Method and apparatus for verifying a KI-based information processing system

By setting up test devices and multi-dimensional data structures on the vehicle, evaluating and updating the test results of the KI-based information processing system, the problem of low transparency and robustness of deep neural networks in autonomous driving is solved, and the system is realized The robustness evaluation and continuous update of data pools reduce development and testing costs.

CN114368390BActive Publication Date: 2024-11-08VOLKSWAGEN AG
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111205068.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2020-10-15
Filing Date
2021-10-15
Publication Date
2024-11-08
Estimated Expiration
2041-10-15

AI Technical Summary

Technical Problem

Existing KI-based information processing systems, especially deep neural networks, have problems such as low transparency, susceptibility to anti-interference, difficulty in measuring robustness and difficulty in maintaining high efficiency in different data domains in autonomous driving applications of vehicles. , resulting in high development and testing costs and limited functional safety.

Method used

Provide a method and equipment to evaluate and robustness test the system by setting up a test device on a vehicle, storing and updating the test results of the KI-based information processing system using a multi-dimensional data structure, and to evaluate and test the system using an unmonitored test method. Includes data augmentation definitions and difference metric definitions to evaluate the system's performance in the face of different interferences and environments.

Benefits of technology

Continuous inspection and data pool updates of KI-based information processing systems during vehicle application can evaluate and improve system robustness, ensure stability and safety under different conditions, and reduce development and testing cost.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114368390B_ABST
    Figure CN114368390B_ABST
Patent Text Reader

Abstract

The invention relates to a method for testing a KI-based information processing system used in the partially or fully automatic control of a vehicle, wherein at least one sensor of the vehicle provides sensor data, the detected sensor data are evaluated by means of a KI-based information processing system arranged in a first control device of the vehicle, and at least one output for controlling the vehicle is generated based on the evaluated sensor data and provided to a control unit of the vehicle, wherein the KI-based information processing system is tested by means of at least one test method by means of a test device arranged in a second control device of the vehicle, and wherein the test results of the at least one test method are stored in a multidimensional data structure in a database arranged in the vehicle with reference to the tested KI-based information processing system and the at least one test method used. The invention also relates to a corresponding device.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method and a device for testing a KI-based information processing system used in the partially or fully automatic control of a vehicle. Background Art

[0002] For example, machine learning based on neural networks has great potential for use in modern driver assistance systems and autonomous vehicles. In this case, functions based on deep neural networks process sensor data (e.g. from cameras, radar or lidar sensors) to derive relevant information from them. This information includes, for example, the type or position of objects in the environment of the motor vehicle, the behavior of objects or the geometry or topology of the roadway.

[0003] The main feature when developing a KI-based information processing system (training) is: purely data-driven parameter fitting (Parameterfitting) without expert intervention. In this case, for example in the case of deep neural networks, the deviation of the output of the neural network (for a given parameterization) from the ground truth (the so-called loss) is determined. The loss function used in this case is selected in such a way that the parameters of the neural network depend on the loss function in a differentiable way. Within the scope of the gradient descent method, in each training step, the parameters of the neural network are adapted according to the derivatives of the deviations determined based on a number of examples. These training steps are repeated very frequently until the loss no longer decreases.

[0004] In this action, the parameters of the KI-based information processing system, especially the neural network, are determined without expert estimation or semantically driven modeling. This may cause significant consequences for the characteristics of the KI-based information processing system, especially the neural network.

[0005] In particular, deep neural networks are largely opaque to humans and their calculations are not interpretable. This represents a significant limitation for system testing or formal verification.

[0006] Furthermore, deep neural networks in particular are susceptible to harmful interference, so-called adversarial perturbations: small manipulations of the input data that are barely perceptible to humans or do not change the semantic content can lead to completely different output data. Such manipulations can be both intentional changes to the data ("neural hacking") and randomly occurring changes to the image (sensor noise, weather effects, specific colors or contrasts).

[0007] Furthermore, it is not clear which input features the neural network is sensitive to. This has the result that, for example, synthetic data generated by simulation has so far been almost unsuccessful in the training of neural networks: neural networks trained in simulation or based on other synthetic data have surprisingly poor performance on real sensor data. In addition, the execution of neural networks in other data domains (training in summer, execution in winter, etc.) sometimes also greatly reduces the functional quality. This has the result, in particular, that the possibilities that sound very attractive from a cost point of view for the development and release of neural networks in simulation (eliminating expensive marking and complex actual testing) are not practical.

[0008] The second point is particularly relevant for the possible limitations of powerful neural networks in the context of functional safety. In order to measure functional safety, it is desirable to measure the robustness of the network to slight changes in the input data (e.g., augmentation). Since such changes can be of various kinds (sensor noise, weather influences, image manipulation, semantically meaningless changes in content such as the wall color of a background building), there is no clear and accepted measure for robustness. On the contrary, many robustness values ​​can be measured relative to disturbances of different types and intensities (e.g., augmentation). In addition, the robustness of a neural network is not an absolute variable, but rather depends on the current input data.

[0009] Furthermore, it becomes difficult to produce a KI-based information processing system that is robust with respect to all necessary disturbances. In addition, the space of possible disturbances is infinite, so that new forms of disturbances may occur during the operation of the vehicle.

[0010] A method for generating a robust automatic learning system and testing a trained automatic learning system is known from DE 10 2018 218 586 A1. Summary of the invention

[0011] The object of the present invention is to create a method and a device for testing a KI-based information processing system used in partially or fully automatic control of a vehicle, with which method and device in particular a data base can be continuously maintained and / or updated.

[0012] This object is achieved according to the invention by a method having the features of the invention and a device having the features of the invention. Advantageous further developments of the invention are described below.

[0013] In particular, a method for testing a KI-based information processing system used when partially or fully automatically controlling a vehicle is provided, wherein at least one sensor of the vehicle provides sensor data for environmental detection, the detected sensor data are evaluated by means of the KI-based information processing system arranged in a first control device of the vehicle, and at least one output for partially or fully automatically controlling the vehicle is generated based on the evaluated sensor data and provided to a control unit of the vehicle, wherein the KI-based information processing system is tested by means of at least one test method by means of a test device arranged in a second control device of the vehicle, and wherein the test results of the at least one test method are stored in a multidimensional data structure in a database arranged in the vehicle with a reference to the tested KI-based information processing system and the at least one test method used.

[0014] In addition, a device is created for testing a KI-based information processing system used when partially or fully automatically controlling a vehicle, the device comprising: at least one sensor, which is configured to provide sensor data for environmental detection of the surrounding environment of the vehicle; a KI-based information processing system arranged in a first control device of the vehicle, which is configured to evaluate the detected sensor data and generate and provide at least one output for partially or fully automatically controlling the vehicle based on the evaluated sensor data; a control unit, which uses the provided output as a control parameter for partially or fully automatically controlling the vehicle, and a test device arranged in a second control device of the vehicle, which is configured to test the KI-based information processing system with the aid of at least one test method, and store the test results of the at least one test method in a multidimensional data structure in a database arranged in the vehicle with reference to the tested KI-based information processing system and the at least one test method used.

[0015] The method and device enable a KI-based information processing system used in the partially or fully automatic control of a vehicle to be tested during use, and thereby expand and / or update a data pool for testing and evaluating the KI-based information processing system. In particular, the evaluation of the KI-based information processing system in terms of robustness can be carried out at a later point in time by using test results stored in a multidimensional data structure. Therefore, during the use of the KI-based information processing system, it is provided that the KI-based information processing system is tested by means of at least one test method. The test results of the at least one test method are stored in a multidimensional data structure in a database arranged in the vehicle, wherein a reference to the tested KI-based information processing system and the at least one test method used is associated, so that the test results can be used later to test and / or evaluate the KI-based information processing system and can be called up from the multidimensional data structure in a targeted manner for this purpose.

[0016] The method and the device have the advantage that when using a KI-based information processing system, test results can be continuously checked and collected for a KI-based information processing system used in the partially or fully automatic control of a vehicle. As a result, the data pool for checking and / or evaluating can be expanded and / or updated step by step and in particular continuously.

[0017] The KI-based information processing system is in particular an information processing system based on methods of artificial intelligence (KI). For example, the KI-based information processing system can be designed as a deep neural network or include at least one such deep neural network. However, in principle, the method described in the present disclosure can also be used in other KI-based information processing systems, for example in rule-based information processing systems. In particular, at least one KI-based information processing system is trained and / or finally parameterized. For example, the KI-based information processing system can be a trained neural network. The KI-based information processing system in particular includes a structural description and parameters and / or is defined by a structural description and parameters.

[0018] The test method is in particular an unmonitored test method, i.e., no ground truth is known for the output of the KI-based information processing system during the test. In particular, the test method is performed so that the processes of the KI-based information processing system are not interfered with or such processes are not changed. In particular, the test method is passive compared to the KI-based information processing system to be tested. In other words, the test process only observes the KI-based information processing system.

[0019] The sensor is in particular a camera, a lidar sensor, a radar sensor or an ultrasonic sensor. The sensor data can be one-dimensional or multi-dimensional, in particular two-dimensional or three-dimensional. If the sensor is designed as a camera, the sensor data in particular includes a two-dimensional camera image.

[0020] Parts of the device, in particular the first control device, the second control device, the control unit and / or the testing device, can be designed individually or jointly as a combination of hardware and software, for example as a program code executed on a microcontroller or microprocessor.

[0021] It can be provided that the multidimensional data structure already exists and is transmitted to the device, for example, from a central server, and is used there, for example, to check and assess the robustness of the KI-based information processing system before application. The test results generated according to the method described in the present disclosure are added as new entries to the provided multidimensional data structure, so that the test results are extended to the multidimensional data structure. However, it is also possible to update already existing entries or data points in the multidimensional data structure. The multidimensional data structure is stored in particular in a database. On a central server, the database is in particular referred to as a central database, while in the device, the database is in particular referred to as a decentralized database.

[0022] The multidimensional data structure can be defined, for example, by the dimensions: "KI-based information processing system", "data group", "data enhancement definition" and "difference measure definition". In particular, the multidimensional data structure can also be referred to as a hypercube. The dimension "KI-based information processing system" as a value range includes in particular all set KI-based information processing systems (as long as more than one KI-based information processing system exists or should be applied and / or tested). If only one KI-based information processing system exists or should be applied and / or tested, the dimension "KI-based information processing system" can also be cancelled. The dimension "data group" includes in particular all set data groups as a value range, with which the KI-based information processing system referenced in the multidimensional data structure transmitted to the device has been verified. The dimension "data enhancement definition" as a value range includes in particular all the following data enhancement methods, with which the KI-based information processing system referenced in the multidimensional data structure has been enhanced for verification. The dimension "difference measure definition" as a value range includes in particular all difference measures that have been used for verification. To each combination of values ​​within these dimensions is assigned a data point that includes at least one difference value. The test results that have been obtained with the aid of the method described in the present disclosure are added as data points to such a multidimensional data structure. In this case, it is not necessary to take into account all dimensions present in the multidimensional data structure. In particular, the test method of the method described in the present disclosure does not provide differences because it is an unmonitored test method. In other words, the obtained test results are added to the multidimensional data structure, in particular as at least one new data point. For an unmonitored test method, the multidimensional data structure has in particular the dimensions: KI-based information processing system, test method, characteristics and / or context (Kontext) of sensor data.

[0023] The data set includes in particular data. The data may be in particular one-dimensional or multi-dimensional, in particular two-dimensional. For example, the data may be images of a camera or a lidar sensor. However, in principle any sensor data may be used.

[0024] The data enhancement definition defines in particular a data enhancement or a data enhancement method. The data enhancement definition specifies how the data of a data set should be changed when generating a multidimensional data structure. In this case, a large number of changes can be provided. By way of example, the following should be mentioned: adding noise, adding one or more countermeasure interferences and / or sensor interferences, contrast change, brightness change, color change, weather condition change (e.g. adding snow or rain to a camera image detected in summer when the sun is shining). The data enhancement or the data enhancement method is designed or defined in particular depending on physical sensor properties (interferences, etc.) and / or possible physical interferences and / or technical interferences and / or possible countermeasure interferences of the sensor system.

[0025] The difference metric definition defines, in particular, a difference metric. The difference metric specifies, in particular, how output data generated by a KI-based information processing system for (unenhanced) data of a data group should be compared with output data generated by the KI-based information processing system for enhanced data when generating a multidimensional data structure. If a KI-based information processing system outputs, for example, vectors as output data, the difference metric may include a comparison of vectors, for example by determining the difference between the vectors. A simple example for other difference metrics is as follows: If a KI-based information processing system outputs, for example, as output data how many pedestrians there are in a detected camera image, the numbers outputted respectively for the data and the enhanced data may be compared with each other (for example, 3 pedestrians compared to 5 pedestrians, so that the difference is equal to 2 pedestrians).

[0026] If the chronological data are stored in a data set, the difference measure can also relate to the chronological data, i.e., the temporally adjacent data. Thus, a data pool can be generated and made available by a KI-based information processing system for robustness evaluation when processing video sequences (or other chronological data). For example, within the scope of the robustness evaluation, it can be checked whether a pedestrian in a video sequence spanning a plurality of individual video images is reliably identified as a pedestrian by a KI-based information processing system.

[0027] In particular, all artifacts that are decisive in the generation of the multidimensional data structure can also be stored as metadata and / or headers in the multidimensional data structure. These artifacts include, for example: a reference to the software code used, a reference to at least one KI-based information processing system and the hyperparameters used for training, a reference to one or more used data sets (including descriptive data if necessary) and / or a reference to the used initial values ​​("random seeds") for the used random generators.

[0028] The multidimensional data structure in particular forms a test database, with which the test results obtained under different conditions for different KI-based information processing systems can be expanded on a central server by test results obtained in the ongoing operation according to the method described in the present disclosure during the application of the KI-based information processing system. With the aid of the test results in the multidimensional data structure, the KI-based information processing system can be evaluated and assessed in an improved manner with respect to robustness. In particular, the provided multidimensional data structure allows the selection of a robust KI-based information processing system for an application under given boundary conditions (characteristics of the sensor data, context, etc.).

[0029] Before executing the method described in the present disclosure, a multidimensional data structure can be generated, in particular on a central server, for example, by a method for providing a data pool for evaluating the robustness of at least one KI-based information processing system, wherein at least one KI-based information processing system, at least one data group, at least one data enhancement definition and at least one difference metric definition are received as input parameters, wherein the multidimensional data structure is generated based on these input parameters, wherein these dimensions of the multidimensional data structure and the value ranges of these dimensions are set by the received input parameters, and wherein each data point of the multidimensional data structure includes a difference value determined by means of at least one defined difference metric, and the difference value is determined in such a way that at least one defined difference metric is formed between the following output data, wherein the output data is generated by at least one KI-based information processing system for data of at least one data group and for the same data enhanced by means of at least one defined data enhancement, and wherein the generated multidimensional data structure is provided so that the robustness of the at least one KI-based information processing system can be evaluated based on the difference values ​​included by the multidimensional data structure. The multidimensional data structure is subsequently transmitted to the device in order to be expanded there with further test results and / or updated by means of the method described in the present disclosure for testing a KI-based information processing system used in partially or fully automatic control of a vehicle.

[0030] The method and the device can be used in particular in a vehicle. The vehicle is in particular a motor vehicle. However, in principle, the vehicle can also be another land, water, air or space vehicle. An air vehicle is in particular a drone or a partially automated or automated air bus. In principle, the method can also be used in the production or logistics sector, for example in monitoring or in robotics.

[0031] In one specific embodiment, it is provided that the output is at least one object classified from the sensor data by the KI-based information processing system or is a control output, such as a steering angle, a speed or a trajectory.

[0032] In one embodiment, it is provided that at least one context information of the context in which the sensor data is detected is determined, wherein in addition to the test results, the determined at least one context information is additionally stored in a multidimensional data structure. Thus, the test results can be stored in the following multidimensional data structure, which is associated with the context in which the sensor data is detected. At a later point in time, the already existing test results can be called from the multidimensional data structure for a given context, and used to evaluate and assess the robustness of the information processing system based on KI in the same context. This in particular allows the robustness of the information processing system based on KI to be evaluated and assessed in relation to the context. The context can in particular include information about the environment in which the sensor data has been detected, such as the speed of the vehicle, the time of the day, the day of the week, the season, the surrounding environment and / or the location (e.g., a city, a country, a street for children to play, a highway, etc.) and / or weather conditions (e.g., rain, snow, fog, etc.). For example, at least one context information is determined by means of a context determination device provided for this purpose of the device.

[0033] In one embodiment, it is provided that the characteristics of the detected sensor data are determined, wherein in addition to the test results, the determined characteristics are additionally stored in a multidimensional data structure. Thus, the test results can be stored in a multidimensional data structure, which is associated with the characteristics of the detected sensor data. At a later time point, the existing test results can be called from the multidimensional data structure for similar sensor data, i.e. sensor data with the same characteristics, and used to evaluate and assess the robustness of the information processing system based on the KI when similar sensor data is conveyed. This is especially allowed to evaluate and assess the robustness of the information processing system based on the KI according to the characteristics of the sensor data. The characteristics of the sensor data can, for example, include one or more of the following: interference in the sensor data, such as countermeasure interference or noise, in addition to photometric characteristics, such as brightness, contrast, color saturation, hue, etc. The characteristics of the sensor data can be determined, for example, by means of a computer vision method or by means of a trained machine learning method, such as an autoencoder or a generative adversarial network (GAN). For example, these characteristics are determined by means of a characteristic determination device provided for this purpose and / or by means of at least one interference recognition device of the device.

[0034] In one embodiment, it is provided that the at least one test method includes: determining and / or checking the consistency of at least one output during a time change. In this way, the reliability or robustness of the output during a time change can be evaluated. In particular, it is checked to what extent the output of the KI-based information processing system changes over time or remains the same. In this case, for example, it can be checked and evaluated whether the output changes in a jumpy manner and / or oscillates over time. If the KI-based information processing system under consideration is used, for example, for object recognition and / or object classification in the detected camera image, it can be checked for the recognized object whether the object class assigned to the recognized object changes or remains constant during the time change. If the object class changes multiple times, the consistency of the KI-based information processing system is low, and on the contrary, if the assigned object class remains constant during the time change, the consistency is high. High consistency indicates that the KI-based information processing system works well, that is, robustly, under given circumstances (such as characteristics and / or context of sensor data, etc.). One or more corresponding test results for the tested KI-based information processing system are then stored in a multidimensional data structure, for example with reference to the test method “(time) consistency”.

[0035] In one embodiment, it is provided that at least one test method includes: determining and / or checking the confidence level of at least one output. This makes it possible to evaluate how well, i.e., how reliably or robustly, the KI-based information processing system works, in particular under given conditions. This confidence level is estimated and provided, in particular, by the KI-based information processing system itself. If the KI-based information processing system is, for example, a neural network that performs object classification in the detected camera data, the neural network provides estimated values ​​for different object classes as outputs. If a high estimated value (e.g., 99%) is assigned to an object class, the confidence level is greater than when only a lower estimated value (e.g., 30%) is assigned to the object class and the remaining probability is distributed over many other object classes. Then, for example, one or more corresponding test results for the tested KI-based information processing system are stored in a multidimensional data structure with reference to the “confidence level” of the test method.

[0036] In one embodiment, it is provided that at least one test method includes a rationality check on at least one output. Thus, the output of a KI-based information processing system can be compared with the outputs of other KI-based information processing systems and / or the (overall) context. The outputs generated for the detected sensor data of different sensors can also be compared with each other. When the degree of conformity (Übereinstimmung) of the output is large, the rationality of the output of the tested KI-based information processing system is large, and when the degree of conformity is low, the rationality of the output of the tested KI-based information processing system is correspondingly low. Then, for example, one or more corresponding test results for the tested KI-based information processing system are stored in a multidimensional data structure using a reference to the "rationality" of the test method. In this case, the reference may include other information about the type of rationality check.

[0037] In one embodiment, it is provided that the multidimensional data structure is synchronized with a central multidimensional data structure stored in a central database on a central server and / or with a decentralized multidimensional data structure stored in a decentralized database. As a result, newly acquired test results can be shared with a common infrastructure, in particular with other vehicles. In this way, the data pool for evaluating and assessing the robustness of KI-based information processing systems can be continuously expanded and / or updated. In this case, synchronization includes in particular the exchange of information in both directions.

[0038] In one specific embodiment, it is provided that the KI-based information processing system provides functions for autonomous driving of a vehicle and / or for driver assistance of a vehicle and / or for environmental detection and / or environmental perception.

[0039] In one embodiment, it is provided that the KI-based information processing system is a neural network and / or comprises at least one neural network. The neural network is in particular a deep neural network, in particular a convolutional network.

[0040] The other features for configuring the device are derived from the description of the configuration of the method. The advantages of the device are in this case the same as in the configuration of the method.

[0041] In addition, in particular, a system is created for verifying a KI-based information processing system used in partially or fully automatic control of a vehicle, the system comprising: at least one, in particular distributed, device for verifying a KI-based information processing system used in partially or fully automatic control of a vehicle according to one of the described embodiments, and a central server, wherein at least one multidimensional data structure is stored in a central database on the central server, and wherein the central server is configured to generate the multidimensional data structure and / or transmit the multidimensional data structure to the in particular distributed device and / or synchronize the at least one multidimensional data structure with the multidimensional data structure of the in particular distributed device.

[0042] The method, device and system may be configured as part of a device, system or method for partially or fully automatic control of a vehicle.

[0043] The method in this case provides, for example, that sensor data for environmental detection are provided by means of at least one sensor of the vehicle, the detected sensor data are evaluated by means of a KI-based information processing system arranged in a first control device of the vehicle, and based on the evaluated sensor data at least one output for partially or fully automatic control of the vehicle is generated and provided to a control unit of the vehicle, wherein properties of the sensor data and / or a current context in which the sensor data are detected are determined by means of an analysis device arranged in a second control device of the vehicle, wherein a KI-based information processing system is selected from a plurality of available KI-based information processing systems by means of a selection device arranged in the second control device, wherein the selection is performed using a multidimensional data structure in which the plurality of KI-based information processing systems are selected. The information processing system stores the test results for at least one test method in the multidimensional data structure according to the characteristics of the sensor data and / or according to the context, and the multidimensional data structure is stored in a database arranged in the vehicle, wherein the robustness of multiple KI-based information processing systems is determined and evaluated based on the multidimensional data structure under the condition of specific characteristics and / or specific context, and the most robust KI-based information processing system is selected and used to evaluate the detected sensor data in the first control device.

[0044] The robustness is determined in particular on the basis of at least one robustness measure. The robustness measure in particular specifies how the test results stored in the multidimensional data structure, in particular the difference values ​​and the test results of the unmonitored test method, should be combined into one or more robustness values. The one or more robustness values ​​obtained for the respective KI-based information processing system can then be compared with one another and / or with a threshold value.

[0045] The device may be set up to also carry out the method described in the above paragraphs for partially or fully automatic control of a vehicle by providing a robust KI-based information processing system. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The present invention will be described in more detail below with reference to the accompanying drawings according to a preferred embodiment. In this case:

[0047] Figure 1 A schematic diagram showing an embodiment of a device for testing an information processing system based on KI used when partially or fully automatically controlling a vehicle;

[0048] Figure 2 A schematic diagram is shown for illustrating embedding the method into a system and method for partially or fully automatically controlling a vehicle by providing a robust KI-based information processing system. DETAILED DESCRIPTION

[0049] Figure 1 A schematic diagram of an embodiment of a device 1 for testing a KI-based information processing system 10 used in partially or fully automatic control of a vehicle is shown. The KI-based information processing method 10 is executed in a vehicle 50 for evaluating sensor data. The KI-based information processing system 10 can be, for example, a trained deep neural network.

[0050] The device 1 comprises a sensor 51 which is configured to provide sensor data 11 for environmental detection of the surroundings of the vehicle 50. For example, the sensor 51 can be a camera, with which a camera image of the surroundings of the vehicle 50 is detected as sensor data 11.

[0051] Furthermore, the device 1 comprises a KI-based information processing system 10 arranged in a first control device 52 of the vehicle 50, which is configured to evaluate the detected sensor data 11 and to generate and provide at least one output 30 for partially or fully automatic control of the vehicle 50 based on the evaluated sensor data 11. It can be provided that the control device 52 is part of the device 1.

[0052] In this case, it can be provided that the output 30 is at least one object classified by the KI-based information processing system 10 from the sensor data 11 or is a control output 30 , such as a steering angle, a speed or a trajectory.

[0053] Furthermore, the device 1 comprises a control unit 53 which uses the provided output 30 as a control parameter for partially or fully automatic control of the vehicle 50 . The control unit 53 is, for example, a trajectory planner or an actuator controller of the vehicle 50 .

[0054] The device 1 further comprises a test device 2 which is arranged in a second control device 54 of the vehicle 50 . The second control device 54 is in particular part of the device 1 . The test device 2 comprises a computing device 3 and a memory 4 .

[0055] The device 1 is designed, in particular, to carry out the method described in the publication for testing a KI-based information processing system 10 that is used in the partially or fully automatic control of a vehicle 50 .

[0056] In particular, it is provided that KI-based information processing system 10 provides functions for autonomous driving of vehicle 50 and / or for driver assistance of vehicle 50 and / or for environment detection and / or environment perception.

[0057] The mode of functioning of the device 1 is shown by way of example for a single KI-based information processing system 10. However, in principle, a plurality of KI-based information processing systems 10 operating in parallel can also be tested simultaneously.

[0058] The test device 2 is configured to check a KI-based information processing system 10 by means of at least one test method 12-x and to store test results 13-x of the at least one test method 12-x in a multidimensional data structure 20 in a database 6 arranged in the vehicle 50 with reference to the tested KI-based information processing system 10 and the at least one test method 12-x used. The at least one test method 12-x is in particular an unmonitored test method.

[0059] For this purpose, at least one output 30 is supplied to the testing device 2. This takes place, in particular, continuously or continually during the use of the KI-based information processing system 10.

[0060] It can be provided that the test method 12-1 includes determining and / or checking the consistency of at least one output 30 during the time change. For this purpose, the outputs 30 during the time change are compared with one another with respect to deviations and / or correspondences. If the outputs 30 include, for example, object classes estimated by the KI-based information processing system 10, then within the scope of the test method 12-1 it is checked whether the object class estimated for the object changes during the time change. The test result 13-1 is stored in the multidimensional data structure 20.

[0061] It may be provided that the test method 12-2 comprises determining and / or checking a confidence level of the at least one output 30. In general, the output 30 generated by the KI-based information processing system 10 comprises a statement about the probability of the respective output 30. In the example of estimating the object class of an object imaged in a camera image by means of a neural network, the neural network provides, for each of the possible object classes, a probability that the imaged object belongs to one of these object classes as output 30. These probabilities can be used as confidence levels in order to check the reliability or robustness of the neural network. The test result 13-2, i.e., the confidence level(s), is stored in the multidimensional data structure 20.

[0062] It may be provided that the testing method 12 - 3 includes performing a rationality check on at least one output 30 . The rationality check may be performed depending on the context and / or on outputs 30 of other KI-based information processing systems 10 .

[0063] It can be provided that at least one context information 15 of the context of the detected sensor data 11 is determined, wherein in addition to the test results 13-x, the at least one determined context information 15 is stored in the multidimensional data structure 20. For this purpose, the detected sensor data 11 are fed to the second control device 54, in particular the computing device 3, wherein the context can be recognized by means of a pattern recognition method, for example based on a trained machine learning method, and at least one context information 15 can be determined. For example, the context information can include information about the location (city, country, highway, street for children to play, garage entrance, etc.), time (day, night, summer, winter, etc.) and / or situation (weather, lighting conditions, traffic flow, etc.). In order to determine the at least one context information 15, the device 1, in particular the computing device 3, in particular includes a context determination device 19.

[0064] It can be provided that the properties 16 of the detected sensor data 11 are determined, wherein the determined properties 16 are stored in addition to the test results 13-x in the multidimensional data structure 20. If the sensor 51 is, for example, a camera, the properties can include, in particular, photometric properties of the detected camera image, such as brightness, contrast, etc. For determining the properties 16, the device 1, in particular the computing device 3, includes, in particular, a property determination device 18. Furthermore, the device 1, in particular the computing device 3, can include a disturbance detection device 17 for detecting countermeasures as properties of the sensor data 11.

[0065] In one specific embodiment, it is provided that the multidimensional data structure 20 is synchronized with a central multidimensional data structure 20 stored in a central database 41 on a central server 40 and / or with a decentralized multidimensional data structure 20 stored in a decentralized database 6 of another vehicle (not shown). In this case, it can be provided in particular that only expanded and / or changed parts of the multidimensional data structure 20 are transmitted for synchronization in order to reduce the amount of data.

[0066] exist Figure 2 Schematic diagram for illustrating embedding the method into a system 80 and method for partially or fully automatically controlling a vehicle 50 by providing a robust KI-based information processing system 10 r is shown in .

[0067] Within the scope of a method for partially or fully automatically controlling a vehicle by providing a robust KI-based information processing system 10r, a multidimensional data structure 20 is generated on a central server 40 based on a plurality of KI-based information processing systems 10, different test methods 12-x and different data sets 5-x.

[0068] This is carried out in measure 90, in which in particular a method for providing a data pool for evaluating the robustness of at least one KI-based information processing system 10 is performed, wherein at least one KI-based information processing system 10, at least one data group 5-x, at least one data enhancement definition and at least one difference metric definition are received as input parameters, wherein a multidimensional data structure 20 is generated based on the input parameters, wherein the dimensions of the multidimensional data structure 20 and the value ranges of the dimensions are set by the received input parameters, and wherein each data point of the multidimensional data structure includes a difference value determined by means of at least one defined difference metric, the difference value being determined in such a way that at least one defined difference metric is formed between the following output data, wherein the output data is generated by at least one KI-based information processing system 10 for data of at least one data group 5-x and for the same data enhanced by means of at least one defined data enhancement, and wherein the generated multidimensional data structure 20 is provided so that the robustness of at least one KI-based information processing system 10 can be evaluated based on the difference values ​​included by the multidimensional data structure 20.

[0069] Based on the multidimensional data structure 20 and the test results (i.e., differences) stored therein and, if necessary, on the basis of other meta-information (context information, characteristics of sensor data or data sets, etc.), a robust KI-based information processing system 10r can then be determined for specific boundary conditions by correspondingly evaluating the test results, in particular the stored differences, in the multidimensional data structure 20. For this purpose, the test results, in particular the differences, are summarized and compared with a robustness or reliability threshold value, for example. Depending on the comparison result, the KI-based information processing system 10 is classified as robust or non-robust.

[0070] In measure 95, a KI-based information processing system 10 on a central server 40 is tested and / or evaluated. In this measure 95, in particular, a method for evaluating and certifying the robustness of a KI-based information processing system 10 is performed, wherein a multidimensional data structure 20 is received in a manner subordinate to the KI-based information processing system 10, wherein at least the following difference values ​​are stored in the multidimensional data structure 20 at least according to the dimensions of the data set 5-x, (one or more) data enhancement definitions and (one or more) difference measures, respectively, wherein the difference values ​​are the difference values ​​between the output data obtained by the KI-based information processing system 10 for the data and the enhanced data determined by means of at least one difference measure, wherein at least one robustness of the KI-based information processing system 10 is determined from at least one selection of these difference values, and the at least one robustness is compared with at least one robustness requirement, and wherein the KI-based information processing system 10 is discarded (verwerfen) based on the comparison result, the KI-based information processing system 10 is re-evaluated using a modified multidimensional data structure 20, or the KI-based information processing system 10 is certified as robust. This is done for all KI-based information processing systems 10 .

[0071] The generated multidimensional data structure 20 and the KI-based information processing system 10r classified as robust in measure 95 are transmitted to at least one vehicle 50. The vehicle 50 can be a test vehicle or a vehicle currently in production.

[0072] In the measure 100, the method described in the present disclosure for testing a KI-based information processing system 10a used or activated when partially or fully automatically controlling a vehicle 50 is performed during application. In this case, the output of the activated KI-based information processing system 10a is tested with the aid of at least one test method 12-x as already described above. The obtained test results are stored in the multidimensional data structure 20. In particular, the multidimensional data structure 20 is supplemented with test results 13-x with references to the activated KI-based information processing system 10a and the respectively used test methods 12-x. The multidimensional data structure 20 can then be synchronized with a multidimensional data structure 20 on a central server 40 and / or with multidimensional data structures 20 of other vehicles.

[0073] Furthermore, in the measure 200, it is provided that, for the evaluation of the detected sensor data 11, for example a camera image detected by means of a camera, the most robust KI-based information processing method 10 under given conditions, in particular given context and / or characteristics of the sensor data, is selected and activated from the KI-based information processing methods 10. Characteristics in the sensor data 11 can be identified, for example, by means of a characteristic identification device 18 and / or a disturbance identification device 17, which for example identifies adversarial disturbances in the sensor data 11. In this case, the selection is made based on the test results 13-x stored in the multidimensional data structure 20, wherein the robustness of the KI-based information processing system 10r considered for the evaluation is determined and evaluated based on the respectively stored test results 13-x. In particular, at least one robustness measure can be determined from the difference values ​​and / or the test results stored in the multidimensional data structure 20. The at least one determined robustness measure can be compared with at least one robustness requirement. The most robust KI-based information processing system 10r, for example the one with the highest robustness value, is then activated for use with the currently detected sensor data 11 and, for this purpose, is loaded, for example, into the first control device 52 ( Figure 1 ) in a memory of the computer or in the control device 52 to be activated for application, and then an activated KI-based information processing system 10a is provided.

[0074] exist Figure 2The system 80 or method described in the specification allows the use of the most robust KI-based information processing system 10r under given conditions (characteristics and / or context of the sensor data 11, etc.) for evaluating the detected sensor data 11. In this case, in particular, the selection is always made taking into account the current conditions in the vehicle 50. With the help of the method and the associated device described in the present disclosure for testing the KI-based information processing system 10r used when the vehicle 50 is partially or fully automatically controlled, the multidimensional data structure 20 used in this case can be continuously expanded with other test results of the unmonitored test method during the application of the KI-based information processing system 10a that is just activated. In this way, the data pool for testing and / or evaluating the robustness of the KI-based information processing systems 10, 10r, 10a can be continuously or stepwise expanded and / or updated.

[0075] Reference numerals list

[0076] 1 Equipment

[0077] 2 Test equipment

[0078] 3 Computing Devices

[0079] 4 Memory

[0080] 5-x Data Group

[0081] 6 (Decentralized) Database

[0082] 10 Information Processing System Based on KI

[0083] 10a Activated KI-based information processing system

[0084] 10r Robust KI-based information processing system

[0085] 11 Sensor Data

[0086] 12-x Test Method

[0087] 13-x Test Results

[0088] 15 Contextual Information

[0089] 16 Features

[0090] 17 Interference identification device

[0091] 18. Characteristic determination device

[0092] 19 Context determination device

[0093] 20 Multidimensional Data Structures

[0094] 30 Output

[0095] 40 Central Server

[0096] 41 (Central) Database

[0097] 50 Transportation

[0098] 51 Sensors

[0099] 52 First control device

[0100] 53 Control Unit

[0101] 54 Second control device

[0102] 80 Systems

[0103] 90 measures

[0104] 95 Measures

[0105] 100 measures

[0106] 200 measures.

Claims

1. A method for testing a KI-based information processing system (10) used in partially or fully automatic control of a vehicle (50), wherein - at least one sensor (51) of the vehicle (50) provides sensor data (11) for detecting the environment, - evaluating the detected sensor data (11) by means of a KI-based information processing system (10) arranged in a first control device (52) of the vehicle (50), and - generating at least one output (30) for partially or fully automatic control of the vehicle (50) based on the evaluated sensor data (11), and - a control unit (53) provided to said vehicle (50), wherein the KI-based information processing system (10) is tested by means of at least one test method (12-x) by means of a test device (2) arranged in a second control device (54) of the vehicle (50), and The test results (13-x) of the at least one test method (12-x) are stored in a multidimensional data structure (20) in a database (6) arranged in the vehicle (50) with reference to the tested KI-based information processing system (10) and the at least one test method (12-x) used, so that the data pool for checking and / or evaluating the KI-based information processing system is continuously expanded and / or updated, wherein the robustness of the KI-based information processing system is evaluated at a later point in time by using the test results stored in the multidimensional data structure.

2. The method according to claim 1, characterized in that The output (30) is at least one object classified from the sensor data (11) by the KI-based information processing system (10) or a control output (30).

3. The method according to claim 1 or 2, characterized in that: At least one piece of context information (15) of a context in which the sensor data (11) was detected is determined, wherein the at least one piece of context information (15) determined is stored in addition to the test result (13-x) in the multidimensional data structure (20).

4. The method according to claim 1 or 2, characterized in that: A property (16) of the detected sensor data (11) is determined, wherein the determined property (16) is stored in addition to the test result (13-x) in the multidimensional data structure (20).

5. The method according to claim 1 or 2, characterized in that: The at least one test method (12-x) comprises determining and / or verifying the consistency of the at least one output (30) over time.

6. The method according to claim 1 or 2, characterized in that: The at least one testing method (12-x) comprises determining and / or verifying a confidence level of the at least one output (30).

7. The method according to claim 1 or 2, characterized in that: The at least one testing method (12-x) comprises performing a plausibility check on the at least one output (30).

8. The method according to claim 1 or 2, characterized in that: The multidimensional data structure (20) is synchronized with a central multidimensional data structure (20) stored in a central database (41) on a central server (40) and / or with a decentralized multidimensional data structure (20) stored in a decentralized database (6).

9. The method according to claim 1 or 2, characterized in that: The KI-based information processing system (10) provides functions for automatic driving of the vehicle (50) and / or for driver assistance of the vehicle (50) and / or for environmental detection and / or environmental perception.

10. The method according to claim 1 or 2, characterized in that: The KI-based information processing system (10) is a neural network and / or includes at least one neural network.

11. The method according to claim 2, characterized in that The output (30) is a steering angle, a speed or a trajectory.

12. A device (1) for testing a KI-based information processing system (10) used in partially or fully automatic control of a vehicle (50), the device (1) comprising: at least one sensor (51) configured to provide sensor data (11) for environmental detection of the surroundings of the vehicle (50); a KI-based information processing system (10) arranged in a first control device (52) of the vehicle (50), the KI-based information processing system being configured to evaluate the detected sensor data (11) and to generate and provide at least one output (30) for partially or fully automatic control of the vehicle (50) based on the evaluated sensor data (11); a control unit (53) which uses the provided output (30) as a control parameter for partially or fully automatically controlling the vehicle (50); and a test device (2) arranged in a second control device (54) of the vehicle (50), the test device being configured to test the KI-based information processing system (10) by means of at least one test method (12-x), and storing the test results (13-x) of the at least one test method (12-x) in a multidimensional data structure (20) in a database (6) arranged in the vehicle (50) with reference to the tested KI-based information processing system (10) and the at least one test method (12-x) used, so that the data pool for testing and / or evaluating the KI-based information processing system is continuously expanded and / or updated, and on this basis, an evaluation of the robustness of the KI-based information processing system can be performed at a later point in time.

13. A system (80) for checking a KI-based information processing system (10) used when partially or fully automatically controlling a vehicle (50), the system comprising at least one device (1) according to claim 12 and a central server (40), wherein at least one multidimensional data structure (20) is stored in a central database (41) on the central server (40), and wherein the central server (40) is configured to generate the multidimensional data structure (20) and / or transmit the multidimensional data structure to the device (1) and / or synchronize the at least one multidimensional data structure (20) with the multidimensional data structure (20) of the at least one device (1).