Training method and method for detecting a system state of a physical system

Abstract

The invention relates to a computer-implemented method for training a machine-learning algorithm to detect whether a physical system is in a fault-free state or a faulty state; the method comprising: providing (S1) training data which have at least one measurement signal relating to a measurement variable of the physical system; extracting (S2) at least one characteristic feature from the at least one measurement signal; determining (S3) a limit function on the basis of the at least one characteristic feature and / or on the basis of the at least one measurement signal by the machine-learning algorithm; and training (S4) the machine-learning algorithm on the basis of the training data in order to detect whether the physical system is in a fault-free state or a faulty state, depending on the determined limit function by ascertaining at least one fault indicator.

Description

[0001] Description

[0002] title

[0003] Training methods and methods for detecting a system state of a physical system

[0004] The invention relates to a computer-implemented method for training a machine learning algorithm to detect whether a physical system is in a fault-free state or a faulty state. Furthermore, the invention relates to a computer-implemented method and a device for detecting whether a physical system is in a fault-free state or a faulty state.

[0005] State of the art

[0006] Crankcase ventilation is a critical component of the engine system of modern vehicles. It plays a crucial role in maintaining engine performance and efficiency by removing excess gases and contaminants from the crankcase. Without properly functioning crankcase ventilation, harmful gases can build up in the engine, leading to loss of power, increased wear, and even serious engine damage.

[0007] Crankcase ventilation is typically achieved via a hose or pipe connection connected to the engine's intake tract. According to legal regulations regarding on-board diagnostics in a vehicle, the loss of such hoses must be diagnosed unless the hoses are permanently connected to the ventilation system. When the engine is throttled, "blow-by gases" or oil vapors, among other things, can be discharged and burned due to the negative pressure in the intake manifold. "Blow-by gases" refer to the gases and vapors generated by the combustion process in the engine that enter the engine's crankcase. These gases include hydrocarbons, carbon monoxide, nitrogen oxides, sulfur oxides, and water vapor.

[0008] In turbocharged engines, an additional connection to the intake tract is used upstream of the compressor, creating a vacuum through the Venturi effect. The Venturi effect is a physical phenomenon in which the flow velocity of liquids or gases in a pipe is increased by a constriction. This means that when a fluid flows through a narrow section in a pipe, the fluid's velocity increases while the pressure decreases.

[0009] Diagnosis of hose disconnection is typically performed by detecting a vacuum in the crankcase. If the measured pressure approaches ambient pressure, it is concluded that the hose connection has disconnected. This is output by the detection system, for example, in the form of a "Not OK" (NOK) status. However, if an operating-point-dependent vacuum is measured, the hose connection is correctly connected (OK status).

[0010] However, in engines with large flow cross-sections in the intake tract, robustness problems can arise when diagnosing the hose failure, since negative pressure is no longer reliably generated and therefore measurable, particularly in the medium load range of the engine.

[0011] Instead, pressure measurements in the vent hose often show oscillations. Since these oscillations generally disappear when the hoses become loose, the diagnosis is performed, for example, based on the sum of the signal changes of a relative pressure signal. If this sum is above a limit value, the hoses are still connected (and thus oscillations can be measured). If, however, the sum is below the limit value, the hoses are no longer connected but may have fallen off. Since the characteristics of the pressure oscillations can change significantly across the relevant operating range or load range of the engine, this type of diagnosis can result in very limited operating windows within which the hoses become loose can be detected with sufficient accuracy, or in robustness problems when detecting the hoses becoming loose.

[0012] Similar methods and / or approaches can also be used to generally check a system state of a physical system to determine whether a malfunction and / or a faulty state exists or whether the system is in a fault-free state.

[0013] One object of the invention is to provide an improved method for detecting a system state using a trained machine learning algorithm. Furthermore, one object of the invention is to provide a training method for training the machine learning algorithm to detect the system state.

[0014] The problem is solved by a computer-implemented method for training a machine learning algorithm for detecting whether a physical system is in a fault-free state or a faulty state, according to the features of patent claim 1. The problem is solved by a computer-implemented method for detecting whether a physical system is in a fault-free state or a faulty state.

[0015] Disclosure of the invention

[0016] In a first aspect, a computer-implemented method is specified for training a machine learning algorithm to detect whether a physical system is in a fault-free state or a faulty state; the method comprising: providing training data comprising at least one measurement signal about a measured variable of the physical system; extracting at least one characteristic feature from the at least one measurement signal; determining a limit function based on the at least one characteristic feature and / or based on the at least one measurement signal by the machine learning algorithm; and training the machine learning algorithm based on the training data to detect, depending on the determined limit function, whether the physical system is in a fault-free state or a faulty state by determining at least one error indicator.In this case, after the step of determining, a step of determining at least one error indicator as a function of the determined limit function by the machine learning algorithm and a step of training the machine learning algorithm on the basis of the training data can be provided in order to determine, as a function of the error indicator, whether the physical system is in a fault-free state or a faulty state.

[0017] In a second aspect, a computer-implemented method is specified for detecting whether a physical system is in a fault-free state or a faulty state, using a machine learning algorithm trained according to the method according to the first aspect; the method comprising: providing the at least one measurement signal via the at least one measured variable of the physical system; extracting at least one characteristic feature from the at least one measurement signal; determining the at least one fault indicator based on the at least one measurement signal and / or based on the at least one characteristic feature; and detecting whether the physical system is in a fault-free state or a faulty state based on the at least one fault indicator.

[0018] It is understood that the steps according to the invention and further optional steps do not necessarily have to be carried out in the order shown, but can also be carried out in a different order. Furthermore, further intermediate steps can be provided. The individual steps can also comprise one or more sub-steps without thereby departing from the scope of the method according to the invention. The detection of the system state is described again below for the case in which the physical system is a ventilation device for ventilating a crankcase. In this case, the at least one measurement signal is particularly preferably a pressure measurement signal detected by a pressure sensor of the ventilation device.In the faulty state of the venting device, it is then detected that a hose of the venting device is incorrectly connected or has fallen off. It is understood, however, that the embodiments are also adaptable to other physical systems in which a fault condition is to be detected. Thus, the present disclosure is not limited to venting devices. Rather, these are mentioned purely as examples and relate only to a particularly preferred embodiment of the physical system. The system state can be detected particularly preferably when an internal combustion engine of a motor vehicle is in supercharged operation, i.e., preferably a throttle valve is fully open and a turbocharger of the internal combustion engine is running.In this case, before beginning a particularly ongoing and / or continuous and / or sequential detection of the system state, a predetermined number of measured values ​​of the at least one measurement signal should first be measured. Once this predetermined number has been recorded and / or measured, the characteristic features can preferably be extracted at least from this number of determined measured values. The predetermined number is then preferably always supplemented by the current measured value(s) in order to enable real-time monitoring and / or detection of the system state.

[0019] Furthermore, the trained machine learning algorithm can determine whether the current state of the physical system represents a transitional state, in which, among other things, the OK state and the nOK state are difficult or impossible to distinguish. The trained machine learning algorithm can also indicate whether the respective measured values ​​of at least one measurement signal deviate significantly from the training data used to train the algorithm. If such a deviation is detected, the model can be retrained, for example. The detection of such a transitional state and / or the detection of a measured value deviation can preferably be learned during the training phase.

[0020] The inventive method for detecting the system state preferably captures complex physical relationships of the physical system through the use of machine learning. The trained machine learning algorithm is then preferably stored in the form of a detection model. The inventive procedure represents a machine-learning alternative to the analytical, operating-point-dependent weighting of various features of a recorded measurement signal used in conventional software development. The preferred definition of associated limit values ​​for distinguishing an OK state from a noOK state preferably reduces the present method essentially to two parameters, via which diagnostic results of the at least one measurement signal can be accepted for state detection or rejected for such state detection.Measured values ​​of the at least one measurement signal are preferably rejected if the measured values ​​are subject to a high degree of uncertainty. Diagnostic results or measured values ​​of the at least one measurement signal that are in a transition region between an OK state and a nOK state are preferably subject to a high degree of uncertainty. In this transition region, the two states (OK, nOK) are difficult or impossible to distinguish from one another. Measured values ​​whose characteristic features are unknown to the detection model are also subject to a high degree of uncertainty, for example because they were not recorded during the measurements to generate the training data or were not used to train the machine learning algorithm.During operation, the detection model preferably diagnoses the state of the physical system, such as the state of a hose connection of a venting device, based on sequences of the at least one measurement signal that have a predetermined length. It also uses at least one limit value to determine the reliability and / or accuracy with which the state is determined. Measurement signals and / or sequences of measurement signals that are subject to high uncertainty are preferably discarded. The described method for state detection makes it possible to capture the characteristic features of the at least one measurement signal, particularly across a relevant operating range of the physical system. This enables a clearer distinction between an OK state and a nOK state and expands the operating range of the diagnosis compared to the prior art.Together with their continuous running capability, this increases the robustness of the diagnosis compared to the state of the art.

[0021] In one embodiment, the trained machine learning algorithm comprises a support vector machine and a Gaussian process regressor, wherein the boundary function is determined by the support vector machine, and wherein the at least one error indicator is determined by the Gaussian process regressor as a function of the boundary function. Particularly preferably, the machine learning algorithm comprises two sub-algorithms, namely the support vector machine and a Gaussian process regressor, at least for training purposes. In the trained state of the algorithm, preferably only the Gaussian process regressor is used for implementation on a control unit. The support vector machine, on the other hand, is preferably only required during training to determine the boundary function. By advantageously choosing the model type, the transition region in which the physical behavior of OK and nOK is similar or identical can be detected particularly preferably.Alternatively or additionally, it is possible for the detection model to at least classify data with an unknown characteristic as such, namely if such data was not present in the training data. During training, the present method preferably uses a sub-algorithm of the support vector machine (SVM) type to determine, by means of machine learning, a boundary function, for example a boundary curve and / or a boundary surface, by which respective measured values ​​of the at least one measurement signal for determining the OK state and the nOK state can be separated or demarcated from one another. The boundary function preferably specifies how the measured values ​​can be categorized in order to be as clearly demarcated as possible. Furthermore, the present training method uses a Gaussian process regressor (GPR) and trains it to predict the shortest distance between the measured values ​​or measuring points and the determined boundary function.The present method can preferably be executed within a few seconds to minutes on a control unit and / or a computer or laptop. The SVM preferably serves as an auxiliary model during training, based on which the GPR is trained.

[0022] A support vector machine (SVM) is a machine learning model used for classification and regression. SVMs belong to the so-called "supervised learning" algorithms and are based on the idea of ​​finding a dividing line or surface between different classes of data points. The goal of the SVM is to find the dividing line in such a way that the distance between the nearest data points and the dividing line, also known as the margin, is maximized. The data points closest to the dividing line are called support vectors. The SVM model is also capable of handling nonlinear data by projecting the data into a higher-dimensional space and finding a dividing line or surface there. This process is called the kernel trick.

[0023] In one embodiment, the at least one error indicator has a magnitude distance of a respective measurement point of the at least one measurement signal from the limit function and / or sign information of the respective measurement point of the at least one measurement signal relative to the limit function and / or a standard deviation of the predicted distance of the respective measurement point of the at least one measurement signal from the limit function. During inference, the GPR preferably predicts, from a measured value of the at least one measurement signal, its distance from the limit function and / or a standard deviation from the distance prediction made. The following information can preferably be derived from these two variables. On the one hand, the sign of the distance from the limit function can indicate whether it is more of an OK state or more of a nOK state.This can be done because the boundary function preferably represents a separation barrier for distinguishing the OK state from the nOK state. Furthermore, it can preferably be determined from an absolute value of the distance, in particular by comparison with a predetermined distance limit value, whether the respective measured value is located in a transition range in which the OK state and the nOK state are similar, such that a clear distinction between the OK state and the nOK state is not possible. Alternatively or additionally, the standard deviation from the distance prediction made, in particular in comparison with a predetermined, applicable limit value, can be used to indicate whether a respective measured value lies within a predetermined data range in which it is sufficiently similar to the training data used for model training.If a measured value lies within the transition range or the standard deviation of the prediction is too large, then the prediction is preferably rejected; otherwise, it is preferably recognized as a valid diagnostic result for condition diagnosis.

[0024] In one embodiment, the at least one characteristic feature has at least one oscillation characteristic and / or at least one autocorrelation of the at least one measurement signal and / or a sum of magnitudes of at least one signal change within the at least one measurement signal and / or a mean value of the at least one measurement signal and / or a standard deviation. Other characteristic features are also conceivable. An autocorrelation can be determined using the example of crankcase ventilation, for example, by considering an air extraction frequency. Using the example of crankcase ventilation, the at least one oscillation characteristic can be induced, for example, by an intake system of an engine.

[0025] In one embodiment, the at least one measurement signal comprises a time-series-based measurement signal, wherein the machine learning algorithm is trained to detect the state of the physical system based on a predetermined time window of the time-series-based measurement signal. From a physical perspective, for example, a ring buffer can be provided in which a specific number of measurement values ​​of the at least one measurement signal is always stored in order to then be available for further signal processing, in particular for provision to the machine learning algorithm. For example, a number of X measurement values ​​can be stored in such a ring buffer, which can then be further processed, for example to extract the characteristic features therefrom.Such a ring buffer is preferably continuously updated with a current measured value, so that the current X measured values ​​are always available for evaluation. The detection model preferably requires only short measurement windows for condition diagnosis and can, in particular, be implemented and / or executed continuously.

[0026] In one embodiment, the machine learning algorithm can be retrained in the event of a change in the physical system and / or a change in the at least one measurement signal. A change in the oscillation characteristic in the at least one measurement signal, for example caused by an adjustment and / or change to a part and / or component of the physical system, including a change in the piping of the venting device, particularly for packaging reasons, can be captured by retraining the model and incorporated directly into a preferred control unit software in the form of a model update.

[0027] In one embodiment, the at least one measurement signal is acquired by at least one measurement sensor of the physical system. The at least one measurement signal can particularly preferably already be filtered and / or preprocessed and / or evaluated in order to be provided to the machine learning algorithm. This can, for example, simplify the extraction of the at least one characteristic feature, since the at least one measurement signal has already been filtered to remove signal noise.

[0028] In one embodiment, the measurement signal comprises a pressure measurement signal and / or a temperature measurement signal and / or an electrical measurement signal and / or an acceleration measurement signal and / or a speed measurement signal and / or a position measurement signal and / or another physical measurement signal. It is understood that all measurement signals of a physical system not explicitly listed here are also included without being mentioned separately.

[0029] In one embodiment, the state of the physical system is detected continuously and / or at predetermined intervals. The time steps can, for example, be specified by a respective sensor clock time of the measuring sensor used to acquire the at least one measurement signal. The detection model preferably runs continuously during operation of the physical system, so that with a high probability a state of the physical system is detected and / or measured and / or assumed by the system in which the detection model can successfully perform state detection. The present method for state detection thus particularly increases the certainty that state detection or state diagnosis occurs in every cycle of the physical system, thus always ensuring that the physical system is operating in an OK state.

[0030] In a particularly preferred embodiment, the physical system is a ventilation device for ventilating a crankcase. The at least one measurement signal then preferably comprises a pressure measurement signal detected by a pressure sensor of the ventilation device. In the faulty state of the ventilation device, it is then preferably detected that a hose of the ventilation device is incorrectly connected or has dropped out. Particularly preferably, the present method for detecting a system state can be used to model a pressure oscillation characteristic, in particular one that is dependent on the operating point. Machine learning is used to assist in this process. The model-based diagnosis of the ventilation device for ventilating a crankcase preferably detects the drop in the hose from the intake tract upstream of the turbocharger, in particular by means of a machine learning algorithm.As training data or

[0031] Input variables for training the machine learning algorithm are features of the at least one pressure measurement signal, preferably detected by a pressure sensor of the venting device. The at least one characteristic feature is, for example, an autocorrelation and / or a sum of the magnitudes of the signal changes of the at least one measurement signal and / or at least one mean value. The model preferably predicts, based on at least one of these pieces of information, whether the tubing has fallen off or not. The prediction is preferably made using the at least one error indicator, which indicates whether the venting device is in the OK or the nOK state.

[0032] The present method particularly preferably relates to the systematic measurement of a load and / or speed operating range of an internal combustion engine in which the hose or hose connection to be diagnosed is used for crankcase ventilation. To provide the training data, pressure signals are preferably recorded at various positions along the hose for both the OK state (hose connected) and the nOK state (hose connection open or disconnected). Sequences of a predetermined time duration, for example, from a few tens to a few hundred milliseconds, are preferably extracted from these measurements, and the at least one characteristic feature is determined and / or calculated from them.

[0033] At least one characteristic feature is required as an input variable for model training. For this purpose, an autocorrelation of the relative pressure between the oil tank and the hose relative to the combustion frequency can be used as a characteristic feature, preferably in the case of the venting device. Alternatively or additionally, the sum of the absolute changes in the relative pressure between the oil tank and the hose can be used as a characteristic feature. Furthermore, alternatively or additionally, the sum of the absolute changes in the pressure at the hose under investigation can be used as a characteristic feature.

[0034] The present invention also proposes a device for training a machine learning algorithm for detecting whether a physical system is in a fault-free state or a faulty state; the device comprising: a provision device configured to provide training data comprising at least one measurement signal relating to a measured variable of the physical system; and an evaluation and computing device configured to extract at least one characteristic feature from the at least one measurement signal; to determine a limit function based on the at least one characteristic feature and / or based on the at least one measurement signal by the machine learning algorithm;and to train the machine learning algorithm on the basis of the training data in order to detect, depending on the determined limit function, whether the physical system is in a fault-free state or a faulty state by determining at least one fault indicator;

[0035] The present invention also proposes a device for detecting whether a physical system is in a fault-free state or a faulty state, the device comprising: a provision device designed to provide the at least one measurement signal via the at least one measured variable of the physical system and the machine learning algorithm trained according to the method according to one of claims 1 to 8; an evaluation and computing device designed to extract at least one characteristic feature from the at least one measurement signal; to determine the at least one error indicator on the basis of the at least one measurement signal and / or on the basis of the at least one characteristic feature; and to detect, on the basis of the at least one error indicator, whether the physical system is in a fault-free state or a faulty state.

[0036] The present invention also proposes a control unit and / or an evaluation and computing device configured to execute a method according to one of the present aspects and / or embodiments, if the respective method is implemented in the form of program code on the control unit and / or the evaluation and computing device. Preferably, only the trained GPR model is stored on the control unit itself and executed for state detection. The detection model can preferably be trained on a control unit using a hardware acceleration unit. State detection can also be executed directly on the control unit. The detection model can, in principle, be divided into several components for deployment on the control unit.For example, a submodel can be implemented for predicting the distance of a measured value from the limit function, as well as a submodel for predicting the standard deviation. The models are called on the ECU, for example, by a wrapper function. Diagnostics preferably run continuously during operation of the physical system.

[0037] The present invention also claims a computer program with program code for executing at least parts of the method according to the invention in one of its embodiments when the computer program is executed on a computer. In other words, the invention relates to a computer program (product) comprising instructions that, when executed by a computer, cause the computer to execute the method / steps of the method according to the invention in one of its embodiments.

[0038] The present invention also proposes a computer-readable data carrier containing program code of a computer program for executing at least parts of the method according to the invention in one of its embodiments when the computer program is executed on a computer. In other words, the invention relates to a computer-readable (storage) medium comprising instructions which, when executed by a computer, cause the computer to execute the method / steps of the method according to the invention in one of its embodiments.

[0039] The described designs and further training courses can be combined as desired.

[0040] Further possible embodiments, developments and implementations of the invention also include combinations of features of the invention described previously or below with regard to the embodiments that are not explicitly mentioned.

[0041] Short description of the drawings

[0042] The accompanying drawings are intended to provide a further understanding of embodiments of the invention. They illustrate embodiments and, in the context of the description, serve to explain principles and concepts of the invention.

[0043] Other embodiments and many of the aforementioned advantages will become apparent upon review of the drawings. The elements illustrated in the drawings are not necessarily drawn to scale.

[0044] They show:

[0045] Fig. 1 is a schematic flow diagram of an embodiment of a training method;

[0046] Fig. 2 is a schematic flow diagram of an embodiment of a method for detecting a system state;

[0047] Fig. 3 is a schematic representation of an exemplary physical system; and

[0048] Fig. 4 is a schematic diagram for the detection of the

[0049] System state using the trained machine learning algorithm.

[0050] In the figures of the drawings, the same reference symbols designate the same or functionally equivalent elements, parts or components, unless otherwise stated.

[0051] Figure 1 shows a schematic flow diagram of a computer-implemented method for training a machine learning algorithm to detect whether a physical system is in a fault-free state OK or a faulty state nOK.

[0052] In any embodiment, the method can be carried out at least partially by a device 1, which for this purpose can comprise several components not shown in detail, for example one or more provision devices and / or at least one evaluation and computing device. It is understood that the provision device can be designed jointly with the evaluation and computing device or can be different from it. Furthermore, the system can comprise a memory device and / or an output device and / or a display device and / or an input device. The device 1 can be, for example, a control unit.

[0053] According to the invention, the computer-implemented method comprises at least the following steps:

[0054] In a step S1, training data is provided which includes at least one measurement signal about a measured variable of the physical system.

[0055] In a step S2, at least one characteristic feature is extracted from the at least one measurement signal.

[0056] In a step S3, a limit function is determined on the basis of the at least one characteristic feature and / or on the basis of the at least one measurement signal by the machine learning algorithm.

[0057] In a step S4, the machine learning algorithm is trained on the basis of the training data in order to detect, depending on the determined limit function, by determining at least one error indicator, whether the physical system is in a fault-free state OK or a faulty state nOK.

[0058] Figure 2 shows a schematic flow diagram of a computer-implemented method for detecting whether a physical system is in a fault-free state OK or a faulty state nOK by a trained machine learning algorithm.

[0059] In any embodiment, the detection method can be carried out at least partially by a device 1, which for this purpose can comprise several components not shown in detail, for example one or more provision devices and / or at least one evaluation and computing device. It is understood that the provision device can be designed jointly with the evaluation and computing device or can be different from it. Furthermore, the system can comprise a memory device and / or an output device and / or a display device and / or an input device. The device 1 can be, for example, a control unit.

[0060] In a step S10, the at least one measurement signal is provided via the at least one measured variable of the physical system.

[0061] In a step S20, at least one characteristic feature is extracted from the at least one measurement signal.

[0062] In a step S30, the at least one error indicator is determined on the basis of the at least one measurement signal and / or on the basis of the at least one characteristic feature.

[0063] In a step S40, a detection is carried out as to whether the physical system is in a fault-free state OK or a faulty state nOK, based on the at least one fault indicator.

[0064] Figure 3 shows an exemplary embodiment of a physical system 1000. The physical system comprises a ventilation device 100 for ventilating a crankcase 102. The at least one measurement signal is preferably a pressure measurement signal that is detected by at least one pressure measurement sensor 104, 105 of the ventilation device 100. In the faulty state nOK of the ventilation device 100, the detection model detects that a hose 106 of the ventilation device 100 is incorrectly connected or has dropped out. The pressure measurement sensor 104 is configured to measure and / or detect the pressure in the hose 106 to be tested in the form of at least one first pressure measurement signal. The physical system 1000 further comprises an internal combustion engine 108, which is connected to an oil tank 110.The additional pressure sensor 105 is configured to measure a pressure at the oil tank 110 in the form of a second pressure measurement signal. The physical system 1000 further includes an air filter 112, which is connected in the region of the hose system 106. The physical system 1000 includes a compressor 114 and a valve 116. The hose system 106 is connected upstream of the compressor 114. The valve 116 is arranged between the internal combustion engine 108 and the compressor 114. The physical system 1000 further comprises two check valves 118 and 120. The check valve 118 is arranged between the pressure sensor 104, the hose 106, and the oil tank 110 and, during throttled operation, blocks the flow of intake air from the intake system via the hose 106 toward the intake system via the check valve 120 or via the oil tank toward the crankcase of the internal combustion engine 108.The check valve 120 is arranged between a hose connection leading from the valve 116 to the internal combustion engine 108 and a hose connection leading from the check valve 118 to the oil tank 110. During supercharged operation, it blocks a backflow of intake air from the intake system via the check valve 118 via the hose connection 106 into the intake tract between the air filter 112 and the compressor 114, as well as the flow of intake air from the intake system via the oil tank 110 into the crankcase of the internal combustion engine 108. The condition of the hose connection 106, i.e., whether it is correctly connected or not, can be checked by the device 1 or the control unit. For this purpose, the detection model trained according to the invention or the trained machine learning algorithm DM is executed on the device 1. The trained detection model DM is preferably implemented in the form of program code on the device 1.

[0065] Figure 4 shows a schematic diagram for detecting the system state using the trained machine learning algorithm or the detection model. In the example shown, the detection model DM was trained based on three characteristic features generated from the at least one measurement signal. First, an autocorrelation of the at least one measurement signal was determined as a characteristic feature. The autocorrelation is plotted on a z-axis. A sum of magnitudes of at least one signal change within the at least one measurement signal generated by the pressure sensor 104 is plotted on a y-axis as a further characteristic feature.

[0066] The two features on the y- and z-axes refer to a signal that is generated by subtracting the measurement signals from pressure sensors 104 and 105. However, the use of different sensor values ​​is fundamentally a plausible and applicable solution.

[0067] As a further characteristic feature, a sum of the amounts of at least one signal change within the at least one measurement signal generated by the pressure measuring sensor 105 is plotted on an x-axis. The respective characteristic features are preferably each normalized to values ​​between 0 and 1. The diagram in Figure 4 shows a boundary surface 400 specified by the limit function determined according to the invention. The boundary surface 400 separates a value range in which measured values ​​indicate an OK state from a value range in which measured values ​​indicate a nOK state. Furthermore, a comfort zone 402 is shown in which the measured values ​​lie in a range that was recorded by the original training data. Measured values ​​orAssociated characteristic features that lie outside this comfort zone 402 cannot, for example, be used with sufficient accuracy to classify the system state into OK and nOK, since it must be assumed that the underlying characteristics were not included in the original training data. To include such "outliers", the detection model DM can, for example, be retrained, particularly based on the "outliers". The outliers are identified in Figure 4 with the reference numeral 404. A group of measured values ​​or associated characteristic features, by which an nOK state is indicated, is provided with the reference numeral 406 and is drawn to the right of the boundary surface 400. Furthermore, an exemplary distance d1 of one of these right-hand measured values ​​or features 406 from the boundary surface 400 is drawn. A group of measured values ​​orof associated characteristic features, by which an OK state is indicated, is provided with the reference symbol 410 and is drawn to the left of the interface 400. Furthermore, an exemplary distance d2 and an exemplary distance d3 of two of these left-side measured values ​​or features 410 from the interface 400 are drawn. Also drawn to the left of the interface 400 are validation data for validating the correct function of the detection model DM. This validation data is provided with the reference symbol 412. An exemplary distance from a measured value or feature of this validation data 412 is marked with d4.

Claims

Claims 1 . A computer-implemented method for training a machine learning algorithm to detect whether a physical system (1000) is in a fault-free state (OK) or a faulty state (nOK); the method comprising: Providing (S1) training data comprising at least one measurement signal about a measured variable of the physical system (1000); Extracting (S2) at least one characteristic feature from the at least one measurement signal; Determining (S3) a limit function based on the at least one characteristic feature and / or on the basis of the at least one measurement signal by the machine learning algorithm; and Training (S4) the machine learning algorithm on the basis of the training data in order to detect, depending on the determined limit function, by determining at least one error indicator, whether the physical system (1000) is in an error-free state (OK) or an error-affected state (nOK).

2. The method according to claim 1, wherein the trained machine learning algorithm comprises a support vector machine and a Gaussian process regressor, wherein the boundary function is determined by the support vector machine, and wherein the at least one error indicator is determined by the Gaussian process regressor as a function of the boundary function.

3. Method according to claim 1 or 2, wherein the at least one error indicator represents a magnitude distance of a respective measuring point of the at least one measuring signal from the limit function and / or a sign information of the respective measuring point of the at least one measuring signal related to the limit function and / or a standard deviation from a made distance prediction of the respective measuring point of the at least one measuring signal from the limit function.

4. Method according to one of the preceding claims, wherein the at least one characteristic feature comprises at least one oscillation characteristic and / or at least one autocorrelation of the at least one measurement signal and / or a sum of amounts of at least one signal change within the at least one measurement signal and / or a mean value of the at least one measurement signal and / or a standard deviation.

5. The method according to any one of the preceding claims, wherein the at least one measurement signal comprises a time-series-based measurement signal, wherein the machine learning algorithm is trained to detect the state of the physical system based on a predetermined time window of the time-series-based measurement signal.

6. Method according to one of the preceding claims, wherein the machine learning algorithm can be retrained in the event of a change in the physical system and / or in the event of a change in the at least one measurement signal.

7. Method according to one of the preceding claims, wherein the at least one measurement signal is detected by at least one measurement sensor of the physical system and is preferably filtered and / or preprocessed and / or evaluated in order to be provided by the machine learning algorithm.

8. Method according to one of the preceding claims, wherein the measurement signal is a pressure measurement signal and / or a temperature measurement signal and / or an electrical measurement signal and / or an acceleration measurement signal and / or a speed measurement signal and / or a position measuring signal and / or another physical measuring signal.

9. A computer-implemented method for detecting whether a physical system (1000) is in a fault-free state (OK) or a faulty state (nOK) by a machine learning algorithm trained according to the method of any one of claims 1 to 8; the method comprising: Providing (S10) the at least one measurement signal via the at least one measured variable of the physical system; Extracting (S20) at least one characteristic feature from the at least one measurement signal; Determining (S30) the at least one error indicator on the basis of the at least one measurement signal and / or on the basis of the at least one characteristic feature; and Detecting (S40) whether the physical system is in a fault-free state (OK) or a faulty state (nOK) based on the at least one fault indicator.

10. The method according to claim 9, wherein the detection of the state of the physical system occurs continuously and / or in predetermined periods.

11. Method according to one of the preceding claims, wherein the physical system (1000) is a ventilation device for ventilating a crankcase, wherein the at least one measurement signal comprises a pressure measurement signal detected by a pressure measurement sensor of the ventilation device, and wherein in the faulty state (nOK) of the ventilation device it is detected that a hose of the ventilation device is incorrectly connected or has fallen off.

12. Control device and / or evaluation and computing device which is designed to carry out a method according to one of the preceding claims to be carried out if the method according to one of the preceding claims is implemented in the form of program code.

13. Device (1) for detecting whether a physical system is in a fault-free state (OK) or a faulty state (nOK), the device comprising: a provision device which is designed to provide the at least one measurement signal via the at least one measured variable of the physical system and the machine learning algorithm trained according to the method according to one of claims 1 to 8; an evaluation and computing device which is designed to extract at least one characteristic feature from the at least one measurement signal; to determine the at least one fault indicator on the basis of the at least one measurement signal and / or on the basis of the at least one characteristic feature; and to detect, on the basis of the at least one fault indicator, whether the physical system is in a fault-free state (OK) or a faulty state (nOK).

14. A computer program comprising program code for executing at least parts of a method according to any one of claims 1 to 11 when the computer program is executed on a computer.

15. A computer-readable data carrier with program code of a computer program for carrying out at least parts of a method according to one of claims 1 to 11 when the computer program is executed on a computer.

Images