Method and apparatus for providing a control system

Abstract

Method for providing a control system for regulating and / or controlling a technical system, in particular a valve, wherein control parameters of the control system, which characterize the behavior of the control system, are selected depending on compressed system parameters, which characterize the behavior of the technical system, wherein the compressed system parameters are determined from system parameters, which characterize the behavior of the technical system, by means of a dimensionality reduction system, in particular a machine learning system, wherein the system parameters characterize the behavior of the technical system at an earlier time.

Description

[0001] The invention relates to a method for providing a control system, a computer program, a computer, a control system and a control device. State of the art

[0002] A controller for controlling an actuator is known from DE 10 2013 217 105 A1. Advantages of the invention

[0003] The invention with the features of independent claim 1 has the advantage that the controller can be particularly easily adapted individually to the technical system to be controlled.

[0004] Further aspects of the invention are the subject of the independent claims. Advantageous further developments are the subject of the dependent claims. Disclosure of the invention

[0005] In a first aspect, the invention relates to a method for providing a control system for regulating and / or controlling a technical system, in particular a valve, wherein control parameters of the control system that characterize the behavior of the control system are selected depending on compressed system parameters that characterize the behavior of the technical system, wherein the compressed system parameters are determined by means of a dimensionality reduction system, in particular a machine learning system, from system parameters that characterize the behavior of the technical system, wherein the system parameters characterize the behavior of the technical system, preferably at a particular predeterminable earlier point in time (relative to a point in time at which the system parameters are provided).

[0006] This makes it particularly easy to adapt the control parameters of the control system to the individual properties of each technical system, thus effectively compensating for inevitable fluctuations even between similar components. By reducing the dimension of the system parameters, which, for example, can characterize the behavior of the system in the case of a predefined excitation as a time series and are therefore potentially very high-dimensional, the problem of selecting suitable system parameters (which can, for example, be output from a machine learning system depending on the

[0007] In a further development, it can be provided that the system parameters are determined by a measurement on the technical system, in particular during production of the technical system, preferably at the end of this production line. The system parameters thus characterize the behavior of the technical system at the time of production. This enables standardized measurements of the properties of the technical system to be carried out in a particularly simple manner. In some embodiments, the technical system can be understood as the set of its components, so that in the case of distributed production of components of the technical system, the overall view of the production of all components can be understood as the production of the technical system.

[0008] In a further development, it can be provided that the measurement is carried out in such a way (ie that measurement constraints are selected in such a way) that measurements on different examples of similar technical systems have the highest possible scatter.

[0009] A measurement on a technical system can be understood as the entire process of interaction with the system to determine a system response, up to and including the determination of the measurement result based on the determined system response. An appropriate selection of measurement conditions can therefore either mean that the conditions under which the interaction with the technical system is carried out to determine the system response are selected accordingly. An appropriate selection of measurement conditions can also mean that the respective interactions with the technical system are carried out under a large number of conditions, and the measurement results are determined only for a portion of the interactions carried out.

[0010] In particular, it can be provided that the measurement is carried out in such a way (ie that measurement conditions are selected in such a way) that measurements on different examples of similar technical systems have a scatter that is greater than a predeterminable threshold scatter.

[0011] Alternatively or additionally, it can be provided that in a test measurement preceding the measurement on the technical system, identical measurements (i.e. identical for a plurality of specimens) are first carried out on a plurality of specimens of similar technical systems and a scatter of the measured values ​​between the measurements on different specimens is determined, and wherein the measurement on the technical system whose system parameters are determined is dependent on the scatter determined, in particular that the measurement conditions under which this measurement is carried out are dependent on the scatter determined.

[0012] In a further development of this, it can be provided that, given a plurality of measurement conditions, the plurality of measurements are carried out and the respective scatter is determined, and wherein the measurement on the technical system whose system parameters are determined is dependent on the scatter determined, in particular that the measurement conditions under which this measurement is carried out are dependent on the scatter determined.

[0013] In some embodiments, it may be provided that the dimension reduction system is an autoencoder and the compressed system parameters are determined at a bottleneck of the autoencoder when the system parameters are supplied to the autoencoder.

[0014] It can be provided that the compressed system parameters determined at the bottleneck of the autoencoder are a compressed representation of the supplied system parameters determined by the autoencoder.

[0015] In other embodiments, it can be provided that the autoencoder is a variational autoencoder, and the determined compressed system parameters then represent mean values ​​or standard deviations of a probability density distribution of the system parameters processed by an encoder of the variational autoencoder.

[0016] In still further embodiments, other dimension reduction systems such as a normalizing flow or a principal component analysis are conceivable.

[0017] In further developments of the previous embodiments, it can be provided that the control parameters of the control system are selected by means of a second machine learning system, for example, a neural network, depending on the compressed system parameters. By compressing the system parameters into a smaller latent space, it is particularly easy to train the machine learning system, since not all control parameters of the training data set need to be known in advance; rather, some of the control parameters can be determined by interpolation or extrapolation from known control parameters.

[0018] Embodiments of the invention are explained in more detail below with reference to the accompanying drawings. In the drawings: Fig. 1 schematically shows an exemplary structure of an autoencoder; Fig. 2 shows an example of a result of measurements; Fig. 3 shows an example of a motor vehicle with a regulator; Fig. 4 schematically shows a production line; Fig. 5 shows a flowchart of embodiments for training the autoencoder; Fig. 6 shows a flowchart of embodiments for compressing the measurement data; Fig. 7 shows a flowchart of embodiments for training the machine learning system; Fig. 8 shows a flowchart showing a procedure for parameterizing the controller.

[0019] Fig. Figure 1 schematically shows an exemplary structure of an autoencoder for encoding and decoding a time series of data (E). The autoencoder comprises an encoder (1001) and a decoder (2001). Encoder (1001) and decoder (2001) are advantageously each configured as a neural network.

[0020] The encoder (1001) comprises an input (1100) and an output (1500). The decoder (2001) comprises an input (2100) and an output (2500). The output (1500) of the encoder (1001) is coupled to the input (2100) of the decoder (2001), so that data (L) (hereinafter "latent codes") provided at the output (1500) of the encoder (1000) are present at the input (2100) of the decoder (2000).

[0021] The encoder (1001) and decoder (2001) are configured such that data (E) present at the input (1100) of the encoder (1001), hereinafter referred to as "input data," are converted by the encoder (1001) into latent codes (L) and by the decoder (2001) into data (A) (hereinafter referred to as "output data"), which are provided at its output (2500). The encoder (1001) and decoder (2001), coupled together in this manner, form an autoencoder (5001) in a known manner.

[0022] Input data (E) and output data (A) have the same dimensionality, whereas the latent codes (L) have a lower dimensionality than these data. The layer (1700) of the autoencoder (5000) at whose output the latent codes (L) are provided is therefore also referred to as the "bottleneck."

[0023] When training the encoder (1001) and decoder (2001), the parameters of the encoder (1001) and decoder (2001) are adjusted in such a way that the input data (E) and output data (A) match as closely as possible, i.e., the data present at the input (1100) of the autoencoder (5001) are reconstructed as well as possible at its output (2500) (whereby, as usual, regularization terms are of course possible for the training target).

[0024] Fig. Figure 1 illustrates, by way of example, the result of a measurement performed during the production of a valve. The example shows the attraction force (F) of a valve's magnet armature versus a valve's attraction path (w) (although the abscissa represents the attraction path (w) and not time, such sequences of measurement points ordered along a dimension are referred to as "time series"). A multitude of such time series for a multitude of valves are shown.

[0025] Up to a first tightening path (w1), the scatter (ΔF) between the time series is quite small for each tightening path and then increases rapidly. Between a second tightening path (w2) and a third tightening path (w3), this scatter (ΔF) is large and then decreases again.

[0026] If an autoencoder (5001) is used as in Fig. 1, trained with the complete time series, the error in the reconstruction of the output variable (A) from the input variable (E) in this range, with the dimension of the latent codes (L) remaining constant, does not differ significantly from an autoencoder (5001) that was only trained on this range of high variance, for example, the range between the second pull path (w2) and the third pull path (w3). In other words, the autoencoder (5001) is trained during training to focus on those areas subject to high fluctuations.

[0027] Fig. Figure 3 shows an example of a motor vehicle (100) with a braking system (130) comprising a valve (140) for controlling a flow rate of brake fluid, thereby controlling the strength of a brake actuation of the braking system (130). The motor vehicle (100) has a control unit (120) that controls the braking system (130) with a control algorithm based on signals from a sensor (110). For this purpose, a controller (125) is provided on the control unit (e.g., implemented in software), which executes this control algorithm.

[0028] Fig. Figure 4 shows an example of a production line (300) on which the valve (140) is manufactured. By means of a measuring device (310) at the end of the line, the Fig. 2 are recorded and fed to a computer (200). The computer is configured to determine parameters of the controller (125) based on the determined time series so that the controller can optimally control the braking system (130) with the valve (140). For this purpose, the computer (200) preferably comprises a machine learning system, in some embodiments a neural network.

[0029] The technical properties, which may be subject to fluctuations during production processes on the production line (300) and during production-related tests of the measuring device (310), are also simultaneously properties that affect the behavior of the controller (125). For example, for pressure control in the brake system (130), the flow rate of the valve (140) under different actuating currents is crucial.

[0030] The deviations (variances) of these physical properties are encoded in latent codes (L) in a low-dimensional form, as described below. The controller (125), which is adjusted accordingly, is thus able to optimize the control behavior for each individual component.

[0031] Fig. 5 illustrates in a flowchart embodiments for training the Fig. 1. First, a plurality of components manufactured on the production line (300)—in some embodiments, the valves (140)—are provided (1000) and measured (1010) by means of the measuring device (310). Subsequently, a fluctuation range (Δw) of this plurality of measurements is determined (1020), and a starting point (w2) and an end point (w3) of a range are identified (1030) in which this variability is sufficiently large, in some embodiments greater than a predeterminable threshold value. Subsequently, the time series identified are fed to the autoencoder (5000) as input data (E), and the autoencoder (5001) is trained (1040) to reconstruct this input data (E) as best as possible as corresponding output data (A). This concludes the method.

[0032] In some embodiments, it may be provided that steps (1020) and (1030) are not performed and the entire time series is used to train the autoencoder (5001).

[0033] Fig. 6 illustrates, in a flowchart, embodiments for compressing the measurement data of the individual component manufactured on the production line (300)—in some embodiments, the valve (140). First, the individual component, manufactured or still in production, is provided (2000). Using the measuring device (310), a property of the component is measured (2010), and the resulting time series is restricted (2020) to the range to which the time series with which the autoencoder (5001) was trained were restricted—for example, to the range between w2 and w3. In some alternative embodiments, this range can initially be specified, and the measurement can be restricted exclusively to this range. The time series thus restricted is fed as input data (E) to the encoder (1001) of the autoencoder (5001), and latent codes (L) are determined (2030) using the encoder (1001).These latent codes (L) are the compressed measurement data.

[0034] Fig. 7 illustrates, in a flowchart, embodiments for training the machine learning system of the computer (200) for determining the parameters of the controller (125). First, for a first plurality of components manufactured on the production line (300) - in some embodiments, the valves (140) - the latent code (L) determined by means of the measuring device (310) is determined - for example, with the Fig. 6 illustrated method (3000). Subsequently, the components are put into operation (3010) in their respective target state in interaction with the controller (125), in some embodiments by installing the valves (150) in the braking system (130) of the respective motor vehicle (100). For the motor vehicles (100) equipped in this way, the controllers (125) are parameterized (3020), for example in the so-called application process. Since this only needs to be done for a small number of motor vehicles (100), the effort is manageable. Thus, there are pairs of latent codes (L) and associated parameters of the controller (125).

[0035] Subsequently, for a second plurality of further components manufactured on the production line (300) - in some embodiments of the valves (140) - the latent code (L) determined by means of the measuring device (310) is determined - for example with the Fig. 6 illustrated method (3030). Subsequently, the parameters of the controller (125) are determined (3040) by means of interpolation and, if necessary, extrapolation, in particular linear interpolation or linear extrapolation, from the pairs of latent codes (L) and parameters of the controller (125) determined for the first plurality of components. These parameters can, for example, include a pre-control value (which, if necessary, is also selected depending on further variables that characterize an operating state of the controlled technical system, for example the braking system (130)) or, for example, gain factors of a PID controller. The first plurality and the second plurality thus result in pairs of latent codes (L) and associated parameters of the controller (125). This entirety is provided as a training data set (3050). The machine learning system is then trained with this training data set in order to determine the parameters of the controller (125) from the latent codes (3060).This concludes the procedure. This exploits the fact that the latent space of latent codes (L) adequately reflects the mathematical structure of the manifold of time series in a high-dimensional space, thus allowing meaningful interpolation between points in this latent space.

[0036] Fig. Figure 8 illustrates a flowchart of a method for parameterizing the controller (125) in the motor vehicle (100). For the component, in some embodiments the valve (140), the time series determined for the component with the measuring device (310) are, for example, Fig. 6 and the associated latent codes (L) are provided (4000). Then, using the method shown in Fig.7, the computer's (200) trained machine learning system determines (4010) the parameters of the controller (125) from the latent codes (L) and provides them (4030) to the controller (125). During operation of the motor vehicle (100), the controller (125) controls the braking system (130) using these parameters (4040).

[0037] It is understood that this method is not limited to the regulation of braking systems of motor vehicles, but can be transferred to any controlled technical system whose behavior depends on variables that can be determined in the technical system (or a component of the technical system). QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This 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

[0000] DE 10 2013 217 105 A1

[0002]

Claims

[1] Method for providing a control system (125) for regulating and / or controlling a technical system (130), in particular a braking system comprising at least one valve (140), wherein control parameters of the control system (125) which characterize the behavior of the control system (125) are selected as a function of compressed system parameters (L) which characterize the behavior of the technical system (130), wherein the compressed system parameters (L) are determined by means of a dimensionality reduction system, in particular a machine learning system, from system parameters (E) which characterize the behavior of the technical system (130), wherein the system parameters (E) characterize the behavior of the technical system (130), in particular at an earlier point in time. [2] Method according to claim 1, wherein the system parameters are determined as a function of a measurement on the technical system (130), in particular during manufacture of the technical system (130). [3] Method according to claim 2, wherein the measurement is carried out in such a way that measurements on different examples of similar technical systems (130) have a sufficiently high scatter (ΔF). [4] Method according to claim 3, wherein the measurement is carried out in such a way that measurements on different examples of similar technical systems (125) have a scatter (ΔF) which is greater than a predeterminable threshold scatter. [5] Method according to claim 3 or 4, wherein in a test measurement preceding the measurement on the technical system (125), identical measurements are first carried out on a plurality of specimens of similar technical systems and a scatter (ΔF) of the measured values ​​between the measurements on different specimens is determined, and wherein the measurement on the technical system (125) is dependent on the determined scatter (ΔF). [6] Method according to one of the preceding claims, wherein the dimension reduction system is an autoencoder (5001) and the compressed system parameters (L) are determined at a bottleneck of the autoencoder (5001) when the system parameters (E) are supplied to the autoencoder (5001). [7] Method according to one of the preceding claims, wherein the control parameters of the control system (125) are selected by means of a second machine learning system depending on the compressed system parameters (L). [8] A computer program arranged to cause your computer to carry out the method according to any one of the preceding claims when executed on the computer (200). [9] Computer (200) on which the computer program according to claim 7 is stored and which is arranged to carry out the method according to one of claims 1 to 6. [10] Control system (125) determined by the method according to one of claims 1 to 6. [11] Control device (120) comprising the control system (125) according to claim 9.

Images