Method and apparatus for providing and evaluating a sensor model for change-point detection

Abstract

Method for evaluating a data-based sensor model (5) for determining a change-point time in a sensor signal time series (S), comprising the following steps: - Providing (S2) an evaluation signal time series (A) in an evaluation time window of a sensor signal time series (S); - Determining (S3) sensor signal segments (F1, F2, F3, F4) from the evaluation signal time series (A) that are temporally shifted or offset from each other by a number of sampling steps, wherein the sensor signal segments (F1, F2, F3, F4) have a shorter length than the evaluation signal time series; - Determining (S4) one or more frequency contributions from the sensor signal segments (F1, F2, F3, F4) using an FFT or a Goertzel algorithm; - Evaluating (S5, S6) the frequency contributions in a trained data-based sensor model (5) to determine a change-point time within the evaluation time window.

Description

Technical field The invention relates to a method for providing and evaluating a sensor model for detecting a change point time in a sensor signal time series, and in particular measures for providing a data-based sensor model for evaluating explainable physically motivated features. State of the art DE 10 2016 216 945 A1 discloses a method for executing a function based on a model value of a data-based functional model. DE 10 2018 217 310 A1 discloses a method for evaluating time series signals, in which a representation of a time series signal is analyzed to determine points of greater change, each of which is assigned a probability of being a point of greater change, and the determined points of greater change are examined using an empirical probability distribution to classify the determined points of greater change into recurrent and non-recurrent points of greater change. DE 10 2019 209 690 A1 discloses methods for controlling the fuel metering in an internal combustion engine. DE 10 2019 220 530 A1 discloses a method for operating a manufacturing system for the successive production of components. US patent 2007 / 0203616 A1 discloses a vehicle control device with a neural network. Technical background Sensors for measuring physical quantities are often continuously sampled. For example, a suitable sensor can be used to measure pressure, mass flow, acceleration, temperature, vibration, or similar parameters. The output of the sensor or sensor system typically provides a sensor signal time series as an electrical or digitized signal at predefined sampling times. This time series represents the temporal evolution of a sensor signal. For evaluation purposes, such a sensor signal time series can be analyzed, allowing specific characteristics of a technical system to be identified based on the course of the sensor signal. While sensor signals can be evaluated in a variety of ways, one application is to determine the point in time of a significant change in a system state, a so-called change point, by evaluating the sensor signal time series. For this purpose, a sensor model is typically provided that assigns information to a segment of the sensor signal time series, indicating a change point. Disclosure of the invention According to the invention, a method for evaluating a data-based sensor model for change-point detection according to claim 1, a method for training a data-based sensor model to provide a change-point time based on a predetermined sensor signal time series, and corresponding devices according to the dependent claims are provided. Further details are specified in the dependent claims. According to a first aspect, a method for evaluating a data-based sensor model to determine a change point time in a sensor signal time series is provided, comprising the following steps: - Providing an evaluation signal time series in an evaluation time window of a sensor signal time series; - Determining sensor signal segments from the evaluation signal time series that are time-shifted or offset from each other by a number of sampling steps, wherein the sensor signal segments have a shorter length than the evaluation signal time series; - Determining one or more frequency contributions from the sensor signal segments, in particular using an FFT (Fast Fourier Transform), DFT (Discrete Fourier Transform) or a Goertzel algorithm; - Evaluating the frequency contributions in a trained data-based sensor model to determine a change point time within the evaluation time window. As described at the beginning, the above method concerns a sensor model for evaluating a sensor signal time series from a conventional sensor that is continuously sampled in sampling steps. Such a sensor could be, for example, a pressure sensor, a mass flow sensor, an accelerometer, a vibration sensor, a radiation sensor, or the like. These types of sensors are typically sampled continuously at a predetermined sampling frequency to monitor changes over time, thus providing a sensor signal time series in analog or digital form. Such a sensor signal time series can be evaluated in a variety of ways. To monitor system states, it is often necessary to detect a point in time at which a significant change of state occurs in the technical system being measured. Such a point in time is called the change point. A group of data-driven sensor models has proven particularly effective for evaluating a sensor signal time series to determine a change point time. For this purpose, the sensor signal time series is sampled, and a time segment of the sensor signal is selected using an evaluation time window. The segment of the sensor signal time series captured within the evaluation time window is fed to the sensor model as an input vector. This evaluation signal time series can be configured as a data-driven classification model, so that, depending on the input vector, an output vector is generated, which is configured as a classification vector. This classification vector typically has a dimensionality with a number of elements, each of which is assigned a class and each of which is assigned to a specific time within the evaluation time window of the sensor signal time series.The argmax of the classification vector corresponds to the classification to be determined; that is, the index value of the relevant element in the output vector corresponds to a specific, predefined time within the evaluation time window. Thus, the sensor model can be configured to specify the change point time as a classification vector, where the change point time is defined as the argmax of the classification vector. By using the sensor model as a classification model, an evaluation signal time series is classified, and a change point time within the selected evaluation signal window is determined according to a trained sensor model. The value of the element of the classification vector, i.e., usually the element with the highest value, then has an index value that determines the time in the sensor signal time series that corresponds to the change point time. Training such a data-driven sensor model is typically performed using predefined training datasets in a manner known per se. The training datasets assign a classification vector as a label to an input vector (evaluation signal time series), which can be obtained by sampling a sensor signal within a predefined evaluation signal time window. One problem with purely neural network-based, data-driven sensor models is that their behavior is difficult to predict, and an output within a specific range cannot be guaranteed. This generally precludes their use in safety-critical systems, such as those relevant to driving in motor vehicles. The sensor model can be trained to assign a change point time to each frequency contribution in an evaluation point time series. The frequency contributions can refer to one or more predetermined frequencies. The above procedure involves preprocessing the evaluation signal time series to determine frequency characteristics that are physically explainable. If the data-based sensor model is evaluated with the frequency-based frequency characteristics, the behavior can be explained and the sensor model can thus be applied to safety-critical systems. The sensor model becomes explainable because the specific frequency features are physically motivated; that is, such a frequency feature indicates or is detected when the frequency dominates at a given point. These features are then combined using a linear function and output to Argmax as the detected class. This means the classification is based on the linear combination of the physical features. For each prediction, it can be determined which features (and with what weighting) it is based. According to the above procedure, an evaluation signal time series derived from a sensor signal time series is analyzed frequency-based to obtain section-by-section frequency contributions for one or more predefined frequencies. For this purpose, the evaluation signal time series is decomposed into several mutually offset sampling time windows, and the corresponding one or more frequency contributions are determined from the resulting signal time series segments. It may be possible to determine the frequency contributions based on one or more predefined frequencies and, in particular, the phase of an underlying sine or cosine signal. The frequency contributions can, for example, represent amplitude values ​​of a frequency spectrum at predefined frequencies, obtained, for instance, by an FFT or a Goertzel algorithm.For example, the frequency signal used to analyze the signal time series segment can correspond to a cosine signal with a predefined phase and frequency. These represent hyperparameters of the sensor model. Each individual signal time series segment yields one or more frequency contributions, each representing an input feature for the evaluation signal time series. These compress the information within the sensor signal time series into a comprehensible frequency contribution. The frequency contributions are then further processed using one or more neuronal layers of the sensor model. A regression value or a classification vector can be output. The regression value can directly specify the change point time, while the classification vector can specify the change point time via the index value as `argmax`. To train such a data-driven sensor model, corresponding frequency contributions are extracted as features from the training data time series and assigned to a corresponding label, namely a change-point time. In this case, the sensor model only needs to be trained by adjusting the model parameters of the neuronal layer. The training can be performed in a known manner using a gradient-based method. According to another aspect, a method for training a data-based sensor model to evaluate a point time series in order to determine a change point time is provided, comprising the following steps: - Providing training datasets, each specifying a point time series and a label in the form of a change point time; - Determining sensor signal segments from the point time series that are time-shifted or offset by a number of sampling steps, where the sensor signal segments have a shorter length than the point time series; - Determining one or more frequency contributions from the sensor signal segments using an FFT, DFT, or Goertzel algorithm; - Training the data-based sensor model with the frequency contributions and the change point times assigned to them. According to another aspect, a device for carrying out one of the above procedures is provided. Brief description of the drawings Embodiments are explained in more detail below with reference to the accompanying drawings. These show: Fig. 1 a schematic representation of a sensor system for acquiring sensor signal time series; Fig. 2 a flowchart illustrating a method for evaluating a data-based sensor model with prior frequency-based feature extraction; Fig. 3 a representation of an evaluation signal time series with sensor signal segments; Fig. 4 a flowchart illustrating a method for training a data-based sensor model with prior frequency-based feature extraction; and Fig. 5 a schematic representation of an injection system for injecting fuel into the cylinder of an internal combustion engine with a control unit in which the sensor model is implemented. Description of embodiments The following section describes in more detail the evaluation of a sensor model using a block diagram in Fig. 1 and a flowchart in Fig. 2. Fig. 1 shows a sensor system 1 with a sensor 2 designed for recording and acquiring continuous measurement signals. The sensor 2 can be, for example, a pressure sensor, a mass flow sensor, a temperature sensor, an accelerometer, a vibration sensor, a radiation sensor, or the like, and is sampled in step S1 at a sampling rate to obtain a continuous sensor signal time series S at discrete sampling steps. The sensor signal time series S can correspond to the detection of a varying physical quantity that changes, for example, according to a cyclic process. The cyclic process is detected and involves a cyclic change of state, which is reflected in a change of the physical quantity. In step S2, the sensor signal time series S is fed to a preprocessing block 3, which cyclically applies an evaluation time window to the sensor signal time series S to determine an evaluation signal time series A. The evaluation signal time series has a predefined number of samples generated from the sensor signal time series S. Depending on the specified evaluation time window, preprocessing block 3 creates the evaluation signal time series A as a vector of predetermined length. The evaluation signal time series A is positioned temporally relative to the sensor signal time series in such a way that it encompasses, as far as possible, the recurring state change of the change point time. In step S3, the evaluation signal time series A is fed to a feature extraction block 4. Feature extraction block 4 extracts signal time series segments from the evaluation signal time series A. Each segment corresponds to a section of the evaluation signal time series A and has a shorter length, e.g., between 30% and 70% of the length of the evaluation signal time series A. The signal time series segments are offset from each other, for example, by one or a predetermined number of samples. Figure 3 shows an example of how the signal time series segments F1, F2, F3, F4 can be positioned over the evaluation signal time series A. In feature extraction block 4, a frequency analysis function is applied to each of the signal time series segments F1, F2, F3, F4 in step S4, for example in the form of an FFT (Fast Fourier Transform), DFT (Discrete Fourier Transform), or a Goertzel algorithm. The Goertzel algorithm represents a special form of the discrete Fourier transform with which individual discrete spectral components can be efficiently calculated. Frequency analysis can be used to determine the spectral component, i.e., the frequency contribution, of one or more predetermined frequencies for each of the signal time series segments F1, F2, F3, F4. These predetermined frequencies correspond to predefined hyperparameters of the data-based sensor model. The one or more frequency contributions F for each of the signal time series segments F1, F2, F3, F4 are now fed in step S5 to a sensor model 5 in the form of a single- or multi-layer neural network. The neuron functions of the neural network are defined, in a known manner, as the sum of the output values ​​of the preceding neuron layer or the frequency contributions, weighted by a weighting vector, and a corresponding bias value. This sum can be applied a nonlinear activation function. The results can be output as a vector for further processing in a subsequent neuron layer or as a classification result. The sensor model 5 can therefore output an output vector O in step S6, which corresponds to a classification output. As described above, the output vector O has elements whose index value indicates a point in time or period within the evaluation time window and is permanently assigned to it. Figure 4 shows a flowchart illustrating the training of the data-based sensor model 5. Starting with training datasets provided in step S11, each comprising an evaluation signal time series A and optionally one or more further state variables of the technical system as well as an associated label in the form of a classification vector, these are first fed to the feature extraction block 4, as is also used for the evaluation of evaluation signal time series A described above. In step S12, the feature extraction block 4 divides the evaluation signal time series A into the signal time series segments F1, F2, F3, F4, as illustrated in Fig. 3, for each of which a frequency contribution or contributions are determined in the manner described above. From the signal time series segments F1, F2, F3, F4, corresponding one or more frequency contributions for predetermined frequencies and phases are obtained in step S13 through frequency analysis. These are fed as input variables in the form of an input vector to the sensor model 5, which is designed as a neural network. The neural network of sensor model 5 is now trained in step S14 according to the resulting frequency contributions. This means that, during training, the evaluation signal time series A, provided with a training dataset, is divided into several signal time series segments F1, F2, F3, F4, each representing a temporal segment of the evaluation signal time series A, offset from one another. The evaluation signal time series A is assigned, for example, a label in the form of a change point time, specifically in the form of a classification vector whose argmax specifies a change point time. The classification vector used for training can have the entry 1 at an index position corresponding to the label's change point time, while the value 0 is assigned at the other positions. The neural network is trained using well-known gradient-based methods, such as backpropagation, to adjust the model parameters, i.e., the weights and bias values ​​of the artificial neurons. Preferably, the neural network has two layers of neurons, with the output layer being configured solely to perform a dimensional reduction to the dimension of the classification vector as the output vector O. Fig. 5 shows, as an example of a sensor system 1, an injection system 10 for an internal combustion engine 12 of a motor vehicle, for which one cylinder 13 (of, in particular, several cylinders) is shown by way of example. The internal combustion engine 12 is preferably designed as a diesel engine with direct injection, but can also be designed as a gasoline engine. Cylinder 13 has an inlet valve 14 and an exhaust valve 15 for supplying fresh air and for removing combustion exhaust gas. Furthermore, fuel for the operation of the internal combustion engine 12 is injected into a combustion chamber 17 of the cylinder 13 via an injection valve 16. Fuel is supplied to the injection valve via a fuel supply line 18, through which fuel is provided at high fuel pressure in a manner known per se (e.g., common rail). The injection valve 16 has an electromagnetically or piezoelectrically controlled actuator unit 21, which is coupled to a valve needle 22. In the closed state of the injection valve 6, the valve needle 22 sits on a needle seat 23. By actuating the actuator unit 21, the valve needle 22 is moved longitudinally and opens part of a valve opening in the needle seat 23 to inject the pressurized fuel into the combustion chamber 17 of the cylinder 13. The injector 16 further comprises a piezoelectric sensor 25, which is arranged in the injector 6. The piezoelectric sensor 25 is deformed by pressure changes in the fuel flowing through the injector 6, and a voltage signal is generated as the sensor signal. The injection is controlled by a control unit 30, which specifies the amount of fuel to be injected by energizing the actuator unit 21. The sensor signal is sampled in time using an A / D converter 31 in the control unit 30, specifically at a sampling rate of 0.5 to 5 MHz. This results in a sensor signal time series. Furthermore, a pressure sensor 18 is provided to determine the fuel pressure upstream of the injection valve 16. During operation of the internal combustion engine 12, the sensor signal is used to determine the correct opening or closing time of the injection valve 16. For this purpose, the sensor signal is digitized into a corresponding evaluation point time series A using the A / D converter 31 and by specifying an evaluation time window. This series is then evaluated using the feature extraction described above and the trained data-based sensor model 5. From this, the opening time of the injection valve 16 and, accordingly, the injected fuel quantity can be determined, depending on the fuel pressure and other operating parameters. To determine the opening time, an opening time and a closing time are required in order to calculate the opening time as the time difference between these values. In conjunction with the sensor system 1 described above, the sampled pressure signal corresponds to the sensor signal time series, where the activation time for opening or closing the injection valve can be assumed to be the change point time for the label. The evaluation time window results from the cyclic repetition of the injection process in an internal combustion engine with a temporal position that essentially begins a predetermined time before the activated opening time and can be determined as the crankshaft angle.

Claims

Method for evaluating a data-based sensor model (5) for determining a change-point time in a sensor signal time series (S), comprising the following steps: - Providing (S2) an evaluation signal time series (A) in an evaluation time window of a sensor signal time series (S); - Determining (S3) sensor signal segments (F1, F2, F3, F4) from the evaluation signal time series (A) that are temporally shifted or offset from each other by a number of sampling steps, wherein the sensor signal segments (F1, F2, F3, F4) have a shorter length than the evaluation signal time series; - Determining (S4) one or more frequency contributions from the sensor signal segments (F1, F2, F3, F4) using an FFT or a Goertzel algorithm; - Evaluating (S5, S6) the frequency contributions in a trained data-based sensor model (5) to determine a change-point time within the evaluation time window. Method according to claim 1, wherein the sensor model (5) is trained to assign a change point time to each of the frequency contributions of an evaluation point time series (A). Method according to claim 1 or 2, wherein the frequency contributions are determined based on one or more predetermined frequencies and in particular a phase position of an underlying sine or cosine signal. Method according to any one of claims 1 to 3, wherein the sensor model (5) is provided as a single- or multi-layer neural network. Method for training a data-based sensor model (5) for evaluating a point-time series (A) to determine a change-point time, comprising the following steps: - Providing (S11) training datasets, each specifying a point-time series (A) and a label in the form of a change-point time; - Determining (S12) sensor signal segments (F1, F2, F3, F4) from the point-time series (A) that are time-shifted or offset by a number of sampling steps, wherein the sensor signal segments (F1, F2, F3, F4) have a shorter length than the point-time series (A); - Determining (S13) one or more frequency contributions from the sensor signal segments (F1, F2, F3, F4) using an FFT or a Goertzel algorithm; - Training (S14) the data-based sensor model (5) with the frequency contributions and the associated change-point time. Change point times. Method according to claim 5, wherein the data-based sensor model (5) is designed in the form of a deep neural network and is trained using a backpropagation-based training method. Method according to any one of claims 1 to 6, wherein the sensor model (5) is configured to specify the change point time as a classification vector, wherein the change point time is specified as argmax of the classification vector. Device for carrying out one of the methods according to one of claims 1 to 7. Computer program product comprising instructions which, when the program is executed by a computer, cause it to perform the steps of the method according to any one of claims 1 to 7. Machine-readable storage medium comprising instructions which, when executed by a computer, cause the computer to perform the steps of the method according to any one of claims 1 to 7.

Images