retraining, updating, or retraining.

[0092] The adaptation of the control signal model is used to correct inaccuracies in the circuit simulation and the component model on which the circuit simulation is based if, for example, due to component aging, wear, or other systematic changes in operating conditions, the taught control signal (adaptation) is no longer optimal.

[0093] The goal of adapting the control signal model is to form a modified control signal waveform based on a pre-given control signal waveform that optimizes the behavioral metrics. Specifically, the cost function of the behavioral metrics and the system of equations from circuit simulation can be used in conjunction with gradient descent to further train the model parameters of the control signal model.

[0094] The initial creation of the control signal model can be achieved using a training dataset, which consists of the provided control signal waveform S and the predicted control signal waveform S. t+1 ...S t+T This is constituted by the assigned modified control signal S'. Here, the modified control signal waveform is determined based on a pre-given control signal waveform and according to a behavioral metric, which is determined based on one or more behavioral parameters and a pre-given cost function, wherein the one or more behavioral parameters are determined for the modified control signal by means of circuit simulation.

[0095] Subsequently, another scheme for training the control signal model is referenced. Figure 5 The block diagram is used to describe this. Block 11 represents the control signal model to be trained. This control signal model is based on the provided control signal waveform S. t-n ...S t To determine the predicted control signal waveform S t+1 ...S t+T The modified control signal waveform S' is determined based on the current training state. Simulation block 12 is used to simulate one or more electrical behavior parameters VG, which simulate the operation of the technical device 6 based on the modified control signal waveform S'. Circuit simulation is well known and is based on generally differentiable differential equations.

[0096] The cost function is pre-defined in cost function block 13, which combines the behavior parameter VG of each considered time step of the modified control signal waveform S' into the behavior metric VM. Thus, the total value of the behavior metrics, especially the sum of the behavior metrics VM, represents the overall behavior metric VM. ges This overall behavior metric represents the loss during the training of the control signal model.

[0097] This training can be achieved through backpropagation, where the control signal model is differentiated using simulation blocks to adapt its parameters. This requires that the applied cost function also be differentiable.

[0098] In order to construct training data, the control signal waveform can be given in advance in as many variations as possible within the framework of possible control signals used to manipulate the converter device 3.

[0099] Circuit simulation can determine one or more corresponding behavioral parameters VG or the resulting behavioral metrics based on the parameterization of the modified control signal waveform S' (according to a pre-given cost function).

[0100] By employing optimization methods, such as stochastic gradient descent, the training dataset for the control signal model can be used to measure the overall behavior of the VM through aspects related to the provided control signal waveform S. ges The optimization (minimization) is performed, especially through backpropagation. This is possible because the functions upon which the circuit simulation is based and the cost function can be differentiated, allowing the optimized behavioral metric to be determined by differentiating the cost function and the circuit simulation function, which is then used to update the model parameters of the control signal model.

[0101] To ensure that the function induced by the control signal waveform in the electronic system can also be achieved through a modified control signal waveform S', this optimization can be performed using appropriate auxiliary conditions. Alternatively or additionally, in the case of a cost function, to calculate the behavioral metric, a parameter can be considered that evaluates how the function induced by the control signal waveform is achieved through the modified control signal waveform.

[0102] In particular, control signal models can be constructed in the form of neural networks, such as recurrent neural networks (LSTM, GRU), multilayer perceptrons, etc. Therefore, training the control signal model can be achieved using backpropagation based on a VM (monitoring of the overall behavior). ges To execute by minimizing.

[0103] To reduce the burden on the control unit 4 of electronic system 1, the training or adaptation of the control signal model can also be performed outside of electronic system 1. For this purpose, the parameterized original and modified control signal waveforms S, S', and, if necessary, operating parameters, operating characteristics, and system characteristics can be transmitted to an external computing unit. This external computing unit, with knowledge of the electronic circuitry, performs circuit simulation to calculate behavioral parameters or behavioral metrics. Therefore, the retraining of the neural network model parameters can be performed externally, and these model parameters can be transmitted back to electronic system 1 so that they can be applied later.

[0104] In another embodiment, the control signal model can also be implemented as a lookup table in the technical system, such that the control signal parameters of the original control signal waveform are assigned to the modified control signal parameters according to the lookup table, and these modified control signal parameters represent the modified control signal waveform S'. Here, the lookup table is created based on the control signal model, which can be implemented outside of the technical system 1. In this way, the computational cost in the electronic system 1 can be significantly reduced.

Claims

1. A method for operating a technical apparatus (6) having an electronic converter (2) controlled by at least one control signal, the method comprising the following steps: - Provide (S1) the control signal waveform (S) that should be used to manipulate the electronic converter (2); - Predict the predicted control signal waveform (S) based on the provided control signal waveform (S). t+1 ...S t+T ), which is the predicted future waveform of the control signal; - Based on a trainable data-based control signal model, the provided control signal waveform (S) and the predicted control signal waveform (S) are used. t+1 ...S t+T The provided control signal waveform (S) is modified to obtain the modified control signal waveform (S'). The control signal model is trained based on the provided control signal waveform (S) and the predicted control signal waveform (S). t+1 ...S t+T To determine the modified control signal waveform (S'); - The electronic converter (2) is operated according to the modified control signal waveform (S').

2. The method according to claim 1, wherein the control signal model is constructed to determine the modified control signal waveform (S') based on one or more of the following parameters: one or more operating parameters of the electronic converter (2) and / or the technical device (6) to be controlled; one or more operating characteristics of the electronic converter (2) and / or the technical device (6) to be controlled; one or more system characteristics of the electronic converter (2) and / or the technical device (6) to be controlled; and one or more system parameters of the technical device (6), wherein the operating parameters affect the operation of the technical device (6).

3. The method of claim 2, wherein the operating parameters include one or more of the following parameters: the thermal resistance of the overall structure; the capacitance of the standby capacitor coupled to the electronic components; and the variance of these parameters.

4. The method according to any one of claims 1 to 3, wherein the control signal model is constructed as a trainable data-based model.

5. The method according to any one of claims 1 to 3, wherein the control signal model is constructed as an artificial neural network.

6. The method according to claim 5, wherein the artificial neural network is a multilayer perceptron or a recurrent neural network.

7. The method according to any one of claims 1 to 3, wherein the provided control signal waveform (S) and the predicted control signal waveform (S) for manipulating the electronic converter (2) t+1 ...S t+T The modified control signal waveform (S') is parameterized or limited by control signal parameters, and / or the modified control signal waveform (S') is parameterized or limited by corresponding modified control signal parameters.

8. The method of claim 7, wherein parameterization of the provided control signal waveform and / or the modified control signal waveform (S') is performed by means of a time period and the values ​​of electrical parameters assigned to the time period.

9. The method of claim 8, wherein the electrical parameter is voltage or current.

10. The method of claim 7, wherein parameterization of the provided control signal waveform and / or the modified control signal waveform (S') is performed by periodically manipulating one or more parameters.

11. The method of claim 10, wherein the one or more parameters include period frequency, frequency and / or pulse width modulation, modulation rate, duty cycle, pulse duration and / or pulse shape.

12. The method according to any one of claims 1 to 3, wherein the predicted control signal waveform (S) t+1 ...S t+T The prediction is performed using a data-based prediction model, which is trained to determine the predicted control signal waveform (S) based on the provided control signal waveform (S). t+1 ...S t+T ).

13. The method of claim 12, wherein the data-based prediction model includes a recurrent neural network, a state-space model, a sequence2sequence model, or a NARXGP model.

14. The method according to any one of claims 1 to 3, wherein the modification of the provided control signal waveform (S) is performed according to the trainable data-based control signal model based on the provided control signal waveform (S) and based on the predicted control signal waveform (S). t+1 ...S t+T It also performs according to one or more operating parameters of the technical device (6).

15. The method of claim 14, wherein the one or more operating parameters describe the state of the technical device (6) to be controlled based on the predicted control signal waveform.

16. A method for training a control signal model, wherein the control signal model is constructed for predicting a control signal waveform (S) based on a provided control signal waveform (S). t+1 ...S t+T To determine the modified control signal waveform (S'), the method comprises the following steps: - Provide a training dataset, which includes the provided control signal waveform (S) that should be used to manipulate the electronic converter (2) and the predicted control signal waveform (S). t+1 ...S t+T ); - The control signal model is trained based on the loss function, such that the provided control signal waveform (S) and the predicted control signal waveform (S) are... t+1 ...S t+T The loss function is mapped to the modified control signal waveform (S'), wherein the loss function depends on one or more behavioral measures of the technical device (6) when the converter (2) is manipulated using the modified control signal waveform (S').

17. The method of claim 16, wherein the control signal model is configured for use in the method of any one of claims 1 to 15.

18. The method of claim 16, wherein the one or more behavioral metrics (VMs) characterize the technical device (6) depending on one or more behavioral parameters (VGs), the one or more behavioral parameters describing the behavior of the electronic circuitry of the technical device (6) depending on the modified control signal waveform (S').

19. The method of claim 18, wherein the corresponding behavioral metric (VM) is determined according to a pre-given cost function for evaluating behavioral parameters (VG).

20. The method of claim 18, wherein the one or more behavioral parameters (VG) are determined by means of measurement and / or circuit simulation.

21. The method of any one of claims 18 to 20, wherein the cost function maps a plurality of behavioral parameters (VG) to the behavioral metric (VM) and is differentiable, such that the loss function can be applied in conjunction with gradient-based training methods to train the control signal model.

22. The method according to any one of claims 18 to 20, wherein the one or more behavioral parameters (VG) describe the functional capability of the technical device (6) when manipulated by the modified control signal waveform (S').

23. The method of claim 22, wherein the one or more behavioral parameters (VG) describe power loss, interference metrics, thermal load on the device (6) and / or load metrics affecting the converter device (3) or the device (6) for the desired expected service life.

24. The method of claim 23, wherein the disturbance metric is a metric relating to the occurrence of oscillations or overshoot.

25. An apparatus (6) for operating a technical device having an electronic converter (2) controlled by at least one control signal, wherein the apparatus is configured to: - Provide a control signal waveform that should be used to operate the electronic converter (2); - Predict the predicted control signal waveform (S) based on the provided control signal waveform (S). t+1 ...S t+T ), which is the predicted future waveform of the control signal; - Modify the provided control signal waveform (S) according to the provided control signal waveform and the predicted control signal waveform according to the trainable data-based control signal model, so as to obtain the modified control signal waveform (S'). The control signal model is trained based on the provided control signal waveform (S) and the predicted control signal waveform (S). t+1 ...S t+T To determine the modified control signal waveform (S'); - The electronic converter (2) is operated according to the modified control signal waveform (S').

26. An apparatus for training a control signal model, wherein the control signal model is constructed to predict a control signal waveform (S) based on a provided control signal waveform (S) and based on the provided control signal waveform (S). t+1 ...S t+T ) to determine the modified control signal waveform (S'), wherein the device is configured to: - Provide a training dataset, which includes the provided control signal waveform that should be used to manipulate the electronic converter (2) and the predicted control signal waveform (S). t+1 ...S t+T ); - The control signal model is trained based on the loss function, such that the provided control signal waveform (S) and the predicted control signal waveform (S) are... t+1 ...S t+T The loss function is mapped to the modified control signal waveform (S'), wherein the loss function depends on one or more behavioral measures of the technical device (6) when the converter (2) is manipulated using the modified control signal waveform (S').

27. The method of claim 26, wherein the control signal model is configured for use in the method of any one of claims 1 to 15.

28. An electronic system (1) having a technical device (6) and a device according to claim 25, the technical device comprising a converter device (3) of an electronic converter (2).

29. A computer program product comprising instructions that, when implemented by a computer, cause the computer to perform the steps of the method according to any one of claims 1 to 24.

30. A machine-readable storage medium comprising instructions that, when implemented by a computer, cause the computer to perform the steps of the method according to any one of claims 1 to 24.

Images