Method and device for improved evaluation of measurement signals from a sensor

Abstract

Method for training a machine learning system (200) for evaluating a measurement signal (M) of a sensor (2) which is configured to determine at least one variable (A) characterizing an operating state of a technical system (11), wherein the machine learning system (200) comprises a first subsystem (201) which is configured to determine a latent representation (L) from the measurement signal (M) supplied to it, and wherein the machine learning system comprises a second subsystem (202) which is configured to determine the variable (A) characterizing the operating state of the technical system (11) from this latent representation (L), and wherein the first subsystem (201) is trained in an unsupervised or self-supervised manner, and wherein the machine learning system (200) is subsequently trained in a supervised manner.

Description

The invention relates to a method for evaluating a measurement signal of a sensor, a computer program, a machine-readable storage medium and a control device.Prior ArtAccording to the prior art, a virtual sensor can be used to determine a target variable which depends on one or more correlating measured variables. In contrast to conventional physical sensors, the target variable is not measured directly, but rather is mapped by software on the basis of correlations with other measured variables. Mathematical models, simulations or artificial intelligence can be used here.For example, DE 10 2023 206 032.9, which is not prepublished, discloses a method for evaluating a measurement signal of a sensor which is configured to determine at least one variable characterizing an operating state of a technical system, wherein a low-frequency component of the measurement signal is determined for this purpose and / or a high-frequency component of the measurement signal is determined, and wherein the evaluation of the low-frequency component and the evaluation of the high-frequency component are carried out using different evaluation methods for determining the variable characterizing the operating state of the technical system.Advantages of the InventionThe invention with the features of independent claim 1 has the advantage over the related art that it can achieve very good results with very few labeled data.Further aspects of the invention are the subject of the subordinate claims. Advantageous further developments are the subject of the dependent claims.Disclosure of the InventionIn a first aspect, the invention therefore relates to a method for training a machine learning system for evaluating a measurement signal of a sensor, which is configured to ascertain at least one variable characterizing an operating state of a technical system, wherein the machine learning system comprises a first subsystem configured to ascertain a latent representation of the measurement signal supplied to it, and wherein the machine learning system comprises a second subsystem configured to ascertain the variable characterizing the operating state of the technical system from this latent representation, and wherein the first subsystem is trained unsupervised or self-supervised, and wherein the machine learning system is subsequently trained supervised.A "latent representation" refers in the usual way to a (typically vector-valued) compressed representation of the (likewise typically vector-valued) measurement signal.This has the advantage that only a few labeled data are necessary for training. This means that for a given amount of training data, the evaluation of the measurement signal works particularly well.The machine learning system may be, for example, a neural network.It can be provided that during the monitored training only the second subsystem is changed, that is to say that the first subsystem remains unchanged.This has the advantage that particularly few labeled data are necessary for training.Here, "changing a subsystem" can mean in the usual way that parameters characterizing the behavior of this subsystem are changed.In further developments, it can be provided that during unsupervised or self-supervised training of the first subsystem, parts of the measurement signal are used weighted by weighting factors, wherein the weighting factors have been identified as being particularly relevant for the correct determination of the variable characterizing the operating state of the technical system by means of a model which is configured to determine the variable characterizing the operating state of the technical system from the measurement signal.In particular, it is possible in some embodiments that the weighting factors for sampling points in the time series are determined from the model by means of a so-called "SAP" method, and in unsupervised or self-supervised learning a cost function characterizing this learning has a term in which contributions associated with the respective sampling points are weighted with these weighting factors.As a result, unsupervised or self-supervised training becomes particularly efficient.If the measurement signal is given as a time series, the parts of the measurement signal can be sections, in particular a coherent section, of the measurement signal in some embodiments. In particular, it can be provided that for this purpose a start time and an end time of the time series are predetermined in order to unambiguously characterize the part of the measurement signal.The model can be a machine learning system, in particular a neural network, which has been trained by means of parallel pairs of variable characterizing the measurement signal and the operating state of the technical system. Since the model is used merely for identifying the parts of the measurement signal that are most relevant for the training of the first subsystem, fewer pairs are required here than for a completely correct reconstruction of the variable that characterizes the operating state of the technical system.Embodiments of the invention are explained in more detail below with reference to the attached drawings. In the drawings, there are shown: FIG. 1 schematically shows a sensor in a brake system; FIG. 2 schematically shows a structure of a training device of the machine learning system; FIG. 3 schematically illustrates a flow of information through the machine learning system; FIG. 4 is a schematic flow chart of a process according to an embodiment of the invention.FIG. 1 schematically shows a brake system (1) of a motor vehicle (11) having, in the exemplary embodiment, a plurality of sensors (2) comprising a pressure sensor in a hydraulic feed line (3), with which a contact pressure force of a brake caliper (4) is controlled, a sensor for determining a longitudinal acceleration of the motor vehicle (11), a sensor for detecting a temperature of the liquid in the hydraulic feed line (3) and a voltage sensor on a pump (not shown) for conveying the liquid. A multidimensional measurement signal (M) comprising the measurement signals of these sensors is supplied to a control unit (10), which comprises a computer-readable storage medium (20), on which a computer program is stored, on which a program for evaluating the measurement signal (M) is stored in order to determine a torque that the brake caliper (4) transmits to a wheel (not shown) of the motor vehicle (11). The control device (10) further comprises a processor (21) which is configured to execute computer programs stored on the storage medium (20).FIG. 2 schematically branches a training device ( 100) for training a machine learning system ( 200), which comprises a first subsystem ( 201) and a second subsystem ( 202). In the exemplary embodiment, the machine learning system is a neural network. The measuring device 100) comprises a data memory (105) in which a plurality of measurement signals (M) are provided. The training device is configured to select measurement signals (M) and supply them to the machine learning system (200).The measurement signal (M) is supplied there to the first subsystem ( 201), which is configured to determine a latent representation (L), i.e. a vector whose dimensionality is lower than the dimensionality of the measurement signal (M), therefrom.The machine learning system ( 200) is furthermore configured to supply the determined latent representation (L) to the second subsystem ( 202), which uses this to determine an output variable (A) that characterizes an operating state of the technical system ( 11), in the exemplary embodiment an estimated torque transmitted by the brake caliper ( 4). In preferred embodiments, the variable characterizing the operating state of the technical system ( 11) is given by a time series.The training device ( 100) is configured to transmit the latent variable (L) determined by the machine learning system ( 200) and the output variable (A) determined by the machine learning system ( 200) to an evaluation unit ( 110).Furthermore, the training device ( 100) is configured to train the first subsystem ( 201). In some embodiments, the training device ( 100) is configured to provide an autoencoder (German: autoencoder) for this purposeFIG. 3 schematically shows a flow of information during the operation of the machine learning system ( 200). A provision unit ( 300) provides measured variables ( M1,...,M 4) as time series of variables determined from sensor data, in the exemplary embodiment a torque determined from a current, a temperature, a voltage and a position of the brake caliper ( 4). At a predeterminable point in time (t0), these measured variables therefore correspond to a four-dimensional vector. These time series are supplied to the first subsystem ( 201), which determines the latent representation (L) therefrom, in the illustrated example a three-dimensional vector at the respective time (t 0). The time series profile of the measured variables (M1,...,M4) correspondingly produces a profile in three-dimensional space in the latent representation (L). This latent representation (L) is supplied to the second subsystem ( 202), which determines the output variable (A) from this. In the exemplary embodiment, the output variable (A) is one-dimensional, i.e., it corresponds to a scalar value at the time (t 0).Measured variables (M1,...,M4) and latent representation (L) are also fed in some embodiments to a further machine learning system ( 203), which is trained to estimate an uncertainty of the output variable (ΔA) from these variables. This training of this further machine learning system ( 203) can be carried out, for example, in the training method of the machine learning system ( 200) illustrated in FIG. 4 by the extent to which the machine learning system ( 200) is able to actually reconstruct the output variable (A) from the respective measurement signals (M) in such a way that it actually corresponds to the provided measured or simulatively provided value of the output variable (A) at the respective time (t 0). This degree of uncertainty can be used as a target specification in the monitored training of the further machine learning system ( 203).FIG. 4 is a flow chart showing an exemplary flow of the computer program. First (1000), a plurality of measurement signals (M) are received by the pressure sensor (2) and provided as time-series data.Subsequently, variables (A) parallel to these respective time series characterizing the operating state of the technical system ( 11), in the exemplary embodiment of the brake system ( 1), are provided ( 1100). These variables (A) characterizing the operating state of the technical system can be determined simulatively, for example, or preferably measured. In the exemplary embodiment, these variables (A) characterizing the operating state are again given as time series.A (recurrent in the exemplary embodiment) neural network is now trained ( 1200) using the pairs of measurement signals (M) and parallel variables (A) characterizing the operating state of the technical system ( 1) to reconstruct the respective parallel variables (A) characterizing the operating state of the technical system ( 1) from the measurement signals (M) which are provided to the recurrent neural network as input variables as output variables at an output of the recurrent neural network.In this case, the recurrent NN is likewise trained to identify, for measurement data, the most important parts of the time series for the ascertainment of the operating data. In this case, in the exemplary embodiment using the "SAP" method, a weighting factor is assigned to each sampling time of the time series. This method is known, for example, from A unified approach to interpreting model predictions. S. Lundberg, SI Lee, arXiv preprint arXiv:1705.07874, 2017.Now a second set of measurement signals (M) is provided (1300), which in some embodiments is greater than the set of measurement signals provided in step (1000).Using this second set of measurement signals (M), the first subsystem ( 201) is trained ( 1400), in some embodiments unsupervised, in some embodiments self-supervised, for example using an autoencoder or by means of contrasting learning. For this purpose, the weighting factors identified with respect to the provided measurement data (M) by means of the recurrent NN are provided and used as weighting of the respective measurement data (M) in order to train the first subsystem ( 201). In some exemplary embodiments, this training is carried out as a function of a quadratic norm of a difference between a vector representing the measurement data (M) and a vector representing the measurement data reconstructed by the autoencoder, wherein each dimension of the vectors corresponds to a sampling time and these quadratic distances of the individual sampling times are weighted with the respective identified weighting factors in order to yield a cost function for the training.A third set of measurement signals (M) is now provided ( 1500), and variables (A) characterizing the operating state of the technical system ( 11) parallel to these measurement signals (M) are provided ( 1600), analogous to step ( 1100). These variables (A) characterizing the operating state of the technical system can be determined simulatively, for example, or preferably measured. In the exemplary embodiment, these variables (A) characterizing the operating state are again given as time series.Preferably, from these provided measurement signals (M), the most relevant parts are identified by means of the trained recurrent neural network and only these most relevant parts are provided for the subsequent training (1700).The machine learning system ( 200) is now trained ( 1800) in a monitored manner with the measurement signals (M) thus provided and the parallel variables (A) characterizing the operating state of the technical system ( 11), to reconstruct the variables (A) characterizing the operating state of the technical system ( 11) from the respective measurement signals (M). In the exemplary embodiment, this is done by adapting only the parameters of the second subsystem ( 202)The method ends with this.References included in the specificationThis list of documents cited by the applicant has been produced in an automated manner and is only included for the better information of the reader. The list is not part of the German patent application or utility model application. The DPMA does not take any adhesion for any faults or omissions.Patent Literature citedDE 10 2023 206 032.9

[0003] Cited Non-Patent LiteratureA unified approach to interpreting model predictions. S Lundberg, SI Lee, arXiv preprint arXiv:1705.07874, 2017

[0030]

Claims

Method for training a machine learning system (200) for evaluating a measurement signal (M) of a sensor (2) which is configured to determine at least one variable (A) characterizing an operating state of a technical system (11), wherein the machine learning system (200) comprises a first subsystem (201) which is configured to determine a latent representation (L) from the measurement signal (M) supplied to it, and wherein the machine learning system comprises a second subsystem (202) which is configured to determine the variable (A) characterizing the operating state of the technical system (11) from this latent representation (L), and wherein the first subsystem (201) is trained unsupervised or self-supervised and wherein the machine learning system (200) is subsequently trained supervised.The method of claim 1, wherein only the second subsystem (202) is modified during the monitored training.Method according to one of the preceding claims, wherein, when the first subsystem (201) is trained unsupervised or self-supervised, parts of the measurement signal (M) are used weighted by means of weighting factors, wherein the weighting factors have been identified as being particularly relevant for the correct determination of the variable characterizing the operating state of the technical system (11) by means of a model which is configured to determine the variable characterizing the operating state of the technical system (11) from the measurement signal (M).Method according to Claim 3, wherein the model is a neural network, in particular a recurrent neural network, which has been trained by means of parallel pairs of variable (A) characterizing in each case the measurement signal (M) and the operating state of the technical system (11).Measurement signal evaluator comprising the machine learning system trained according to Claim 1, and which is configured to supply a measurement signal (M) supplied to it to the machine learning system and to determine the variable characterizing the operating state of the technical system by means of the machine learning system and to provide it at the output of the measurement signal evaluator.A training system configured to perform the method of any one of claims 1 to 4.Method for determining the variable characterizing the operating state of the technical system as a function of the measurement signal, wherein the measurement signal is fed to the measurement signal evaluator and the variable characterizing the operating state of the technical system is provided by means of the measurement signal evaluator.A computer program adapted to cause a computer to execute the method of any one of claims 1 to 5 or 7 when executed on the computer (20).A machine readable storage medium (21) on which the computer program of claim 8 is stored.A computer (10) configured to perform the method of any one of claims 1 to 5 or 7.

Images