Method and device for recognizing a manipulation at a technical device in a motor vehicle

Abstract

Method and device for manipulation recognition at a technical device in a motor vehicle by means of an artificial intelligence method. The invention relates to a method for manipulation recognition of a technical device (1), in particular in a motor vehicle, in particular an exhaust gas aftertreatment device, having the following steps: - providing (S1) a time series of input vectors (E) having one or more system variables (G) and having at least one setting variable (S) for influencing the technical device (1); - using (S2, S3) a data-based manipulation recognition model, which includes a recurrent neural network (61) configured to determine a state vector (Z) from the input vectors (E) and an autoencoder (62) configured to determine a reconstruction vector (E') from the state vector (Z); - recognizing (S4, S5) an anomaly from a reconstruction error depending on the reconstruction vector (E'); - recognizing (S8) a manipulation from the reconstruction error.

Description

TECHNICAL FIELD

[0001] The present application relates to a motor vehicle and, in particular, to a method for manipulation recognition of a device of a motor vehicle. The present application also relates to an exhaust gas aftertreatment device and a method for recognizing a manipulation and for diagnosing an exhaust gas aftertreatment device. BACKGROUND

[0002] Technical devices in a motor vehicle can be manipulated in an impermissible manner in order to achieve a favorable operation for the driver. In this way, an exhaust gas aftertreatment device can be manipulated for a power increase of the engine system or for a reduction of material consumption, in particular of urea.

[0003] In general, methods for recognizing a manipulation are based on rules. Rule-based manipulation monitoring methods have the disadvantage that only known manipulation strategies can be recognized or only known manipulations can be intercepted. This defense strategy is therefore not discriminative for new types of manipulation. Furthermore, it is costly to detect complex technical systems together with their dependencies in the regulation system and to create corresponding rules for recognizing a manipulation.

[0004] For example, for an exhaust gas aftertreatment device, the operating state is manifold due to its dynamic behavior and cannot be unambiguously linked to the presence of a manipulation, in particular in the case of rarely occurring system states. For example, today's SCR exhaust gas aftertreatment systems for deoxidation (reduction of nitrogen by injection of urea into the exhaust gas) have a legally prescribed monitoring of system parameters which are important for error-free operation. These system parameters are monitored within the framework of on-board diagnostics whether they follow physically reasonable limit values and are thus checked for plausibility. For system-inherent parameters whose values result from the combination of different setting quantities of the SCR regulation, it can also be checked whether the expected system reaction occurs after a system intervention. Thus, for example, after an increase in the urea dose under defined conditions, a reduction in the nitrogen oxide emissions is expected to be measured by a nitrogen oxide sensor arranged downstream.

[0005] However, so-called SCR emulators are increasingly used, which are able to change data in the program code of the monitoring system or of the sensor values used by the monitoring system, so that even if the SCR system is only marginally active or completely inactive, a false recognition is excluded by the monitoring system. Thereby, in vehicle operation, maintenance costs can be reduced and the costs for refilling urea can be saved in the case of a tolerated increase in nitrogen oxide emissions. The conventional diagnostic functions are deceived by the simulated sensor signals, which makes the recognition of a manipulation difficult. SUMMARY

[0006] According to the application, a method for manipulation recognition in a technical device, in particular in a technical device in a motor vehicle, in particular in an exhaust gas aftertreatment device, and a corresponding device and technical system are specified.

[0007] Further design options are specified in the application.

[0008] According to a first aspect, a method for manipulation recognition in a technical device, in particular in a technical device in a motor vehicle, in particular in an exhaust gas aftertreatment device, is specified, which method has the following steps: providing a time series of input vectors, which input vectors have one or more system variables and have at least one setting variable for influencing the technical device; using a data-based manipulation recognition model, which manipulation recognition model comprises a recurrent neural network, which recurrent neural network is configured to determine a state vector from an input vector, and an autoencoder, which autoencoder is configured to determine a reconstruction vector from a state vector; recognizing an anomaly from a reconstruction error which depends on the reconstruction vector; recognizing a manipulation from the reconstruction error. In particular, the application specifies a method for manipulation recognition in an exhaust gas aftertreatment device in a motor vehicle, which method has the following steps: providing a time series of input vectors, which input vectors have one or more system variables and have at least one setting variable for a urea injection system of the exhaust gas aftertreatment device; providing the acquired time series of input vectors to a data-based manipulation recognition model, which manipulation recognition model comprises a recurrent neural network, which recurrent neural network is configured to determine each respective state vector for a respective consecutive time step of the time series from the input vector, and an autoencoder, which autoencoder is configured to determine a respective reconstruction vector for each respective consecutive time step of the time series from the respective state vector of the respective consecutive time step, wherein the recurrent neural network is trained for normal operating conditions of the exhaust gas aftertreatment device; determining a reconstruction error for each respective consecutive time step of the time series from the respective reconstruction vector of the respective consecutive time step and the input vector in the respective consecutive time step; recognizing a respective anomaly in the behavior of the exhaust gas aftertreatment device from the reconstruction error; counting the recognized respective anomalies using an anomaly counter; determining a proportion from the content of the anomaly counter, which proportion states how many time steps of all considered time steps are recognized as anomalies; comparing the proportion to a pre-specified proportion threshold; and recognizing a manipulation of the exhaust gas aftertreatment device from the result of the comparison.

[0009] It can be specified that the technical device comprises an exhaust gas aftertreatment device, wherein the input vectors comprise setting variables for a urea injection system as the setting variables.

[0010] According to the above-described method, it is proposed to use a machine learning method in order to carry out a manipulation recognition of a technical device in a motor vehicle. By means of a data-based manipulation recognition model, the normal behavior of the underlying technical device is taught and deviations from its normal behavior are considered as a result of a manipulation.

[0011] By means of a deep learning method, dependencies and characteristics of the technical device that are important for the underlying manipulation recognition can be recognized independently. Since the normal behavior of the technical device while it is running is trained in the manipulation recognition model, a deviation from its normal behavior of the technical device can be recognized. This has the advantage that new and so far unknown manipulation attempts can also be recognized by this manipulation recognition model.

[0012] The above-described method provides a manipulation recognition in which a recurrent network (RNN) is used in combination with an autoencoder structure. In principle, the RNN autoencoder structure is trained to use a time series of input vectors and to process this time series into a state vector in a RNN, such as a LSTM (Long Short-Term Memory). The input vector comprises for a time step one or more system variables, at least one manipulated variable for intervening in the technical device and optionally at least one system parameter. The time series of input vectors corresponds to a sequence of input vectors at consecutive time steps in which the described method is implemented periodically. Thus, the time series of input vectors corresponds to a time series of the one or more system variables, the at least one manipulated variable for intervening in the technical device and optionally the at least one system parameter.

[0013] The state vector maps the behavior of the technical device, that is to say, in the trained RNN network, the time series of input vectors is processed into a state vector. Thereby, the state vector depends on the time series of input vectors.

[0014] It can be provided that the state vector for each time step is determined from the input vector by means of a recurrent neural network, wherein an autoencoder is trained to map the state vector onto the input vector, wherein a reconstruction error describes the difference between the input vector and the reconstructed input vector by the autoencoder.

[0015] Thus, by means of the recurrent network, a state vector is formed from the time series of input vectors at each time step, which can preferably have a higher dimension than the input vectors. Then, a conventional autoencoder is used based on a deep neural network in order to recognize anomalies in the time series of these input vectors. The autoencoder is trained to map the state vector from the recurrent network in each time step onto the input vector of the same time step.

[0016] In order to identify anomalies during the operation of the technical device, the reconstruction error of the autoencoder is evaluated. Here, the deviation between the input vector and the reconstructed input vector is evaluated. In particular, the reconstruction error can be determined by means of a predefined error function, which is based, inter alia, on the mean squared error between the original input vector and the reconstructed vector, a Huber loss function or a root-mean squared error.

[0017] The time steps for which the behavior of the technical device is evaluated can be predefined. The behavior is classified as abnormal or normal behavior, respectively, depending on the reconstruction error. Depending on the percentage of anomalies identified during the time steps, a manipulation of the technical device can be identified.

[0018] It can be provided that the reconstruction error is determined depending on the difference between the input vector and the reconstructed input vector in the time steps in order to state whether there is an anomaly in each time step.

[0019] According to one embodiment, the reconstruction error can be determined depending on the reconstructed vector in consecutive time steps in order to find out whether there is an anomaly for each time step, wherein a manipulation of the technical device or an abnormal behavior of the technical device is identified or signaled if the proportion of anomalies during the time steps exceeds a predefined proportion threshold.

[0020] It is also possible to determine a state vector for each time step from the input vector by means of a recurrent network, wherein the autoencoder is trained to map the state vector to itself, wherein the reconstruction error states the difference between the state vector and the reconstructed state vector by the autoencoder.

[0021] It can be provided that the anomalies in the time steps are identified by predefining an individual error threshold for each of the one or more system variables and the at least one manipulated variable, wherein an anomaly in the time steps is identified if a predefined number of the one or more system variables and the at least one manipulated variable exceeds the respective individual error threshold.

[0022] In particular, the individual error thresholds can be determined by means of a validation data set, wherein the error threshold for each of the one or more system variables and the at least one manipulated variable is chosen such that no error is identified corresponding to a predefined percentile.

[0023] According to one embodiment, the identified manipulation can be signaled or the technical device can be operated depending on the identified manipulation.

[0024] According to another aspect, a device for manipulation recognition of a technical installation, in particular a technical installation in a motor vehicle, in particular an exhaust gas aftertreatment installation, is specified, wherein the device is configured to provide a time series of input vectors, the input vectors having one or more system variables and having at least one setting variable for an intervention into the technical installation, to use a data-based manipulation recognition model, the manipulation recognition model comprising a recurrent neural network configured to determine state vectors from the input vectors and an autoencoder configured to determine reconstruction vectors from the state vectors, to identify an anomaly from a reconstruction error depending on the reconstruction vectors, and to identify a manipulation from the reconstruction error. In particular, the invention provides a device for manipulation recognition of an exhaust gas aftertreatment installation in a motor vehicle, wherein the device is configured to provide a time series of input vectors, the input vectors having one or more system variables and having at least one setting variable for a urea injection system of the exhaust gas aftertreatment installation, to provide the acquired time series of input vectors to a data-based manipulation recognition model, the manipulation recognition model comprising a recurrent neural network configured to determine for each respective state vector of a respective consecutive time step of the time series from the input vectors and an autoencoder configured to determine for each respective consecutive time step of the time series a respective reconstruction vector from the respective state vector of the respective consecutive time step, wherein the recurrent neural network is trained for normal operating conditions of the exhaust gas aftertreatment installation, to determine for each respective consecutive time step of the time series a reconstruction error from the respective reconstruction vector of the respective consecutive time step and the input vector in the respective consecutive time step, to identify a respective anomaly of a behavior of the exhaust gas aftertreatment installation from the reconstruction error, to count the identified respective anomalies using an anomaly counter, to determine a proportion from the content of the anomaly counter, which proportion states how many time steps of all considered time steps are identified as anomalies, to compare the proportion to a pre-given proportion threshold, and to identify a manipulation into the exhaust gas aftertreatment installation from a result of the comparison. In addition, the invention provides a computer program product comprising instructions which, when the program is implemented by a computer, cause the computer to implement the steps of the inventive method. A machine-readable storage medium of the invention comprises instructions which, when implemented by a computer, cause the computer to implement the steps of the inventive method. BRIEF DESCRIPTION OF DRAWINGS

[0025] The embodiments are hereinafter described in further detail with reference to the accompanying drawings. In which:

[0026] Figure 1A schematic diagram of an exhaust gas aftertreatment device as an example of a technical system is shown;

[0027] Figure 2 A flow chart of a method for illustrating manipulation recognition for an exhaust gas aftertreatment device is shown; Figure 1 A flow chart of a method for illustrating manipulation recognition for an exhaust gas aftertreatment device is shown; and

[0028] Figure 3 A schematic diagram of a network structure of an autoencoder for use in the case of manipulation recognition, based on an evaluation of a time series of an input vector, is shown. DETAILED DESCRIPTION

[0029] Figure 1 A schematic diagram of an exhaust gas aftertreatment system 2 for an engine system 1 with an internal combustion engine 3 is shown. The exhaust gas aftertreatment device 2 is designed for exhaust gas aftertreatment of the combustion exhaust gas of the internal combustion engine 3. The internal combustion engine 3 can be configured as a diesel engine.

[0030] The exhaust gas aftertreatment device 2 has a particulate filter 21 and an SCR catalytic converter 22. The exhaust gas temperature is measured upstream of the particulate filter 21, downstream of the particulate filter 21 and downstream of the SCR catalytic converter 22 with corresponding temperature sensors 23, 24, 25, and the NOx content is measured upstream and downstream of the SCR catalytic converter 22 with corresponding NOx sensors 26, 27. The exhaust gas temperature and the NOx content are processed in a control unit 4. The sensor signals are provided as system variables G to the control unit. x x The exhaust gas temperature and the NOx content are processed in a control unit 4. The sensor signals are provided as system variables G to the control unit.

[0031] A urea reservoir 51, a urea pump 52 and a controllable injection system 53 for urea are provided. The injection system 53 enables a predetermined amount of urea to be delivered into the combustion exhaust gas upstream of the SCR catalytic converter 22 by means of a set quantity S controllably by the control unit 4.

[0032] The control unit 4 controls the delivery of urea upstream of the SCR catalytic converter 22 by specifying a set quantity for the injection system 53 in accordance with known methods in order to achieve the best possible catalytic purification of the combustion exhaust gas, so that the nitrogen oxide content is reduced as much as possible.

[0033] Conventional manipulation devices manipulate the sensor signals and / or the regulation signals in order to reduce or completely stop the consumption of urea.

[0034] While such manipulations can be recognized by monitoring the operating state of the exhaust gas aftertreatment device on the basis of rules, not all corresponding impermissible operating states can be checked in this way. Thus, a manipulation recognition method based on a manipulation recognition model is proposed. This can be implemented in the control unit 4, as follows in accordance with​Figure 2 The method is illustrated as an example in the flowchart. The method can be implemented in the control unit 4 as software and / or hardware.

[0035] In step S1, a time series of an input vector consisting of the system variables G and the setpoint value S for the injection system 53 of urea is detected. The system variables S can include one or more of the following variables: the exhaust gas temperature; the NOx concentration; the current engine torque; the current charge of the internal combustion engine 3; the rotational speed of the internal combustion engine 3; the amount of fuel injected into the internal combustion engine 3; the pressure in the exhaust gas system; the NH3 concentration; the oxygen concentration in the combustion exhaust gas; the DeNOx efficiency (determined from the NOx concentration before and after the SCR catalytic converter); the engine temperature; the driver's desired torque, for example pre-given by the accelerator pedal position; the vehicle speed; the ambient pressure; the ambient temperature; the selected gear of the gear shift; the vehicle weight; the position of the exhaust gas recirculation valve; and the soot quantity in the combustion exhaust gas.

[0036] The time series of the input vector (consisting of the system variables G and the setpoint value S) is evaluated by means of the manipulation recognition model in order to recognize whether the behavior of the exhaust gas aftertreatment device 2 is abnormal. To this end, the time series of the system variables G and the setpoint value S is fed to the manipulation recognition model in step S2, as is illustrated exemplarily in Figure 3

[0037] The manipulation recognition model 60 has a recurrent neural network 61 in order to assign a state vector Z to the time series of the system variables G and the setpoint value S in a manner known per se. The recurrent neural network 61 can be, for example, an LSTM (Long Short-Term Memory), a GRU unit (Gated Recurrent Unit) or any other recurrent structure that can map temporal dynamics. The recurrent neural networks have in common that they produce a state vector (hidden state) in each time step of the time series of the input vector.

[0038] The dimension of the state vector Z of the recurrent neural network 61 can be chosen arbitrarily, however it has turned out that the reliability of the anomaly recognition is better if the state vector Z has a higher dimension than the input vector consisting of the system variables G and the setpoint value S.

[0039] The recurrent neural network 61 is trained for normal operating conditions of the exhaust gas aftertreatment device 2. The state vector Z thus characterizes the behavior of the exhaust gas aftertreatment device 2. The recurrent neural network 61 is trained, in particular, to map the time series of the system variables G and the setpoint value S onto the state vector Z, which ideally does not change independently of the dynamic behavior of the exhaust gas aftertreatment device 2.​

[0040] In step S3, the state vector Z is delivered as a result of the autoencoder 62 with the encoder part 63 and the decoder part 64. The encoder part 63 and the decoder part 64 can be constituted as a deep neural network consisting of fully connected layers. The activation function can likewise be freely chosen, such as Leaky-ReLU or the like.

[0041] The encoder part 63 of the autoencoder 62 compresses the state vector Z into a feature vector M of reduced dimension (compared to the state vector Z), and the decoder part 64 expands the feature vector M into the original input vector of the same time step.

[0042] In step S4, in the anomaly recognition unit 65, a reconstruction error R between the reconstructed input vector E' at the output of the decoder part 64 and the input vector E at the input of the recurrent neural network 61 is determined. The output layer of the decoder part 64 should have the size of the input vector E in order to reconstruct the respective input vector E'.

[0043] The reconstruction error for recognizing an anomaly of the behavior of the exhaust gas aftertreatment device 2 can be calculated in the anomaly recognition unit 65 signal-wise, that is to say for each time step. For example, the reconstruction error can be determined as the L2 norm.

[0044] Now, in the evaluation unit 66, it can be ascertained in step S5 whether an anomaly exists depending on a comparison of the reconstruction error with an error threshold. The error threshold can be fixedly predefined or individually determined signal-wise, that is to say for each system variable G or setting S (input vector variable), as a percentile value. The percentile value corresponds to a parameter which can be determined in a manner known per se by means of a validation data set. The validation data set is a time series of input vectors which map a normal operation of the exhaust gas aftertreatment device 1.

[0045] The validation data set is used to determine (signal-wise) the reconstruction error between the input vector E and its reconstruction. Here, the error threshold for each input vector variable is chosen such that no error is recognized corresponding to a predefined percentile. For example, a percentile of between 99.9% and 99.99% can be predefined for each input vector variable. From this, a percentile value is obtained for each of these input vector variables, respectively.

[0046] If the reconstruction error in the case of at least one of the input vector quantities exceeds the percentile value, an anomaly in the time step concerned is identified. To this end, in step S6, a time step counter can be incremented for each time step and an anomaly counter can be incremented when an anomaly is identified.

[0047] In operation, the proportion / percentage can be determined from the counter content, which states how many of the time steps in which the input vectors are identified as anomalies with regard to all time steps considered. By means of a threshold comparison of the predefined proportion with a proportion threshold in step S7, it can be ascertained what percentage of the journey since the start of operation has been identified as an anomaly. If this proportion value exceeds the predefined proportion threshold (alternative: yes), it can be concluded that there is a manipulation attempt and this is reported accordingly in step S8. Otherwise (alternative: no), the method continues with step S1.

[0048] The manipulation identification model can be trained in a manner known per se during a plurality of epochs. Here, all training data are processed within each epoch. The training data correspond to time series of system variables and setting variables which are recorded in the manipulation-proof operating environment of the exhaust gas aftertreatment device 2. The time series of input vectors (system variables G and setting variables S) of these training data can comprise values of, for example, 500 to 3000 time steps. For training the manipulation identification model, the recurrent neural network 61 and the autoencoder 62 can be trained such that the input vector E for each time step corresponds to the reconstructed input vector E' obtained on the output side of the autoencoder. The error function for training the manipulation identification model can then use the mean squared error or the root mean squared error between the reconstructed input vector and the original input vector.

[0049] In an alternative embodiment, the autoencoder 62 can be constructed or trained for mapping the state vector Z for each time step onto the state vector Z comprising the system variables G and the setting variables S.

[0050] In this case, a reconstruction error between the reconstructed state vector and the original state vector Z can be determined.

[0051] For determining the reconstruction error, a corresponding error function can be determined on the basis of the mean squared error, the Huber loss function or the root mean squared error between the original input vector and the reconstructed input vector. The determined error value can be propagated by means of backpropagation for training the network parameters, wherein a gradient descent method common to neural networks, such as SGD, Adam, AdamW or AdaGrad, can be used as optimization strategy.

Claims

1. A method for manipulation recognition of an exhaust gas aftertreatment device in a motor vehicle, the method having the following steps: - providing (SI) a time series of input vectors (E) having one or more system variables (G) and having at least one setpoint value (S) for a urea injection system of the exhaust gas aftertreatment device (1); - providing the acquired time series of input vectors to a data-based manipulation recognition model, the manipulation recognition model comprising a recurrent neural network (61) configured to determine for the time series a respective continuous time step a respective state vector (Z) from the input vectors (E), and an autoencoder (62) configured to determine for each respective continuous time step of the time series a respective reconstruction vector (E') from the respective state vector (Z) of the respective continuous time step, wherein the recurrent neural network (61) is trained for normal operating conditions of the exhaust gas aftertreatment device (1); - determining for each respective continuous time step of the time series a reconstruction error from the respective reconstruction vector (E') of the respective continuous time step and the input vector (E) in the respective continuous time step; - recognizing (S4, S5) a respective anomaly of the behavior of the exhaust gas aftertreatment device (1) from the reconstruction error; - counting the recognized respective anomalies using an anomaly counter; - determining from the content of the anomaly counter a proportion which states how many time steps of all considered time steps are recognized as anomalies; - comparing the proportion to a pre-given proportion threshold; and - recognizing (S8) a manipulation of the exhaust gas aftertreatment device (1) from the result of the comparison.

2. The method according to claim 1, wherein the reconstruction error states a difference between the input vector (E) and the reconstructed input vector (E') by the autoencoder (62).

3. The method according to claim 1, wherein the reconstruction error is determined by means of a pre-given error function, the error function being based on a mean squared error, a Huber loss function or a root mean squared error between the original input vector and the reconstruction vector.

4. The method according to any one of claims 1 to 3, wherein an anomaly in a time step is recognized by pre-giving a separate error threshold for each of the one or more system variables (G) and the at least one setpoint value (S), wherein an anomaly in a time step is recognized if a pre-determined number of the one or more system variables (G) and the at least one setpoint value (S) exceeds the respective separate error threshold. ​ ​ ​ ​ ​ ​ ​ ​ ​ 5. The method according to claim 4, wherein the individual error threshold is determined by means of a validation data set, wherein the error threshold for each of the one or more system variables (G) and the at least one setting quantity (S) is chosen such that no error is identified corresponding to a pre-given percentile.

6. The method according to any one of claims 1 to 3, wherein the identified maneuver is signaled, or wherein the exhaust gas aftertreatment device (1) is operated in accordance with the identified maneuver.

7. A device for maneuver identification of an exhaust gas aftertreatment device in a motor vehicle, wherein the device is configured for: - providing (S1) a time series of input vectors (E) having one or more system variables (G) and having at least one setting quantity (S) for a urea injection system of the exhaust gas aftertreatment device (1); - providing the acquired time series of input vectors to a data-based maneuver identification model, the maneuver identification model comprising a recurrent neural network (61) configured for determining from the input vectors (E) each respective state vector (Z) for a respective consecutive time step of the time series and an autoencoder (62) configured for determining from the respective state vectors (Z) of the respective consecutive time steps each respective reconstruction vector (E’) for a respective consecutive time step of the time series, wherein the recurrent neural network (61) is trained for normal operating conditions of the exhaust gas aftertreatment device (1); - determining for each respective consecutive time step of the time series a reconstruction error from the respective reconstruction vector (E’) of the respective consecutive time step and the input vector (E) in the respective consecutive time step; - identifying (S4, S5) respective abnormalities of behavior of the exhaust gas aftertreatment device (1) from the reconstruction errors; - counting the identified respective abnormalities using an abnormality counter; - determining from the content of the abnormality counter a proportion which states how many time steps of all considered time steps are identified as abnormal; - comparing the proportion to a pre-given proportion threshold; and - identifying (S8) a maneuver to the exhaust gas aftertreatment device (1) in dependence on the result of the comparison.

8. A computer program product comprising instructions which, when the program is implemented by a computer, cause the computer to carry out the steps of the method according to any one of claims 1 to 6.

9. A machine-readable storage medium comprising instructions which, when implemented by a computer, cause the computer to carry out the steps of the method according to any one of claims 1 to 6. ​ ​ ​ ​ ​ ​ ​ ​ ​

Images