Method and device for improved evaluation of a measurement signal of a sensor
By using unsupervised and self-supervised training methods in machine learning systems, the potential representation of sensor measurement signals is solved, and the dependence on labeled data in the prior art is achieved, and efficient sensor signal evaluation is achieved.
Patent Information
- Application Number
- CN202411807135.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-12-11
- Filing Date
- 2024-12-10
- Publication Date
- 2025-06-13
AI Technical Summary
The prior art requires a large amount of labeled data when evaluating sensor measurement signals, resulting in complex and inefficient training processes.
Using machine learning systems, including the first subsystem and the second subsystem, potential representations are determined through unsupervised training or self-supervised training, and supervised training is used to determine parameters that characterize the operating state of the technical system, reducing dependence on labeled data.
It is realized that efficient sensor measurement signal evaluation results are obtained when a small amount of labeled data is used, and training efficiency and evaluation accuracy are improved.
Smart Images
Figure CN120146217A_ABST
Abstract
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] According to the prior art, virtual sensors can be used to determine a target parameter that depends on one or more relevant measurement parameters. Compared with traditional physical sensors, the target parameter is not directly measured, but is depicted by software based on its correlation with other measurement parameters. In this case, mathematical models, simulations, or artificial intelligence can be used.
[0003] Thus, for example, from the unpublished DE 102023206032.9, a method for evaluating a measurement signal of a sensor is known, the sensor being configured to determine at least one parameter characterizing the operating state of a technical system, wherein, for this purpose, the low-frequency component of the measurement signal is determined and / or the high-frequency component of the measurement signal is determined, and wherein different evaluation methods are used to perform the evaluation of the low-frequency component and the evaluation of the high-frequency component to determine the parameter characterizing the operating state of the technical system. Summary of the Invention
[0004] Advantages of the Invention
[0005] In contrast, the invention having the features of independent claim 1 has the following advantages: very good results can already be obtained with very little labeled data.
[0006] Other aspects of the invention are the subject matter of the independent claims in parallel. Advantageous refinements are the subject matter of the dependent claims.
[0007] Thus, in a first aspect, the invention relates to a method for training a machine learning system to evaluate a measurement signal of a sensor, the sensor being configured to determine at least one parameter characterizing the operating state of a technical system, wherein the machine learning system includes a first subsystem configured to determine a latent representation based on the measurement signal fed to the first subsystem, and wherein the machine learning system includes a second subsystem configured to determine the parameter characterizing the operating state of the technical system based on the latent representation, and wherein the first subsystem is trained unsupervised or self-supervised, and wherein the machine learning system is then trained supervised.
[0008] Generally, by means of a "latent representation", a (usually vector-valued) compressed representation of a (usually vector-valued) measurement signal is described.
[0009] This has the advantage that only a small amount of labeled data is required for training. This means that for a given amount of training data, the evaluation of the measurement signals functions particularly well.
[0010] The machine learning system can be, for example, a neural network.
[0011] It can be stipulated that during supervised training, only the second subsystem is changed, i.e., the first subsystem remains unchanged. This has the advantage that particularly little labeled data is required for training.
[0012] In this case, "changing the subsystem" usually may mean that the parameters characterizing the behavior of the subsystem are changed.
[0013] In an extended scenario, it can be stipulated that during unsupervised or self-supervised training of the first subsystem, parts of the measurement signals are weighted using weighting factors, where these weighting factors have been identified, by means of a model that is set up to determine parameters characterizing the operating state of the technical system based on the measurement signal, as being particularly important for correctly determining the parameters characterizing the operating state of the technical system.
[0014] In particular, in some embodiments, it may be possible to determine, from the model, the weighting factors for the sampling time points of the time series by means of the so-called "SHAP" method, and during unsupervised or self-supervised learning, the cost function characterizing the learning has a term in which these weighting factors are used to weight the contributions associated with the respective sampling time points.
[0015] Thereby, unsupervised or self-supervised training becomes particularly efficient.
[0016] If the measurement signal is given as a time series, parts of the measurement signal can, in some embodiments, be segments, in particular contiguous segments, of the measurement signal. In particular, it can be stipulated that for this purpose, the start time point and the end time point of the time series are specified in order to unambiguously characterize the part of the measurement signal.
[0017] The model can be a machine learning system, in particular a neural network, which is trained by means of parallel pairs of respectively the measurement signal and the parameters characterizing the operating state of the technical system. In this case, since the model is only used to identify or determine the most important part of the measurement signal for training the first subsystem, fewer pairs are required compared to a complete and correct reconstruction of the parameters characterizing the operating state of the technical system. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Subsequently, embodiments of the present invention are explained in more detail with reference to the accompanying drawings. In the
[0019] drawings:
[0020] Figure 1 Schematically shows a sensor in a braking system;
[0021] Figure 2 Schematically shows the structure of a training device of a machine learning system;
[0022] Figure 3 Schematically shows the information flow of a machine learning system;
[0023] Figure 4 Schematically shows a flow chart of a process according to an embodiment of the present invention. Detailed Embodiment
[0024] Figure 1 Schematically shows a braking system (1) of a motor vehicle (11), in this embodiment, the braking system has a plurality of sensors (2), these sensors include: a pressure sensor in a hydraulic supply line (3), using this hydraulic supply line to control the clamping force of a brake caliper (4); a sensor for determining the longitudinal acceleration of the motor vehicle (11); a sensor for detecting the temperature of the liquid in the hydraulic supply line (3); and a voltage sensor at a pump (not shown) for delivering this liquid. A multi-dimensional measurement signal (M) including the measurement signals of these sensors is delivered to a control device (10), this control device includes a computer-readable storage medium (20), on this computer-readable storage medium, a computer program is stored, on this computer program, a program for evaluating the measurement signal (M) is stored, so as to determine the torque that the brake caliper (4) transmits to a wheel (not shown) of the motor vehicle (11). This control device (10) also includes a processor (21), this processor is configured to: execute the computer program stored on this storage medium (20).
[0025] Figure 2 Schematically shows a training device (100) for training a machine learning system (200), this machine learning system includes a first subsystem (201) and a second subsystem (202). In this embodiment, this machine learning system is a neural network. This measurement device (100) includes a data memory (105), in this data memory, a plurality of measurement signals (M) are provided. This training device is configured to: select measurement signals (M) and deliver these measurement signals to this machine learning system (200).
[0026] The measurement signal (M) is delivered there to a first subsystem (201), this first subsystem is configured to: determine a latent representation (L) therefrom, that is, a vector whose dimension is lower than the dimension of the measurement signal (M).
[0027] The machine learning system (200) is also configured to convey the determined latent representation (L) to a second subsystem (202), which determines from it an output parameter (A) that characterizes the operating state of the technical system (11), which in this embodiment characterizes the estimated torque transmitted by the brake caliper (4). In a preferred embodiment, the parameter characterizing the operating state of the technical system (11) is given as a time series.
[0028] The training device (100) is configured to convey the latent parameter (L) determined by the machine learning system (200) and the output parameter (A) determined by the machine learning system (200) to the evaluation unit (110).
[0029] The training device (100) is also configured to train the first subsystem (201). In some embodiments, the training device (100) is configured to provide, for this purpose, an autoencoder (German: Autokodierer).
[0030] Figure 3 The information flow during operation of the machine learning system (200) is schematically illustrated. The providing unit (300) provides the measurement parameters (M1,..., M4) as a time series of parameters determined from sensor data, which in this embodiment are the torque, temperature, voltage, and position of the brake caliper (4) determined from the current. Thus, at a specifiable time point (t0), these measurement parameters correspond to a four-dimensional vector. These time series are conveyed to the first subsystem (201), which determines from them a latent representation (L), which in the illustrated example is a three-dimensional vector at the corresponding time point (t0). In this latent representation (L), the progression of the time series of the measurement parameters (M1,..., M4) correspondingly results in a progression in three-dimensional space. This latent representation (L) is conveyed to the second subsystem (202), which determines from it an output parameter (A). In this embodiment, the output parameter (A) is one-dimensional, and thus, at the time point (t0), this output parameter corresponds to a scalar value.
[0031] In some embodiments, the measurement parameters (M1,..., M4) and the latent representation (L) are also conveyed to another machine learning system (203), which is trained to estimate the uncertainty of the output parameter (ΔA) based on these parameters. The training of this another machine learning system (203) can be carried out, for example, during the operation of the machine learning system (200) Figure 4It is carried out in the training method shown, so that the machine learning system (200) can actually reconstruct the output parameter (A) according to the corresponding measurement signal (M), such that the output parameter actually corresponds to the value of the provided measured or analog-provided output parameter (A) at the corresponding time point (t0). This level of uncertainty can be used as a target in the supervised training of the other machine learning system (203).
[0032] Figure 4 An exemplary flow of a computer program is shown in a flowchart. First (1000), a plurality of measurement signals (M) are received from a pressure sensor (2), and these measurement signals are provided as time series data.
[0033] Then, parameters (A) characterizing the operating state of the technical system (11) (in this embodiment, the braking system (1)) are provided (1100) in parallel with these respective time series. These parameters (A) characterizing the operating state of the technical system can be determined, for example, in an analog manner or preferably measured. In this embodiment, these parameters (A) characterizing the operating state are given as a time series again.
[0034] Now, with the pairs of the measurement signal (M) and the parameters (A) of the operating state of the parallel technical system (1), a neural network (in this embodiment, a recurrent neural network) is trained (1200) to: reconstruct, as output parameters at the output of the recurrent neural network, the parameters (A) of the operating state of the parallel technical system (1) according to the measurement signals (M) respectively provided as input parameters to the recurrent neural network.
[0035] In this case, the recurrent NN is also trained to: for the measurement data, identify the most important part of the time series for determining the operating data. In this case, in this embodiment, using the "SHAP" method, a weighting factor is assigned 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.
[0036] Now, a second set of measurement signals (M) is provided (1300), which in some embodiments is larger than the set of measurement signals provided in step (1000).
[0037] Using the measurement signals (M) of the second set, the first subsystem (201) is trained (1400). In some embodiments, the first subsystem is trained unsupervised, and in some embodiments, the first subsystem is trained self-supervised, for example, using an autoencoder or by contrastive learning. To this end, weighted factors identified by means of a recurrent NN for the provided measurement data (M) are provided and used as weights for the corresponding measurement data (M) in order to train the first subsystem (201). In some embodiments, the training is performed according to 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 these vectors corresponds to a sampling time point, and the respective identified weighted factors are used to weight these squared intervals for the individual sampling time points in order to obtain the cost function for the training.
[0038] Now, a third set of measurement signals (M) is provided (1500), and for these measurement signals (M), similar to step (1100), parameters (A) characterizing the operating state of the parallel representation technology system (11) are provided (1600). These parameters (A) characterizing the operating state of the technology system can be determined, for example, in an analog manner or preferably measured. In this embodiment, these parameters (A) characterizing the operating state are given again as a time series.
[0039] Preferably, from the provided measurement signals (M), the most important parts are identified by means of a trained recurrent neural network, and only these most important parts are provided (1700) for subsequent training.
[0040] Now, the machine learning system (200) is supervised-trained (1800) using the measurement signals (M) provided in this way and the parameters (A) characterizing the operating state of the parallel representation technology system (11) to: reconstruct the parameters (A) characterizing the operating state of the representation technology system (11) according to the corresponding measurement signals (M). In this embodiment, this is achieved by only adjusting the parameters of the second subsystem (202).
[0041] The method ends hereby.
Claims
1. A method for training a machine learning system (200) for evaluating a measurement signal (M) of a sensor (2), the sensor being configured to determine at least one variable (A) characterizing an operating state of a technical system (11), in, The machine learning system (200) comprises a first subsystem (201) configured to determine a latent representation (L) based on a measurement signal (M) supplied to the first subsystem, and The machine learning system comprises a second subsystem (202) configured to determine a parameter (A) characterizing an operating state of the technical system (11) based on the latent representation (L), and wherein the first subsystem (201) is trained unsupervised or self-supervised, and wherein the machine learning system (200) is then trained supervised.
2. The method according to claim 1, wherein: During the supervised training, only the second subsystem (202) is changed.
3. A method according to any one of the preceding claims, wherein: During unsupervised or self-supervised training of the first subsystem (201), parts of the measurement signal (M) are used in a weighted manner with the aid of weighting factors, wherein the weighting factors have been identified as being particularly important for the correct determination of the parameters characterizing the operating state of the technical system (11) with the aid of a model set up to determine the parameters characterizing the operating state of the technical system (11) based on the measurement signal (M).
4. The method according to claim 3, wherein: The model is a neural network, in particular a recurrent neural network, which is trained with the aid of parallel pairs of measured signals (M) and variables (A) characterizing the operating state of the technical system (11), respectively.
5. A measurement signal evaluator, comprising a machine learning system trained according to claim 1, and the measurement signal evaluator is configured to: supply a measurement signal (M) supplied to the measurement signal evaluator to the machine learning system; and determine a parameter characterizing an operating state of the technical system with the aid of the machine learning system; and provide the parameter at an output of the measurement signal evaluator. 6 . A training system, which is configured to carry out the method according to claim 1 .
7. A method for determining a variable characterizing an operating state of a technical system from a measurement signal, in, The measurement signal is supplied to the measurement signal evaluator and a variable characterizing an operating state of the technical system is provided by means of the measurement signal evaluator.
8. A computer program, which is designed to cause a computer (20) to execute the method according to any one of claims 1 to 5 or 7 when the computer program is executed on the computer.
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 7.