(19)
(11)EP 3 654 246 A1

(12)EUROPEAN PATENT APPLICATION

(43)Date of publication:
20.05.2020 Bulletin 2020/21

(21)Application number: 18206154.9

(22)Date of filing:  14.11.2018
(51)Int. Cl.: 
G06N 3/04  (2006.01)
G06N 3/08  (2006.01)
(84)Designated Contracting States:
AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR
Designated Extension States:
BA ME
Designated Validation States:
KH MA MD TN

(71)Applicant: Bayerische Motoren Werke Aktiengesellschaft
80809 München (DE)

(72)Inventors:
  • Notz, Dominik
    65843 Sulzbach (DE)
  • Zhu, Yi
    Elmhurst, New York 11373, (US)

  


(54)METHOD, VEHICLE, SYSTEM, AND STORAGE MEDIUM FOR INDICATING ANOMALOUS VEHICLE SCENARIO USING ENCODER NETWORK AND REPLAY BUFFER


(57) Providing an evaluation machine comprising a discriminator network (D), a generator network (G), and an encoder network (E), it is an object of the present application to generate a signal (160) indicating anomalous vehicle scenario. The object is achieved by a method comprising calibration steps:
a) storing noise signals and output signals of the generator network (G) in a replay buffer (260);
b) receiving previous generator output signals (Grb) and previous noise signals (zrb) from the replay buffer (260);
c) updating the discriminator network (D), the encoder network (E), and the generator network (G) using:
- training signals (x) from vehicle sensors (120) and an output signals of the encoder network (E);
- the previous generator output signals (Grb) and the previous noise signals (Zrb);
- noise signals (z) and an output signals of the generator network (G);

the method further comprising detection steps:
d) generating an anomaly score from an output signal of the encoder network (E), an output signal of the generator network (G), and/or an output signal of the discriminator network (D), wherein an input signal to the encoder network (E) includes data from vehicle sensors (120);
e) comparing the anomaly score with a pre-set threshold; and
f) generating a signal (160) indicating anomalous scenario depending on the comparison, preferably if the anomaly score is above the pre-set threshold.




Description


[0001] The present application relates to a method for generating a signal indicating anomalous scenarios using an evaluation machine comprising a discriminator network, a generator network, and an encoder network, calibrated using a replay buffer.

[0002] US patent application publication US 2016/0272113 A1 relates to a monitoring system aimed at determining that a vehicle is not being operated by a driver. Based on the vehicle not being operated, the monitoring system can analyse data from one or more sensors of the vehicle to monitor a passenger compartment of the vehicle. With data from the one or more sensors, the monitoring system can detect an anomaly within the passenger compartment of the vehicle. Based, at least in part, on detecting the anomaly, the monitoring system can transmit a signal corresponding to an alert to a computing device of the driver, the alert indicating the anomaly.

[0003] It is an objective of the present invention to improve the safety of vehicles operations. A further objective is to indicate a potential threat to the safety of vehicles operations. A further objective is to reduce the cost of vehicle manufacturing. A further objective is to reduce storage requirements for vehicle scenario information. A further objective is to improve the cautiousness of autonomous vehicle behaviour.

[0004] The objectives of the present invention are solved by the subject matter of claims 1, 9, and 10.

[0005] In particular, the object of the present invention is solved by a method for generating a signal indicating anomalous vehicle scenario using an evaluation machine comprising a discriminator network, a generator network, and an encoder network.

[0006] The method may comprise the calibration (or training) steps (also known as networks "training"): storing noise signals and output signals of the generator network in a replay buffer, receiving previous generator output signals and previous noise signals from the replay buffer, updating (i.e. training) the discriminator network, the encoder network, and the generator network using:
  1. a) training data from vehicle sensors and an output signals of the encoder network,
  2. b) the previous generator output signals and the previous noise signals, and
  3. c) noise signals and an output signals of the generator network.


[0007] The method further comprise a detection (or inference) steps: generating an anomaly score from an output signal of the encoder network, an output signal of the generator network, and/or an output signal of the discriminator network, wherein an input signal to the encoder network includes data from vehicle sensors, comparing the anomaly score with a pre-set threshold, and generating a signal indicating anomalous scenario based on the comparison, such as if the generated anomaly score is above and/or below the pre-set threshold.

[0008] The method may improve the safety of vehicles operations and indicate potential threats by generating a signal indicating anomalous scenarios. The method may also reduce the cost of vehicle manufacturing by generating a signal indicating anomalous scenarios and focus development resources on anomalous scenarios rather than on normal scenarios. The method may also improve the cautiousness of autonomous vehicle operation by removing from autonomous vehicle training scenarios, scenarios including vehicle controlled in an irregular way.

[0009] By storing a batch of noise signals and output signals of the generator network in a replay buffer, and receiving previous generator output signals and previous noise signals from the replay buffer, overfitting the discriminator network is decreased or prevented, resulting in an improved accuracy in the signal indicating an anomalous vehicle scenario, such as improved sensitivity and specificity. This reduces the misclassifications in the output signal.

[0010] By using random noise signals for training the evaluation machine, no prior manual classification of the input signal have to be made, allowing unsupervised calibration of the networks. This improves the efficiency of calibration, especially with large numbers of training signals, and allows automated calibration running in real-time, or almost real-time, with low latency. The noise signals may be generated by the noise generator.

[0011] One problem of the training discriminator generator, encoder networks is that the generator may over-fit to the current discriminator network and the discriminator may over-fit to the current generator network. The replay buffer may store noise signals and output signals of the generator network and provide previous generator output signals and previous noise signals. Current noise signals and output from current generator network may be mixed with previous generator output signals and previous noise signals for training discriminator network. The mix may be one to one so that equal amount of current signals and previous signals are used. The mix may also have any other ratio. The previous signals may be from calibrations steps occurring a previous execution of the calibration steps, it may be of earlier execution of the calibration steps. The replay buffer may also provide a mix of signals of several earlier executions of the calibration steps.

[0012] The calibration steps may be carried out iteratively a plurality of times and the discriminator, generator and encoder networks preferably updated during each iteration (such as towards the end of each iteration). The steps may further be repeated until the generator, encoder, and/or discriminator are sufficiently calibrated, e.g. anomalous scenarios are sufficiently detected. By repeating the calibration steps a plurality of timed, an indication of anomalous vehicle scenario may be signalled with improved sensitivity. During each calibration cycle, some of the newly generated images will be stored in the replay buffer. In each training cycle, the signals for the training of the discriminator will be a combination of newly generated signals and old ones from the replay buffer.

[0013] The output signal of the discriminator network may indicate a confidence of the discriminator network to categorize (distinguish between) a first category of signals comprising sensor signals from vehicle sensors and output signals of the encoder network, and a second category of signals comprising noise signals and output signals of the generator network.

[0014] Updating the discriminator network may include affecting (increasing or decreasing) a networks loss value in a first direction (such as increasing and/or maximizing) by calibrating the discriminator network to increase its confidence in distinguishing the two categories. Updating the encoder network and the generator network may include affecting the networks loss value in a second direction (such as decreasing and/or minimizing the network loss value), by calibrating the encoder network and the generator network so that the confidence of the discriminator network to categorize (distinguish between) the first category of signals and the second category of signals is reduced.

[0015] The networks loss value (objective) may be generated via the function (value function V):

where



[0016] The discriminator network may indicate the confidence with a probability estimate, reaching from with a first value (such as one) when the discriminator network indicates an input signal of the discriminator as belonging to one of the first category and the second category, and to a second value (such as zero) when the discriminator network indicates another input signal to the discriminator as clearly belonging to a another of the first category and the second category. The networks loss value may be maximized when the discriminator network correctly categorizes the first category and the second category of signals with high confidence.

[0017] The first category may be (x,z) pair where x comes from real data and z is generated from the encoder network with input x. The second category may be (x,z) pair where z may be a random vector (such as a Gaussian random vector) and x may be generated from the generator network with input z.

[0018] The output signal of the discriminator network may indicate a value approaching a third value (such as one half) that the discriminator network substantially cannot categorize an input signal as belonging either to the first category, or the second category. An output from the discriminator approaching the third value may indicate a low confidence of the discriminator network in categorizing the input signal in either the first category or in the second category. A discriminator network output signal approaching the third value may minimize the networks loss value.

[0019] The calibrations steps may be carried out until a Nash equilibrium of the network loss value is essentially achieved at least locally. When the Nash equilibrium is essentially achieved, the generator network and the encoder network may be calibrated so that the discriminator network substantially cannot distinguish between the first category of signals and the second category of signals. The output signal of the discriminator network may approach the third value, such as one half.

[0020] Updating the discriminator network, the generator network, and the encoder network by affecting a networks loss value in a first and second directions respectively improves the evaluation machine's sensitivity to anomalous vehicle scenarios.

[0021] Affecting the networks loss value (in a first and/or second direction) may include taking a stochastic gradient decent/ascent step. The stochastic gradient decent/ascent step may further be determined by backpropagation of the discriminator network, the encoder network, and/or the generator network.

[0022] Using stochastic gradient decent/ascent and backpropagation improves the efficiency of the networks update and calibration in particular when the number of dimensions of the vehicle sensors data increases.

[0023] The stochastic gradient decent/ascent may be based on a subset of a complete set of training scenarios.

[0024] During updating of the networks during the calibration steps, the ratio between the first category of signals and the second category of signals may be between 1:4 and 4:1, more preferably between 1:2 and 2:1, more preferably between 4:5 and 5:4. The ratio may be close to 1:1. This ensures that the discriminator gives roughly the same weight to the two categories during training. The preferred ratios ensures optimal training conditions.

[0025] The ratio between previous generator output signals and output signals of the generator network may further during networks updating be between 1:4 and 4:1, more preferably between 1:2 and 2:1, more preferably between 4:5 and 5:4. It may be close to 1:1. Keeping the ratio between preferred intervals ensures that a balance between overfitting and giving too much weight to previous generator output signals.

[0026] Generate the anomaly score may include generating a reconstruction loss value and/or generating a discriminator score value. The reconstruction loss value and the discriminator score value may be linearly weighted into the anomaly score.

[0027] The reconstruction loss may be given by the function:

where x is sensor signals captured during a scenario under evaluation.

[0028] For typical (normal) sensor signals x, the reconstruction loss may be small, since both G and E have been trained on typical sensor signals. For anomalous sensor signals x, the reconstruction loss may be large, since neither G nor E have been trained on anomalous sensor signals.

[0029] Using the reconstruction loss (i.e. the amount of deformation of the original data, when encoded with the encoder network and "decoded" with the generator network) to calculate anomaly score further improves the accuracy of indicating anomalous vehicle scenario.

[0030] The discriminator score may be given by the function:

where x is sensor signals captured during a scenario under evaluation.

[0031] The output of the discriminator score function may be close to -1 when the input "x" includes signals from the vehicle sensors during normal scenarios.

[0032] The output of the discriminator score function may be different from -1 (such as close to 0), when the input "x" includes signals from the vehicle sensors captured during anomalous scenario conditions.

[0033] The discriminator score may also be given by the cross entropy function:



[0034] The discriminator score may then be close to 0 when the input signal "x" includes signals from the vehicle sensors during normal scenarios. The discriminator score may be significantly greater than 0 (such exceeding 1, preferably exceeding 10) when the input "x" includes signals from the vehicle sensors captured during anomalous scenario conditions.

[0035] Using the discriminator score (i.e. output of the discriminator network) to calculate anomaly score further improves the accuracy of indicating anomalous vehicle scenario.

[0036] In addition, by linearly weighting the reconstruction loss value and the discriminator score value into the anomaly score, both the discriminator network, generator network, and encoder network are used to assess the anomaly score. This achieves a more stable anomaly score.

[0037] The linearly weighted anomaly score may be given by the function:



[0038] The anomaly score may be weighted using a geometric mean of the reconstruction loss and the discriminator score.

[0039] The discriminator network, the generator network, and/or the encoder network may be convolutional neural networks. The input to the convolutional neural network may then be a signal containing information of scenarios represented in two dimensions. The signal may contain information related to image data.

[0040] By configuring the discriminator network, the generator network, and/or the encoder network to be a convolutional neural network, spatial structure of input signals may be taken into account more efficiently, vastly reducing the number of network parameters, reducing storage and processing requirements.

[0041] The method may further comprise: generating a driver alert of anomalous scenario and/or a request for driver take-over during closed-loop operation, filtering from vehicle training scenario data, scenario data associated with the signal indicating anomalous scenario acquired during open-loop operation, where anomalous scenario indicate irregular driver behaviour, wherein scenario data not associated with the signal indicating anomalous scenario is stored in a storage for vehicle training, and/or reducing a storage volume of a long time storage of vehicle training data, by filtering from the data storage, scenario data associated with the signal indicating anomalous scenario.

[0042] Using the invention, a vehicle controller may be able to distinguish between driving scenarios that the vehicle is able to handle autonomous, from driving scenarios that the vehicle potentially will not be able to handle autonomous in a safe manner. Generating an alert to the driver may make the driver observant on that the vehicle is entering into an environment that the vehicle may not be able to handle autonomous.

[0043] In a closed-loop operation (i.e. autonomous vehicle operation substantially without human direct interaction), this may involve requesting the driver to take over control of the vehicle. Alternatively, the vehicle may enter into a "safe mode" vehicle operation wherein vehicle velocity is decreased or wherein the vehicle is parked in a controlled manner. In an open-loop operation (i.e. vehicle operation involving human direct interaction) this may make the driver aware of a danger scenario that could affect the safe operation of the vehicle and prepare the driver to take responsive actions.

[0044] The method may also be performed on data collected from one or more previous driving occasions. The detections steps may be applied on such data to indicate a potential irregular driving behaviour. This allows scenarios from "bad" driving behaviour to be removed from driving data used to train the vehicle for autonomous driving.

[0045] Further, the method may be applied to remove redundant data in vehicle operation scenario database. Similar scenario data will generate low anomaly score, indicating that limited amount of additional information is contained needed to improve autonomous driving. Redundant scenario data may be removed from storage, such as long term storage.

[0046] The method may further comprise updating a controller of a vehicle in response to the signal indicating anomalous scenario. The signal indicating anomalous vehicle scenario indicates that vehicle controller may have to be amended in order to the vehicle to autonomous handle the indicated scenario in a safe manner in the future. Responsive to the single indicating anomalous vehicle scenario, the autonomous vehicle operation controller may be amended to include the operation of the anomalous scenario. Subsequently, the discriminator network, the generator network, and/or the encoder network may be updated according to the calibrations steps so that the previously anomalous scenario is subsequently considered a normal scenario. In this manner the controller may progressively be improved to safely handle more and more scenarios until human monitoring is substantially obsolete.

[0047] The evaluation machine may be built into the vehicle for the detection steps to be performed in the vehicle. The vehicle may also be part of a system, wherein information regarding vehicle scenario is sent within the system and the signal indicating anomalous vehicle scenario may be performed outside the vehicle. A signal may be sent to the vehicle in order for the vehicle to take appropriate action on the anomalous scenario. An activator in the vehicle such as an indication light might make the passenger aware of anomalous scenario. Vehicle control may also be handed over to remote human operation in case of anomalous scenario.

[0048] The benefits and advantages of the aforementioned vehicle, storage medium and system are equal or similar to the advantages of the above-mentioned method.

[0049] In the following, embodiments of the invention are described with respect to the figures, wherein
Fig. 1
shows a schematic illustration of vehicle operation with data collection and feedback loop;
Fig. 2
illustrates the structure of networks calibration with bidirectional generative networks and data batch with replay buffer;
Fig. 3A
illustrates detection of anomalous sensor data, driver alert and takeover request;
Fig. 3B
illustrates detection of anomalous sensor data and irregular driver behaviour filtering; and
Fig. 3C
illustrates detection of anomalous sensor data and reducing data volume by data deletion.


[0050] As described above, the invention include calibration steps and detection steps. Calibration (or training) is depicted in Fig. 2. Detection (or inference) is primarily depicted in Figs. 1 and 3.

[0051] Fig. 1 depicts a vehicle 100 comprising sensors 120 generating signals 140 indicating current vehicle scenario 110. The signals 140 may be input to an evaluation machine 150 generating signal 160 indicating anomalous scenario. Controller 130 may process signal 160 indicating anomalous scenario. The controller may, in response to signal 160 alert the driver and/or passenger of anomalous scenario. The controller may also put the vehicle in a "safe mode" that may include reducing the speed of the vehicle 100, or parking the vehicle in a controlled manner. Although depicted as outside the vehicle 100, the evaluation machine 150 may be part of the vehicle. The sensors 120 may be part of the vehicle 100, or outside the vehicle 100. The signals 140 may also come from other sources, such as information from third party regarding traffic conditions or navigation assistance.

[0052] The evaluation machine 150 may determine to generate the signal 160 indicating anomalous scenario when an anomaly score is above a pre-set threshold. The anomaly score may be generated from an output signal of the encoder network E, an output signal of the generator network G, and/or an output signal of the discriminator network D.

[0053] Fig. 2 depicts a structure of the evaluation machine during calibration of the discriminator D, generator G, and encoder E. The training procedure for G and E includes calibrating their parameters to maximising the probability of D making a mistake. In a fully trained system, the output of G resembles the training data x, and the output of E resembles the corresponding feature set z, and the discriminator D is not able to distinguish between input data in a first category 250 (including sensor signals x and output signal E(x) of the encoder network with inputs of the sensor signals x), and a second category 240 (including generator output G(z) and noise signals z).

[0054] Whereas generator G transforms (maps) noise signals from the feature space 210 into signal space 220 corresponding to signals from the vehicle sensors 120, encoder network E transforms (maps) signals in signal space into signals in the feature space 210. In a fully trained system, encoder E generated the inverse of generator G for normal training scenarios. Discriminator network D may down-sample input signals into probability estimates. Generator network G may up-sample noise vectors into estimates of signal data. Encoder E may down-sample input signals into feature signals.

[0055] Fig. 2 also depicts a replay buffer 260. Noise signals z and output of generator network G(z) may be stored in the replay buffer 260. Replay buffer 260 may also be used to provide previous generator output signals Grb and previous noise signals zrb. Previous signals 230 may be used to improve the calibration of discriminator network D, and prevent that discriminator D is over-fitted to current input and hence less apt for evaluating anomalous scenarios. Training batch 270 comprises input signals to discriminator network D for calibrating it.

[0056] The training of the networks may be done by determining stochastic gradient (decent or ascent) of the networks loss value (value function or objective) over the training batch 270. The parameters of the discriminator network may be calibrated to increase the networks loss value and the parameters of the generator and encoder networks may be calibrated to decrease the networks loss value.

[0057] Anomaly detection of sensor signals have multiple application in automotive industry. Figs. 3A to 3C shows examples of application of the invention.

[0058] Fig. 3A depicts the evaluation machine 150 used to generate a signal indicating anomalous scenario 310 during vehicle operation. The anomaly score generated by the evaluation machine 150 may indicate how well the current scenario has been covered in previous drives, which have been used for the training of models for autonomous control of the vehicle. If the similar scenario coverage is very low, or not covered at all, the scenario should be classified as anomalous. The more scenarios of similar kinds have been used for training vehicle control models, the better the models may handle them. Hence, the classification into normal and anomalous scenarios may indicate how well driving models (and the vehicle) handle the given situation. During closed-loop operation, i.e. driving without human intervention, if the estimated ability is too low (i.e. the anomaly score is above a pre-set threshold) the signal might indicate that the vehicle may not be confident in its ability to handle the given situation in a safe manner.

[0059] The vehicle may request the driver to take over control of the vehicle. The signal may indicate to the driver that additional precaution is needed, such as that driver's hands are put on the steering wheel and the driver has readiness to brake. The signal may indicate with reducing speed, vibrating steering wheel and/or drifting steering towards the side of the road that additional attention is needed. The vehicle may have sensors for assessing that driver is alert, such as sensor indicating that driver's hands are on the steering wheel, and may enter a safe mode, such as including stopping the vehicle in a controlled manner, such as on the roadside, unless signals indicate that driver is satisfactory controlling the vehicle.

[0060] Fig. 3B depicts the evaluation machine 150 used to filtering vehicle controller training data. Professional drivers operating a vehicle to collect training data for autonomous vehicles to follow (expert data), it may happen that some of the training data is not satisfyingly cautious. The driver may not have been observant during the particular driving session or performed a rough manoeuvre that should not be imitated by the vehicle. The vehicle training scenario data (140) may be input to the evaluation machine 150. Scenario data associated with the signal indicating anomalous scenario 350 acquired may be filtered training data, where anomalous scenario may indicate irregular driver behaviour. Scenario data not associated with the signal indicating anomalous scenario 360 may be stored in a storage 340 for vehicle training.

[0061] Fig. 3C depicts the evaluation machine 150 used to reducing a storage volume of a long time storage 390 of vehicle training data. Non-redundant data, i.e. data that gives additional information of value compared to already stored training data, may be indicated by the evaluation machine as anomalous. In this case, only training data indicating anomalous scenarios 370 are by stored 390 for future training sessions.

[0062] In this case, the pre-defined threshold for the classification of a scenario as an anomaly may be adjusted such that a fraction, such as the top 5% of all collected data (highest anomaly scores) is stored.

Reference numerals:



[0063] 
100
Vehicle
110
Environment
120
Sensors
130
Controller
140
Sensors signals
150
Evaluation machine
160
Signal indicating anomalous scenario
210
Feature space
220
Data space
230
Replay buffer feature set and generator (decoder) generated data
240
Feature set and generator (decoder) generated data (latest generator)
250
"Real" data and encoder generated feature set
260
Replay buffer
270
Batch of training data for networks calibration
310
Driver alert and/or request for driver takeover
340
Storage
350
"Irregular" driver behaviour, filtered from training scenarios
360
Appropriate driver behaviour, used for algorithm training scenarios
370
Sensor data related to anomalous situation (storing for future training)
380
Data representing training scenarios
390
Long term storage
D
Discriminator
G
Generator
E
Encoder
z
Noise signal from random noise generator
zrb
Noise signal from previous iterations stored in replay buffer
G(z)
Output signal of generator network with noise signals input signal
Grb
Generator network output signal from previous iterations stored in replay buffer
x
Data sample from vehicle sensors
E(x)
Output signal of encoder network with data samples input signal
P(y)
Probability of input data being in certain category



Claims

1. Method for generating a signal (160) indicating anomalous vehicle scenario using an evaluation machine comprising a discriminator network (D), a generator network (G), and an encoder network (E), the method comprising calibration steps:

a) storing noise signals and output signals of the generator network (G) in a replay buffer (260);

b) receiving previous generator output signals (Grb) and previous noise signals (zrb) from the replay buffer (260);

c) updating the discriminator network (D), the encoder network (E), and the generator network (G) using:

- training signals (x) from vehicle sensors (120) and an output signals of the encoder network (E);

- the previous generator output signals (Grb) and the previous noise signals (zrb);

- noise signals (z) and an output signals of the generator network (G);

the method further comprising detection steps:

d) generating an anomaly score from an output signal of the encoder network (E), an output signal of the generator network (G), and/or an output signal of the discriminator network (D), wherein an input signal to the encoder network (E) includes data from vehicle sensors (120);

e) comparing the anomaly score with a pre-set threshold; and

f) generating a signal (160) indicating anomalous scenario depending on the comparison, preferably if the anomaly score is above the pre-set threshold.


 
2. The method of claim 1,
wherein method comprises iterating through steps a) to c) a plurality of times,
wherein the discriminator network, generator network and the encoder network are preferably updated within each iteration.
 
3. The method of the claims 1 or 2,
wherein the output signal of the discriminator network (D) indicates a confidence of the discriminator network (D) to categorize:

- a first category of signals (250) comprising sensor signals from vehicle sensors (120) and output signals of the encoder network (E); and

- a second category of signals (240) comprising noise signals and output signals of the generator network (G).


 
4. The method of claim 3,
wherein step c) includes affecting a networks loss value in a first direction by calibrating the discriminator network (D) to increase the confidence; wherein step c) includes affecting the networks loss value in a second direction by calibrating the encoder network (E) and the generator network (G) to decrease the confidence; and
wherein steps a) to c) are preferably iterated until a Nash equilibrium of the network loss value is essentially achieved at least locally.
 
5. The method of any of the preceding claims,
wherein affecting the networks loss value includes determining at least one stochastic gradient,
wherein the stochastic gradient is preferably determined by a backpropagation of the discriminator network, the encoder network, and/or the generator network; and/or
wherein the ratio between the first category of signals and the second category of signals during step c) is preferably between 1:4 and 4:1, more preferably between 1:2 and 2:1, more preferably between 4:5 and 5:4; and/or
wherein the ratio between previous generator output signals (Grb) and output signals of the generator network (G) during step c) is preferably between 1:4 and 4:1, more preferably between 1:2 and 2:1, more preferably between 4:5 and 5:4.
 
6. The method of any of the preceding claims,
wherein generate the anomaly score includes generating a reconstruction loss value and/or generating a discriminator score value; and wherein the reconstruction loss value and the discriminator score value are preferably linearly weighted into the anomaly score.
 
7. The method of any of the preceding claims,
wherein the discriminator network, the generator network, and/or the encoder network is a convolutional neural network, and wherein the input to the convolutional neural network is preferably a signal containing information of scenarios in two dimensions, more preferably a signal containing image data.
 
8. The method of any of the preceding claims,
wherein the method further comprises:

- generating a driver alert of anomalous scenario and/or a request for driver take-over (310) during closed-loop operation;

- filtering from vehicle training scenario data (140), scenario data associated with the signal indicating anomalous scenario (350) acquired during open-loop operation, where anomalous scenario indicate irregular driver behaviour;
wherein scenario data not associated with the signal indicating anomalous scenario (360) is stored in a storage (340) for vehicle training; and/or

- reducing a storage volume of a long time storage (390) of vehicle training data, by filtering from the data (380) storage (340), scenario data associated with the signal indicating anomalous scenario (370).


 
9. The method of any of the preceding claims,
wherein the method further comprises:

- updating a controller (130) of a vehicle (100) in response to the signal (160) indicating anomalous scenario.


 
10. Vehicle (100) or system with a vehicle (100), comprising an evaluation machine with a discriminator network (D), a generator network (G), and an encoder network (E) adapted according to the calibration steps and comprising means for carrying out the detection steps of one of the claims 1 to 9.
 
11. Computer readable storage medium containing instructions which, when executed by at least one processor, cause the processor to carry out the steps of the method of one of the claims 1 to 9.
 




Drawing


























REFERENCES CITED IN THE DESCRIPTION



This list of references cited by the applicant is for the reader's convenience only. It does not form part of the European patent document. Even though great care has been taken in compiling the references, errors or omissions cannot be excluded and the EPO disclaims all liability in this regard.

Patent documents cited in the description