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

Abstract

The invention relates to a method for evaluating a measurement signal (M) of a sensor configured to determine at least one variable (A) that characterizes an operating state of a technical system, the evaluation being performed by means of a model (301) and a trained machine learning system (302), the model (301) and the machine learning system (302) being weighted on the basis of how similar the measurement signal (M) is to training data that were taken as a basis for training the machine learning system (302).

Description

[0001] Description

[0002] title

[0003] Method and device for improved evaluation of measurement signals from a sensor

[0004] 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.

[0005] State of the art

[0006] According to the state of the art, a virtual sensor can be used to determine a target variable that depends on one or more correlated measured variables. Unlike conventional physical sensors, the target variable is not measured directly, but rather mapped by software based on correlations with other measured variables. Mathematical models, simulations, or artificial intelligence can be used for this purpose.

[0007] For example, from the unpublished DE 10 2023 206 032.9, a method for evaluating a measurement signal of a sensor which is configured to determine at least one variable characterising an operating state of a technical system is known, wherein for this purpose a low-frequency component of the measurement signal is determined 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 characterising the operating state of the technical system.

[0008] Advantages of the invention The invention with the features of independent claim 1 has the advantage that it provides particularly reliable results in a wide range of measurement signals.

[0009] Further aspects of the invention are the subject of the independent claims. Advantageous further developments are the subject of the dependent claims.

[0010] Disclosure of the invention

[0011] In a first aspect, the invention relates to a method for evaluating a measurement signal of a sensor to determine a characteristic quantity, wherein the evaluation is carried out by means of a model, in particular a physical model, and a trained machine learning system, wherein a weighting of the model and the machine learning system is carried out depending on how similar the measurement signal is to training data on the basis of which the machine learning system was trained.

[0012] In other words, this method is a virtual sensor for the quantity characterizing the operating state of the technical system.

[0013] The similarity between the measurement signal and the training data is a particularly simple measure for determining the reliability of a machine learning system. This makes it possible to determine the variable characterizing the operating state of the technical system as reliably as possible.

[0014] In some embodiments, it is provided that the measurement signal is provided to the model and the trained machine learning system at a respective input of the model or the trained machine learning system.

[0015] In some embodiments, it is provided that an estimated value of the variable characterizing the operating state of the technical system is provided at an output of the model, and that a correction value of this estimated value is provided at an output of the trained machine learning system. In other embodiments, it can be provided that an estimated value of the variable characterizing the operating state of the technical system is provided at each output of the model and the machine learning system.

[0016] In further developments, it may be provided that the similarity of the measurement signal to the training data is carried out using an out-of-distribution detection method.

[0017] The term “out-of-distribution detection” is derived from English and refers to procedures that detect whether the measurement signal comes from the same statistical distribution as the training data.

[0018] Furthermore, it can be provided that the similarity of the measurement signal to the training data is determined with the aid of an autoencoder trained with the training data.

[0019] This is a particularly simple way to perform out-of-distribution detection. In other words, in some embodiments, the autoencoder and machine learning system are trained using the same training data.

[0020] In some embodiments, it may be provided that the similarity is determined depending on a reconstruction error when the measurement signal is applied to an input of the autoencoder and a reconstruction of this measurement signal is provided at an output of the autoencoder.

[0021] In some embodiments, the contributions of the machine learning system and the model may be combined additively and / or multiplicatively.

[0022] These combination options are particularly simple ways to combine both model contributions and data-based learned contributions.

[0023] In some embodiments, it may be provided that the weighting evaluates a contribution of the machine learning model less, the lower the similarity between the measurement signal and the training data.

[0024] In some embodiments, a “lesser contribution” is to be understood as meaning that a correction of an output of the model caused by the machine learning system is less.

[0025] In other words, this contribution weights how the determined value of the virtual sensor is composed of the machine learning system and the model. In some embodiments, the reconstruction error of the autoencoder determines this weight. The model's contribution to the determined value of the virtual sensor increases with increasing reconstruction error.

[0026] This makes it possible to determine the quantity characterizing the operating state of the technical system as a determined value of the virtual sensor as reliably as possible.

[0027] Embodiments of the invention are explained in more detail below with reference to the accompanying drawings. In the drawings:

[0028] Figure 1 shows schematically a sensor in a braking system;

[0029] Figure 2 schematically shows an embodiment of a device for determining a variable that characterizes the operating state of the technical system;

[0030] Figure 3 schematically shows a further embodiment of a device for determining a variable that characterizes the operating state of the technical system;

[0031] Figure 4 schematically shows a device for training an autoencoder and one of the aforementioned devices;

[0032] Figure 5 shows a schematic flow diagram of a process according to an embodiment of the invention. Figure 1 shows a schematic diagram of a braking system (1) of a motor vehicle (11) with, in the exemplary embodiment, a plurality of sensors (2) comprising a pressure sensor in a hydraulic supply line (3) with which a contact pressure 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 fluid in the hydraulic supply line (3) and a voltage sensor on a pump (not shown) for conveying the fluid.A multidimensional measurement signal (M) comprising the measurement signals from these sensors is fed to a control unit (10) comprising a computer-readable storage medium (20) on which a computer program is stored, which program is stored for evaluating the measurement signal (M) in order to determine a torque that the brake caliper (4) transmits to a wheel (not shown) of the motor vehicle (11). The control unit (10) further comprises a processor (21) configured to execute the computer program stored on the storage medium (20).

[0033] Figure 2 schematically shows an embodiment of a device (300) for evaluating the measurement signal (M) to determine the variable (A) characterizing the operating state of the technical system—in this embodiment, the torque that the brake caliper (4) transmits to the wheel. In this embodiment, this device (300) is stored as a computer program on the storage medium (20) and is executed by the control unit (10).

[0034] This device (300) comprises a physical model (301) that physically models the relationship between the measurement signal (M) and the transmitted torque (A). The measurement signal (M) is fed to an input (N1) of the physical model (301), which, using the physically modeled relationship, provides an estimated value (AP) of the transmitted torque at an output (X1) of the physical model (301).

[0035] The device (300) further comprises a machine learning system (302), in the exemplary embodiment a neural network. The measurement signal (M) is supplied to the neural network (302) at a first input (N2) and the estimated value (AP) of the transmitted torque is supplied to a second input (N3). The neural network (302) determines therefrom a correction value (DA), which is provided at an output (X2) of the neural network (302).

[0036] The estimated value (AP) and the correction value (DA) are fed to a weighted combiner (303) which, in the exemplary embodiment, determines the torque (A) that the brake caliper (4) transmits to the wheel from a weighted sum of the estimated value (AP) and the correction value (DA).

[0037] In embodiments this is done using the formula

[0038] A = AP + b - DA (1) with a predeterminable weighting factor (b).

[0039] The device (300) further comprises a weighting block (250) which specifies the weighting factor (b) depending on the measurement signal (M). In the exemplary embodiment, the weighting block (250) comprises an autoencoder which is configured to determine a reconstruction of the measurement signal (M). The weighting block (250) is configured to determine the weighting factor (b) depending on the reconstruction error between the measurement signal (M) and the reconstruction of the measurement signal (M). In the exemplary embodiment, a characteristic curve is stored in the weighting block (250) which specifies the value b = 1 when the reconstruction error is zero and decreases with increasing reconstruction error, with b -> 0 for a large reconstruction error.

[0040] Figure 3 shows, in one exemplary embodiment, a further embodiment of the device (300) for evaluating the measurement signal (M) to determine the variable (A) characterizing the operating state of the technical system—in this exemplary embodiment, the torque that the brake caliper (4) transmits to the wheel. In this exemplary embodiment, this device (300) is stored as a computer program on the storage medium (20) and is executed by the control unit (10).

[0041] This device (300) also comprises a physical model (301) that physically models the relationship between the measurement signal (M) and the transmitted torque (A). The measurement signal (M) is fed to an input (N1) of the physical model (301), which, using the physically modeled relationship, provides an estimated value (AP) of the transmitted torque at an output (X1) of the physical model (301).

[0042] The device (300) further comprises a machine learning system (302), in the exemplary embodiment a neural network. The measurement signal (M) is supplied to the neural network (302) at a first input (N2). The neural network (302) determines therefrom an estimated value (MA) of the transmitted torque, which is provided at an output (X2) of the neural network (302).

[0043] The two estimated values ​​(AP, AM) are (DA) fed to a weighted combiner (303) which, in the exemplary embodiment, determines the torque (A) that the brake caliper (4) transmits to the wheel from a weighted sum of the estimated values ​​(AP, AM).

[0044] In embodiments this is done using the formula

[0045] A = a • AP + b • AM (2) with predefined weighting factors (a), (b). These are preferably determined under the constraint a + b = 1 (3).

[0046] The device (300) further comprises a weighting block (250) which specifies the weighting factors (a), (b) depending on the measurement signal (M). In the exemplary embodiment, the weighting block (250) comprises an autoencoder which is configured to determine a reconstruction of the measurement signal (M). The weighting block (250) is configured to determine the weighting factors (a), (b) depending on the reconstruction error between the measurement signal (M) and the reconstruction of the measurement signal (M). In the exemplary embodiment, the weighting factor (b) is determined analogously to that discussed in connection with Figure 2. The weighting factor (a) is determined using a = 1 - b (4).

[0047] Figure 4 shows, in one embodiment, a mode of operation of a training device (100) for training the devices (300) illustrated in Figures 2 and 3 for determining the variable (A) characterizing the operating state of the technical system as a function of the measurement signal (M).

[0048] The device (100) comprises a provision block (105) for providing measurement signals (M) from a training data set comprising a plurality of measurement signals (M) and associated target values ​​of variables (AS) characterizing the operating state of the technical system.

[0049] The training device (300) comprises a processor (120) for executing the training algorithm implemented as a computer program, which is stored on a machine-readable storage medium (121) of the training device (300).

[0050] The measurement signals (M) are fed to the device (300), which uses them to determine the variable (A) characterizing the operating state of the technical system. The associated target variable (A') is fed to a comparison device (110), which, depending on a cost function dependent on this variable (A) and the associated target variable (A'), typically in the form of supervised training, adapts parameters of the device (300), and in particular of the machine learning system (302), such that the device (300) is as well-equipped as possible to correctly output the variable (A).

[0051] The training device (100) is also configured to receive the measurement signals

[0052] (M) of the same training data set to train the autoencoder (200) in an unsupervised manner. The autoencoder (200) typically comprises an encoder (201) that converts the measurement signals (M) into a latent vector (L) of lower dimensionality, and it includes a decoder (201) that determines a reconstruction (M') of the measurement signal (M) from this latent vector (L). The measurement signals (M) are transmitted to the comparison device (110). From a comparison of the measurement signal (M) and the associated reconstruction (M'), the parameters of the autoencoder (200) are typically adjusted using a cost function that penalizes the distance between the measurement signal (M) and the reconstruction (M') such that the reconstruction (M') approximates the associated measurement signal (M) as closely as possible.

[0053] Figure 5 shows a flowchart of an embodiment of the method as it can run in the control unit (10). In some embodiments, the method does not include upstream training (1000) of the device (300) and the autoencoder (300) using the training device (100).

[0054] Subsequently, during operation, measurement data (M) are fed to the device (300). (1100) Here, weighting parameters (a), (b) are determined using the weighting block (250) and fed to the combiner (303).

[0055] As illustrated in Figures (2) and (3), the physical model (301) and the machine learning model (302) determine their respective contributions (AP), (AM), (DA) and feed them to the combiner (303) (1200).

[0056] In the combiner (303), these contributions are combined according to formula (1) and formula (2) to determine the quantity (A) characterizing the operating state of the technical system (1300). This concludes the process.

Claims

Claims 1 . Method 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 evaluation is carried out by means of a model (301) and a trained machine learning system (302), wherein a weighting of the model (301) and the machine learning system (302) is carried out depending on how similar the measurement signal (M) is to training data on the basis of which the machine learning system (302) was trained.

2. The method according to claim 1, wherein the similarity of the measurement signal (M) to the training data is carried out by means of an out-of-distribution detection method.

3. The method according to claim 2, wherein the similarity of the measurement signal (M) to the training data is determined with the aid of an autoencoder (200) trained with the training data.

4. The method according to claim 3, wherein the similarity is determined as a function of a reconstruction error when the measurement signal (M) is applied to an input of the autoencoder (200) and a reconstruction of this measurement signal (M') is provided at an output of the autoencoder.

5. Method according to one of the preceding claims, wherein the contributions of the machine learning system (AM, AD) and the model (AP) are combined additively and / or multiplicatively.

6. Method according to one of the preceding claims, wherein the weighting (b) evaluates a contribution of the machine learning model (AM, AD) less, the lower the similarity between the measurement signal (M) and the training data.

7. A computer program configured to cause a computer (10) to execute the method according to any one of claims 1 to 6 when executed on the computer (10).

8. A machine-readable storage medium (21) on which the computer program according to claim 7 is stored.

9. A control device (10) configured to carry out the method according to one of claims 1 to 6.

Images