Method and device for better evaluating measurement signal of sensor

Abstract

To provide a method for training a machine learning system (200) for evaluating a measurement signal (M) of a sensor (2) configured to specify at least one quantity (A) featuring an operation state of a technical system (11).SOLUTION: A machine learning system (200) includes a first partial system (201), which is configured to specify a latent representation (L) from a measurement signal (M) supplied to the first partial system (201). The machine learning system includes a second partial system (202), which is configured to specify a quantity (A) featuring an operation state of a technical system (11) from the latent representation (L). The first partial system (201) is subjected to an unsupervised training or self-supervised training, and then the machine learning system (200) is subjected to a supervised training.SELECTED DRAWING: Figure 4

Description

Technical Field

[0001] The present 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.

Background Art

[0002] Prior Art According to the prior art, virtual sensors can be used to determine a target quantity that depends on one or more correlated measured quantities. Different from conventional physical sensors, the target quantity is not directly measured, but is mapped by software based on its correlation with other measured quantities. In this case, mathematical models, simulations, or artificial intelligence can be used.

[0003] Thus, for example, from unpublished German Patent Application No. 102023206032.9, a method for evaluating the measurement signal of a sensor configured to identify at least one quantity characterizing the operating state of a technical system, for which purpose the low-frequency component of the measurement signal is identified and / or the high-frequency component of the measurement signal is identified, and the evaluation of the low-frequency component and the evaluation of the high-frequency component are each carried out using different evaluation methods to identify the quantity characterizing the operating state of the technical system, is recognized.

Summary of the Invention

Problems to be Solved by the Invention

[0004] Advantages of the Invention In contrast, the present invention having the features described in independent claim 1 has the advantage that very good results can already be achieved using only very little labeled data.

[0005] Further aspects of the present invention are the subject of further independent claims. Advantageous developments are the subject of the dependent claims.

Means for Solving the Problems

[0006] Disclosure of the Invention Thus, in a first aspect, the present invention is a method for training a machine learning system for evaluating a measurement signal of a sensor configured to identify at least one quantity characterizing an operating state of a technical system, wherein the machine learning system includes a first sub-system configured to identify a latent representation from the measurement signals supplied to the first sub-system, the machine learning system includes a second sub-system configured to identify a quantity characterizing the operating state of the technical system from the latent representation, the first sub-system is trained without a teacher or with self-supervision, and subsequently, the machine learning system is trained with a teacher.

[0007] "Latent representation" means a compressed representation of the measurement signal (which is typically also vector-valued), as is usual.

[0008] This has the advantage that only little labeled data is required for training. This means that, given a certain range of training data, the evaluation of the measurement signals functions particularly well.

[0009] The machine learning system may be, for example, a neural network.

[0010] It can be assumed that, during supervised training, only the second sub-system is changed, i.e., the first sub-system remains unchanged. This has the advantage that particularly little labeled data is required for training.

[0011] In this context, "the sub-system is changed" may, as is usual, mean that the parameters characterizing the behavior of this sub-system are changed.

[0012] In a development form, during unsupervised training or self-supervised training of the first partial system, a part of the measurement signal is weighted and used using weighting factors, and the weighting factors are identified as being particularly relevant for accurately identifying a quantity characterizing the operating state of the technical system using a model configured to identify a quantity characterizing the operating state of the technical system from the measurement signal. It can be assumed that this is the case.

[0013] In particular, in some embodiments, it is possible to identify the weighting factors for the sampling times of the time series from the model using the so-called "SHAP" method, and during unsupervised learning or self-supervised learning, the cost function characterizing this learning may have terms such as the contribution associated with each sampling time being weighted by those weighting factors.

[0014] Thereby, unsupervised training or self-supervised training becomes particularly efficient.

[0015] If the measurement signal is given as a time series, in some embodiments, a part of the measurement signal may be, for example, an interval of the measurement signal, particularly a continuous interval. In particular, for this purpose, it can be assumed that the start time and end time of the time series are set in advance in order to uniquely characterize a part of the measurement signal.

[0016] The model may be a machine learning system, particularly a neural network, trained using parallel pairs of the measurement signal and the quantity characterizing the operating state of the technical system, respectively. In this case, since the model is only used to identify the most relevant part of the measurement signal for training the first partial system, the number of pairs required is less than the number required to completely and accurately reconstruct the quantity characterizing the operating state of the technical system.

[0017] Hereinafter, embodiments of the present invention will be described in more detail with reference to the accompanying drawings.

Brief Description of the Drawings

[0018]

Figure 1

Figure 2

Figure 3

Figure 4

Modes for Carrying Out the Invention

[0019] FIG. 1 schematically shows a braking system (1) of an automobile (11). In this embodiment, the braking system (1) includes a plurality of sensors (2). These plurality of sensors (2) include a pressure sensor in a hydraulic supply pipeline (3) for controlling the pressing force of a brake caliper (4), a sensor for specifying the longitudinal and lateral accelerations of the automobile (11), a sensor for detecting the temperature of the liquid in the hydraulic supply pipeline (3), and a voltage sensor provided in a pump (not shown) for pumping the liquid. A multi-dimensional measurement signal (M) including the measurement signals of these plurality of sensors is supplied to a control device (10). The control device (10) includes a computer-readable storage medium (20), and a computer program is stored in the computer-readable storage medium (20). A program for evaluating the measurement signal (M) is stored to specify the torque transmitted to a wheel (not shown) of the automobile (11) by the brake caliper (4). The control device (10) further includes a processor (21) configured to execute the computer program stored in the storage medium (20).

[0020] FIG. 2 schematically shows a training device (100) for training a machine learning system (200), which machine learning system (200) includes a first sub-system (201) and a second sub-system (202). In the present embodiment, the machine learning system is a neural network. The training device (100) includes a data memory (105), and a plurality of measurement signals (M) are provided to this data memory (105). The training device is configured to select the measurement signal (M) and supply it to the machine learning system (200).

[0021] The measurement signal (M) is then supplied to the first sub-system (201), which first sub-system (201) is configured to identify a latent representation (L) from this measurement signal (M), that is, a vector having a dimension lower than that of the measurement signal (M).

[0022] The machine learning system (200) is further configured to supply the identified latent representation (L) to a second sub-system (202), which second sub-system (202) identifies an output quantity (A) characterizing the operating state of the technical system (11) from this identified latent representation (L), and in the present embodiment, identifies an estimated torque transmitted from the brake caliper (4). In a preferred embodiment, the quantity characterizing the operating state of the technical system (11) is given in a time series.

[0023] The training device (100) is configured to transmit the latent quantity (L) identified by the machine learning system (200) and the output quantity (A) identified by the machine learning system (200) to an evaluation unit (110).

[0024] The training device (100) is further configured to train the first sub-system (201). In some embodiments, the training device (100) is configured to provide an autoencoder for this purpose.

[0025] Figure 3 schematically shows the flow of information when operating the machine learning system (200). The providing unit (300) provides measured quantities (M1, ···, M4) as a time series identified from sensor data, that is, in this embodiment, torque identified from current, temperature, voltage, and the position of the brake caliper (4). That is, at a preset time point (t0), these measured quantities correspond to a 4-dimensional vector. These time series are supplied to the first sub-system (201), and this first sub-system (201) identifies a latent representation (L) from these time series, that is, in the illustrated example, a 3-dimensional vector is identified at each time point (t0). The time series transition of the measured quantities (M1, ···, M4) correspondingly brings about a transition in the 3-dimensional space in the latent representation (L). This latent representation (L) is supplied to the second sub-system (202), and this second sub-system (202) identifies the output force amount (A) from this latent representation (L). In this embodiment, the output force amount (A) is 1-dimensional, that is, at the time point (t0), it corresponds to a scalar value.

[0026] The measured quantities (M1, ···, M4) and the latent representation (L) are also supplied to a further machine learning system (203) in some embodiments, and this further machine learning system (203) is trained to estimate the uncertainty (ΔA) of the output force amount from these quantities. The training of this further machine learning system (203) may be implemented as follows, for example, in the training method of the machine learning system (200) shown in FIG. 4. That is, the training of the further machine learning system (203) may be implemented such that the output force amount (A) at each time point (t0) can be actually reconstructed from each measurement signal (M) by the machine learning system (200) so that the output force amount (A) actually corresponds to the provided measured value or the value provided by simulation. This measure of uncertainty can be used as a target setting in the supervised training of the further machine learning system (203).

[0027] FIG. 4 shows a flowchart of an exemplary sequence of a computer program. First (1000), a plurality of measurement signals (M) from the pressure sensor (2) are received and provided as time series data.

[0028] Subsequently, an amount (A) characterizing the operating state of the technical system (11), in this embodiment the operating state of the braking system (1), is provided (1100) in parallel for each of these time series. The amount (A) characterizing the operating state of these technical systems may be, for example, determined by simulation or, preferably, measured. In this embodiment, the amount (A) characterizing these operating states is also given as a time series.

[0029] Here, the (recurrent in this embodiment) neural network is trained (1200) to reconstruct, as output amounts on the output side of the recurrent neural network, the respective parallel amounts (A) characterizing the operating state of the technical system (11) from the measurement signals (M) provided as input amounts to the recurrent neural network using pairs of the measurement signals (M) and the parallel amounts (A) characterizing the operating state of the technical system (11).

[0030] In this case, the recurrent NN is also trained to identify, for the measurement data, the most important parts of the time series for identifying the operating data. In this case, in this embodiment, the "SHAP" method is used to assign weighting coefficients to each sampling time point 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".

[0031] Here, a second set of measurement signals (M) is provided (1300), and in some embodiments, this second set of measurement signals (M) is larger than the set of measurement signals provided in step (1000).

[0032] Using this second set of these measurement signals (M), the first sub-system (201) is trained without a teacher in some embodiments, or with a self-teacher in some embodiments, for example, by an autoencoder or contrastive learning (1400). For this purpose, the weighted coefficients identified using a recurrent NN with respect to the provided measurement data (M) are provided and used as the weights for each measurement data (M) to train the first sub-system (201). In some examples, this training is performed depending on the squared norm of the difference between the vector representing the measurement data (M) and the vector representing the measurement data reconstructed by the autoencoder, where each dimension of the vector corresponds to a sampling time point, and this squared distance at each individual sampling time point is weighted by the respective identified weighted coefficients to obtain a cost function for training.

[0033] Here, a third set of measurement signals (M) is provided (1500), and similar to step (1100), an amount (A) characterizing the operating state of the technical system (11) parallel to these measurement signals (M) is provided (1600). The amount (A) characterizing the operating state of these technical systems may be, for example, determined by simulation or, preferably, measured. In this example, the amount (A) characterizing these operating states is also given as a time series.

[0034] Preferably, the most relevant part is identified from these provided measurement signals (M) using a trained recurrent neural network, and only this most relevant part is provided for subsequent training (1700).

[0035] Here, the machine learning system (200) is trained with a teacher (1800) to reconstruct, from each measurement signal (M), the quantity (A) characterizing the operating state of the technical system (11), using the measurement signal (M) provided in that way and the parallel quantity (A) characterizing the operating state of the technical system (11). In this embodiment, this is carried out by adjusting only the parameters of the second sub-system (202).

[0036] Thereby, the method ends.

Claims

1. A method for training a machine learning system (200) for evaluating a measurement signal (M) of a sensor (2) configured to identify at least one quantity (A) characterizing an operating state of a technical system (11), comprising: The machine learning system (200) includes a first subsystem (201), the first subsystem (201) being configured to identify a latent representation (L) from the measurement signal (M) provided to the first subsystem (201); the machine learning system comprises a second subsystem (202) configured to identify a quantity (A) characterizing an operational state of the technical system (11) from the latent representation (L), said first subsystem (201) being trained in an unsupervised or self-supervised manner, The method further comprises the step of: training the machine learning system (200) in a supervised manner.

2. During the supervised training, only the second partial system (202) is changed. The method of claim 1.

3. During the unsupervised or self-supervised training of the first subsystem (201), a portion of the measurement signals (M) is used, weighted by means of weighting factors, said weighting coefficients being identified as being particularly relevant for accurately determining a quantity characterizing an operating state of said technical system (11) using a model adapted to determine said quantity characterizing an operating state of said technical system (11) from said measurement signals (M), The method according to claim 1 or 2.

4. said model being a neural network, in particular a recurrent neural network, trained with parallel pairs of measurement signals (M) and quantities (A) characterizing the operating state of said technical system (11), The method according to claim 3.

5. A measurement signal evaluator, The measurement signal evaluator comprises a machine learning system trained according to claim 1; The measurement signal evaluator is configured to supply the measurement signal (M) supplied to the measurement signal evaluator to the machine learning system and to identify quantities characterizing an operating state of a technical system using the machine learning system and provide the quantities at the output side of the measurement signal evaluator.

6. A training system configured to carry out the method according to any one of claims 1 to 4.

7. 1. A method for determining a quantity characterizing an operating state of a technical system in dependence on a measurement signal, comprising: The measurement signal is supplied to a measurement signal evaluator, A method, wherein by means of said measurement signal evaluator a quantity characterizing an operating state of said technical system is provided.

8. A computer program arranged to cause a computer (20) to carry out the method according to any one of claims 1 to 5 or the method according to claim 7, when the computer program is executed on the computer (20).

9. A machine-readable storage medium (21) on which a computer program according to claim 8 is stored.

10. A computer (10) configured to carry out the method according to any one of claims 1 to 5 or the method according to claim 7.

Images