Method and system for diagnosing at least one safety function for controlling a system architecture in a vehicle

Abstract

The disclosure includes a method (100) for diagnosing at least one safety function for controlling a system architecture (55) in a vehicle (50) comprising the following steps: - Generating (102) a first meta-information (10) for a first security function (12) based on at least first input data (14), wherein the first meta-information (10) contains a first confidence value (16) which represents a first diagnostic result for the first security function (12); - Generating (104) a second meta-information (20) for a second AI-based security function (22) based on at least two input data (24) containing at least one machine learning model (30), wherein the second meta-information (20) contains a second confidence value (26) representing a second AI-based diagnostic result (27) based on at least one AI-based diagnostic function (28) for the second AI-based security function (22), wherein the second confidence value (26) further includes information about the relevance of the at least one AI-based diagnostic function (28) based on at least one system architecture-relevant requirement parameter (29); - Generating (106) an overall confidence (40) for the two security functions (12, 22) by fusing the two security functions (12, 22), weighting the first security function (12) and the second AI-based security function (22) on the basis of the respective generated meta-information (10, 20); - Transmitting (108) the generated total confidence (40) to a control unit (52) in order to generate a corresponding safety-related instruction (53) for controlling the system architecture (55) of the vehicle (50).

Description

[0001] The present invention relates to a method and a system for diagnosing at least one safety function for controlling a system architecture in a vehicle. State of the art

[0002] State-of-the-art security architectures for automated systems in an open context, such as automated driving, industrial robots or automated machines, are known and are increasingly based on AI-based algorithms.

[0003] These so-called hybrid systems employ a variety of different sensors, which can include sensors with AI-based components as well as sensors without AI-based components, to process information from the system's environment captured by a sensor with elements of artificial intelligence. Particularly for highly automated vehicle driving, the merging or fusion of information captured by various sensors plays a significant role in ensuring the safe control of the system or vehicle.

[0004] However, with these known systems with AI-based components, it is not yet possible to fully diagnose the AI-based algorithms used during operation for weaknesses or functional deficiencies and to systematically evaluate these deficiencies in order to then use these results in a targeted manner for optimized and improved safety-relevant control of these systems.

[0005] It is therefore the object of the present invention to provide a solution by means of which the performance of a system architecture in a vehicle can be improved in an efficient and reliable manner. Disclosure of the invention

[0006] This object is achieved by a method and a system for diagnosing at least one safety function for controlling a system architecture in a vehicle, having the features of the independent claims.

[0007] According to a first aspect, the disclosure relates to a method for diagnosing at least one safety function for controlling a system architecture in a vehicle, which method comprises the following steps: In a first step, a first meta-information for a first safety function is generated on the basis of at least first input data, wherein the first meta-information contains a first confidence value which represents a first diagnostic result for the first safety function.

[0008] In a second step, a second piece of meta information is generated for a second AI-based safety function on the basis of at least second input data, which contains at least one machine learning model, wherein the second meta information contains a second confidence value that represents a second AI-based diagnostic result based on at least one AI-based diagnostic function for the second AI-based safety function, wherein the second confidence value further contains information about a relevance of the at least one AI-based diagnostic function based on at least one system architecture-relevant requirement parameter.

[0009] In a third step, an overall confidence for the two security functions is generated by merging the two security functions, whereby the first security function and the second AI-based security function are weighted based on the meta-information generated.

[0010] In a fourth step, the generated overall confidence is transmitted to a control unit in order to generate a corresponding safety-related instruction for controlling the vehicle's system architecture.

[0011] A fundamental idea of ​​the present invention is that a method and a corresponding safety architecture are provided to combine safety-related calculation results of AI-based functions with safety-related calculation results of other classic model-based or AI-based functions.

[0012] The method and system according to the invention advantageously enable AI-based algorithms to be fully diagnosed during operation for weaknesses or functional deficiencies and, based on this, to determine a so-called aggregated confidence value of the calculation result.

[0013] Furthermore, AI-based calculation results for safety functions are weighted according to their confidence value in a hybrid system architecture.

[0014] In addition, the method according to the invention and the system simultaneously offer a mitigation option or a functional fallback if the determined confidence value falls below a defined limit.

[0015] The present invention contains the following essential building blocks or elements: 1. The reliability (= integrity, confidence) of the calculation result of the AI-based function used in the system is determined on the basis of different diagnoses focused on the typical weaknesses of AI functions, 2. The determined reliability is passed on to a fusion block as so-called metainformation, 3. In the fusion block, the calculation results of the AI-based functions and the other functions are fused, weighted according to the metadata. A low weight is selected for low confidence. 4. The overall system confidence determined from the various calculation results is also passed on by the fusion as meta-information to a processing safety-critical function (e.g. actuator control), which uses this information (“Uncertainty-Aware”), e.g. to determine the strength of a safety-critical intervention in the system architecture of the vehicle.

[0016] The following advantages are achieved with the method and system according to the invention: - The various weaknesses that AI-based functions can exhibit are specifically detected and mitigated by low weighting of the calculation result at the system level (fusion), so that they cannot have dangerous effects. - If the confidence of the AI-based function is high (and possibly supported by the other function paths), its result can be used, thereby increasing the performance of the system (especially for functions in which AI-based functions have higher performance, such as environment detection), without compromising security. - By combining it with various other functional paths (AI-based or model-based) that are not affected by the same weaknesses, the availability of the overall system is still maintained at a high level - if only one path, or in particular the AI-based path, is "uncertain" or produces a calculation result with low confidence, the other paths can compensate for this or be used as a fallback. - By sharing and evaluating the overall system confidence in the “Uncertainty-Aware” safety function, safety-critical system reactions based on unreliable information can be prevented.

[0017] A possible embodiment of the method provides that the at least one AI-based diagnostic function generates a first AI-based diagnostic information for the second AI-based safety function about whether the second AI-based safety function to be tested is suitable for generating a result with a defined quality, for example expressed by the corresponding confidence value, with the second input data on the basis of predetermined environmental information or context information, wherein if the first AI-based diagnostic information deviates from a defined quality, the first AI-based diagnostic information is taken into account in the step of generating the second meta information.This creates the advantage of efficiently determining whether the machine learning module can handle the provided input data appropriately and whether it is able to provide reliable information for further processing.

[0018] One possible embodiment of the method provides for the context information to be transmitted, at least partially, to the at least one AI-based diagnostic function via an external data source. This creates the advantage that context-based information is provided by data sources, allowing existing resources of the system architecture to be efficiently used for tasks other than determining context information.

[0019] A possible embodiment of the method provides that the at least one AI-based diagnostic function generates a second AI-based diagnostic information item that indicates whether the at least one machine learning model of the AI-based safety function is being executed correctly. If the second AI-based diagnostic information item deviates from a defined quality, the second AI-based diagnostic information item is taken into account in the step of generating the second meta-information item. This provides the advantage of efficiently checking whether the machine learning module is being executed correctly and whether technical specifications, such as calculation times, have been met.

[0020] A possible embodiment of the method provides that the at least one AI-based diagnostic function generates a third AI-based diagnostic information item that indicates whether a calculation result exhibits an anomaly during operation of the at least one machine learning model. If the third AI-based diagnostic information item deviates from a defined quality, the third AI-based diagnostic information item is taken into account in the step of generating the second meta-information item. This provides the advantage that the results of the machine learning module are efficiently checked during operation, and potential anomalies are incorporated as information into the determination of the meta-information item, which improves the quality of the generated meta-information item.

[0021] A possible embodiment of the method provides that the at least one AI-based diagnostic function generates a fourth piece of AI-based diagnostic information containing information about at least one confidence value or one confidence metric, wherein the fourth piece of AI-based diagnostic information is taken into account in the step of generating the second piece of meta-information. This provides the advantage that specific safety KPIs (= Key Performance Indicators) or metrics can be directly determined during operation in an efficient manner. These metrics are used to determine the meta-information and improve the quality of the generated meta-information.

[0022] One possible embodiment of the method provides for the first metadata to be generated using the second input data as well. This has the advantage of improving the quality of the first metadata.

[0023] According to a second aspect, the disclosure relates to a system for diagnosing at least one safety function for controlling a safety architecture in a vehicle, comprising: - a non-KI-based safety module configured to generate first meta information for a first non-KI-based safety function, wherein the first meta information contains a first confidence value representing a non-KI-based diagnostic result for the first non-KI-based safety function; - a KI-based security module configured to generate second meta information for a second AI-based security function containing at least one machine learning model, wherein the second meta information contains a second confidence value representing an AI-based diagnostic result generated by a KI-based diagnostic module based on at least one AI-based diagnostic function for the second AI-based security function; - a fusion module designed to generate an overall confidence for the two security functions by fusing the two security functions, wherein the first non-AI based security function and the second AI based security function are weighted on the basis of the meta information generated respectively; and a - a control module which is designed to generate a corresponding safety-related instruction for controlling the system architecture of the vehicle, taking into account the generated overall confidence.

[0024] According to a third aspect, the disclosure relates to a computer program containing machine-readable instructions which, when executed on one or more computers and / or computer instances, cause the computer or computer instances to carry out the method according to the invention.

[0025] According to a fourth aspect, the disclosure relates to a machine-readable data carrier and / or download product comprising the computer program.

[0026] According to a fifth aspect, the disclosure relates to one or more computers and / or computer instances with the computer program, and / or with the machine-readable data carrier and / or the download product.

[0027] Further measures improving the invention are presented in more detail below together with the description of the preferred embodiments of the invention with reference to figures. Examples of implementation

[0028] It shows: Fig. 1 Schematic flow diagram of method 100 for diagnosing at least one safety function for controlling a system architecture in a vehicle; Fig. 2 Schematic representation of a system architecture for carrying out the method 100 according to an embodiment of the present invention, Fig. 3a Schematic representation of a system architecture for carrying out the method 100 according to an embodiment of the present invention, Fig. 3b Schematic representation of a system architecture for carrying out the method 100 according to an embodiment of the present invention, and Fig. 4 Schematic representation of a system for performing the method 100 according to an embodiment of the present invention.

[0029] Fig. 1 Schematic flow diagram of the method 100 for diagnosing at least one safety function for controlling a system architecture 55 in a vehicle 50.

[0030] In a first step 102, a first piece of meta information 10 is generated for a first safety function 12 on the basis of at least first input data 14, wherein the first piece of meta information 10 contains a first confidence value 16 which represents a first diagnostic result for the first safety function 12.

[0031] In a second step 104, a second piece of meta information 20 is generated for a second AI-based safety function 22 on the basis of at least second input data 24, which contains at least one machine learning model 30, wherein the second piece of meta information 20 contains a second confidence value 26, which represents a second AI-based diagnostic result 27 based on at least one AI-based diagnostic function 28 for the second AI-based safety function 22, wherein the second confidence value 26 further contains information about a relevance of the at least one AI-based diagnostic function 28 on the basis of at least one system architecture-relevant requirement parameter (29).

[0032] In a third step 106, an overall confidence 40 is generated for the two security functions 12, 22 by merging the two security functions 12, 22, wherein the first security function 12 and the second AI-based security function 22 are weighted based on the respectively generated meta-information 10, 20. Within the meaning of the present invention, overall confidence can also be understood as the following, but not exhaustive, list of terms, such as: plausibility ranking (possibility theory), belief (Dempster-Shafer theory), likelihood, distance measures (Mahalanobis), covariance (Bayes theory), etc.

[0033] In a fourth step 108, the generated total confidence 106 is transmitted 108 to a control unit 52 in order to generate a corresponding safety-related instruction 53 for controlling the system architecture 55 of the vehicle 50.

[0034] Fig. 2 shows a schematic representation of a hybrid system architecture 55 for carrying out the method 100 according to an embodiment of the present invention.

[0035] Based on first input data 14, a first piece of metainformation 10 is generated for a first safety function 12. The first piece of metainformation 10 contains a first confidence value 16, which represents a first diagnostic result for the first safety function 12. This first safety function 12 can represent a non-AI-based or an AI-based safety function. The first confidence value 16 represents a first contribution to the overall confidence 40 to be formed, which is generated in the fusion module 65.

[0036] Based on second input data 24, a second piece of meta information 20 is formed for a second AI-based safety function 22. This contains a machine learning model 30. The second piece of meta information 20 contains a second confidence value 26, which represents a second AI-based diagnostic result 27 based on at least one AI-based diagnostic function 28 for the second AI-based safety function 22.

[0037] In the fusion module 65, a total confidence 40 is then formed for the two security functions 12, 22 by merging the two security functions 12, 22 from the first confidence value 16 and the second confidence value 26. In addition, the first security function 12 and the second AI-based security function 22 are weighted based on the meta-information 10, 20 respectively generated.

[0038] The second AI-based safety function 22 can also be regarded as an AI-related diagnostic function 28.

[0039] The second confidence value 26 further includes information about a relevance of the at least one AI-based diagnostic function 28 based on at least one system architecture-relevant requirement parameter 29. This requirement parameter 29 can, for example, include information about a timing for providing data, a data integrity, a requirement for data quality or selected functional contents.

[0040] With regard to the above-mentioned information on relevance, it can also be said as follows: Instead of calculating only a single confidence value related to a so-called result date, an additional distinction is made based on the exact evaluation of the diagnoses as to which functional (functional content, data quality) and non-functional system requirements are affected by negative diagnostic results (e.g. timing, integrity), so that the subsequent fusion has more freedom for weighting.

[0041] In the Fig. 2, the AI-based diagnostic function 28 can generate AI-based diagnostic information 23 via a so-called context monitor 41 or context module. This AI-based diagnostic information 23 provides a statement as to whether the second AI-based safety function 22 to be tested is suitable for generating a result with a defined quality using the second input data 24 on the basis of predetermined environmental information or context information 35. If the first AI-based diagnostic information 23 deviates from a defined quality, the first AI-based diagnostic information 23 is taken into account in step 104 of generating the second meta-information 20.

[0042] The context monitor 41 thus determines during operation whether the machine learning module (MLM) can handle the input data at all, e.g., whether it was actually developed for the given context and is fundamentally capable of providing reliable information. An example of this could be an "out-of-distribution" monitor, which determines whether the input data matches the statistical distribution of the data used to train the MLM.

[0043] However, the context can also be checked explicitly, e.g. by determining specific critical context parameters (e.g. sensor quality, brightness, weather, traffic situation, external object constellation, etc.) from the input data.

[0044] The context information 35 regarding whether the MLM is suitable for the context can also be transmitted, at least in part, to the at least one AI-based diagnostic function 28 via an external data source or an external system (context provider) 36. This can be, for example, a function within the system (e.g., in a highly automated vehicle) that only has the function of determining the context and / or the current scenario, or an external system such as cloud / backend, edge, infrastructure, or other nearby systems that provides this information. If a context deviation is detected, this information is passed on to the Integrity & Confidence Assessor 45.

[0045] The Integrity & Confidence Assessor 45 combines the individual diagnostic results 23, 31, 32, 33 from the individual functional paths in order to generate corresponding meta information 20 with the second confidence value 26, which is then passed on to the fusion module 65.

[0046] This meta information can contain, for example, the following content - a binary information “reliable” or “not reliable” - a single continuous value in the range [0,1] to indicate the reliability or confidence of the calculation result, such as Belief (Dempster Shafer) - a graded value, e.g. fuzzy (“very reliable”, “reliable”, “not reliable”), to indicate the reliability or confidence of the calculation result, such as a so-called plausibility ranking (possibility theory) - an array with the individual results of the diagnoses - Information on the basis of which information the reliability or confidence classification was made (e.g. also the number and / or integrity (e.g. ASIL) of the input data).

[0047] The results can be combined, for example, by - OR rating: if any diagnosis detects a problem, it results in a rating of “not reliable”. - MIN / MAX rating: the most serious, critical error is passed on - Fuzzy evaluation: by combining the individual results (“very reliable”, “reliable”, “not reliable”,...). - All individual results will be forwarded.

[0048] The Fig. Figure 2 further shows a fusion module 65. The calculation results of the various functional paths, where the number of different functional paths can be greater than 3, are fused in a weighted manner according to the "safety metadata" or confidence rating. For example, - Prioritization of the Cl-based calculation result if the reliability or confidence exceeds a threshold (e.g., to achieve the best performance in addition to safety) - Check whether another function path, in addition to the AI-based function, also produced the same result. If so, this result is adopted (again, to ensure both reliability and optimal performance). For example, if the confidence value is too low, a further / better result could be requested externally, e.g., via a V2X interface. - Only the n most reliable (highest confidence) calculation results are used for the fusion (best-of filter) - The calculation results are fused based on their confidence weights (e.g. with a Bayes filter) - The final calculation result is determined from the individual calculation results using a “majority vote”, whereby the calculation results are also weighted according to their confidence.

[0049] The result of the fusion consists of the fusion result of the data and again also “safety metadata” or summarized confidence information 67, which can result from the fusion of the (weighted) data itself or from the fusion of the respective confidence information of the individual function paths.

[0050] The Fig. Figure 2 further shows the system architecture 55, which uses the results of "information fusion" and can, for example, directly or indirectly control safety-critical actuators (e.g., planning for highly automated systems, engine control, other actuator control). This system architecture also has the ability to evaluate the "safety metadata" or summarized or final confidence information 67 and to modulate or limit the safety-critical control of the control module 66 of the vehicle 55 accordingly. This can be done, for example, as follows: - High confidence: no limitation - Medium confidence: moderate limitation, e.g. maximum “ASIL B intervention” instead of “ASIL D intervention” (e.g. maximum 6m / s^2 braking intervention instead of even greater than 6m / s^2 braking intervention or minimum 1.5 meters distance to potentially detected cyclist instead of only 1 meter) - Low confidence: strong limitation (e.g. maximum 3.5 m / s^2 braking intervention) or even transition to the safe state, e.g. “Minimal Risk Manoeuvre11” in highly automated systems

[0051] In addition to such a discrete interpretation of confidence, it can also be used as an additional probabilistic parameter in probabilistic functions, e.g., as an increase in the covariance of the input data in a Bayesian filter or as an increasing factor in criticality calculations (e.g., estimation of collision risk, time-to-collision, near miss or ADS safety measures12, surrogate safety measures, etc.).

[0052] Further referring to Fig. 2, the AI-based diagnostic function 28 can further generate a second AI-based diagnostic information 31 via a so-called runtime monitor 42. This indicates whether the at least one machine learning model 30 of the AI-based safety function 22 is being executed correctly. If the second AI-based diagnostic information 31 deviates from a defined quality, the second AI-based diagnostic information 31 is taken into account in the step of generating 104 the second meta-information 20.

[0053] The Runtime Monitor 42 checks whether the MLM is being executed correctly, in particular whether permanent or transient HW, SW or operating system errors have occurred, e.g. whether calculation times and calculation sequences have been adhered to, etc. If errors are detected, they are forwarded to the “Integrity & Confidence Assessor” 45.

[0054] In the Fig. 2, the AI-based diagnostic function 28 can further generate a third AI-based diagnostic information 32 via a so-called anomaly detector 43. This indicates whether a calculation result during operation of the machine learning model 30 exhibits an anomaly. If the third AI-based diagnostic information 32 deviates from a defined quality, the third AI-based diagnostic information 32 is taken into account in step 104 of generating the second meta-information 20.

[0055] The anomaly detector 43 checks the MLM results for anomalies during operation – possibly also by analyzing a temporal sequence of results (e.g., time series). If anomalies are detected, this information is also forwarded to the "Integrity & Confidence Assessor" 45. Fig. 2, the AI-based diagnostic function 28 can further generate, via a so-called Safety KPI Monitor 44, a fourth AI-based diagnostic information 33 which contains information about at least one uncertainty metric, wherein the fourth AI-based diagnostic information 33 is taken into account in step 104 of generating the second meta-information 20.

[0056] The Safety KPI Monitor 44 determines specific safety KPIs (= safety-related performance indicators) or metrics directly in operation.

[0057] These can be determined, for example, either - parallel through additional calculations or - from the calculation result or - also by using additional “test patterns” that are given at the entrance of the MLM (see dotted line 46 in Fig. 2, Fig. 3a, Fig. 3b)

[0058] Uncertainty metrics could be, for example, - SoftMax - Mean temporal coherence - Autoencoder reconstruction error - Expected calibration error - Noise Robustness (e.g. by imposing noise on the input data)

[0059] These safety KPIs or metrics are then passed on to the “Integrity & Confidence Assessor” 45.

[0060] The generated total confidence 40 is then transmitted to a control unit 52 of the vehicle 50 in order to generate a corresponding safety-related instruction 53 for controlling the system architecture 55 of the vehicle 50.

[0061] Fig. 3a shows a schematic representation of a system architecture 55 for implementing the method 100 according to an embodiment of the present invention. The difference from the embodiment according to Fig. 2 consists in that the first meta-information 10 is now generated using the second input data 24. For this purpose, the second input data 24 is used as additional input for the first security function 12.

[0062] Fig. 3b shows a schematic representation of a system architecture 55 for implementing the method 100 according to another embodiment of the present invention. The difference from the embodiment according to Fig. 3a consists in the fact that further first input data 15 are present, which are fed into the first safety function 12 separately or separately from the first input data 14.

[0063] Fig. 4 shows a schematic representation of a system 60 for diagnosing at least one safety function for controlling a safety architecture 55 in a vehicle 50 according to an embodiment of the present invention.

[0064] The System 60 has the following components: - a non-KI-based security module 61 configured to generate first meta information 10 for a first non-KI-based security function 12, wherein the first meta information 10 contains a first confidence value 16 representing a non-KI-based diagnostic result for the first non-KI-based security function 12; - an AI-based security module 63, which is designed to generate a second piece of meta information 20 for a second AI-based security function 22, which contains at least one machine learning model 30, wherein the second piece of meta information 20 contains a second confidence value 26, which represents an AI-based diagnostic result, which is generated by means of an AI-based diagnostic module 63 on the basis of at least one AI-based diagnostic function 64 for the second AI-based security function 22; - a fusion module 65, which is designed to generate an overall confidence 40 for the two security functions 12, 22 by merging the two security functions 12, 22, wherein the first non-AI-based security function 12 and the second AI-based security function 22 are weighted on the basis of the respectively generated meta information 10, 20; and a - a control module 66 which is designed to generate a corresponding safety-related instruction 53 for controlling the system architecture 55 via a control unit 52 of the vehicle 50, taking into account the generated overall confidence 40.

Claims

[1] Method (100) for diagnosing at least one safety function for controlling a system architecture (55) in a vehicle (50), comprising the following steps: - generating (102) a first meta-information (10) for a first safety function (12) on the basis of at least first input data (14), wherein the first meta-information (10) contains a first confidence value (16) representing a first diagnostic result for the first safety function (12); - generating (104) a second piece of meta information (20) for a second AI-based safety function (22) on the basis of at least second input data (24) which contains at least one machine learning model (30), wherein the second piece of meta information (20) contains a second confidence value (26) which represents a second AI-based diagnostic result (27) on the basis of at least one AI-based diagnostic function (28) for the second AI-based safety function (22), wherein the second confidence value (26) further contains information about a relevance of the at least one AI-based diagnostic function (28) on the basis of at least one system architecture-relevant requirement parameter (29); - generating (106) an overall confidence (40) for the two security functions (12, 22) by merging the two security functions (12, 22), wherein the first security function (12) and the second AI-based security function (22) are weighted on the basis of the respectively generated meta information (10, 20); - transmitting (108) the generated total confidence (40) to a control unit (52) in order to generate a corresponding safety-related instruction (53) for controlling the system architecture (55) of the vehicle (50). [2] Method (100) according to claim 1, wherein the at least one AI-based diagnostic function (28) generates a first AI-based diagnostic information (23) for the second AI-based safety function (22) about whether the second AI-based safety function (22) to be tested is suitable for generating a result with a defined quality using the second input data (24) on the basis of a predetermined environmental information or context information (35), wherein in the event of a deviation of the first AI-based diagnostic information (23) from a defined quality, the first AI-based diagnostic information (23) is taken into account in the step (104) of generating the second meta-information (20). [3] Method (100) according to claim 2, wherein the context information (35) is transmitted at least partially via an external data source (36) to the at least one AI-based diagnostic function (28). [4] Method (100) according to one of the preceding claims, wherein the at least one AI-based diagnostic function (28) generates a second AI-based diagnostic information (31) which indicates whether the at least one machine learning model (30) of the AI-based safety function (22) is executed correctly, wherein in the event of a deviation of the second AI-based diagnostic information (31) from a defined quality, the second AI-based diagnostic information (31) is taken into account in the step of generating (104) the second meta-information (20). [5] Method (100) according to one of the preceding claims, wherein the at least one AI-based diagnostic function (28) generates a third AI-based diagnostic information (32) which indicates whether a calculation result in the operation of the at least one machine learning model (30) has an anomaly, wherein in the event of a deviation of the third AI-based diagnostic information (32) from a defined quality, the third AI-based diagnostic information (32) is taken into account in the step (104) of generating the second meta information (20). [6] Method (100) according to one of the preceding claims, wherein the at least one AI-based diagnostic function (28) generates a fourth AI-based diagnostic information (33) containing information about at least one uncertainty metric, wherein the fourth AI-based diagnostic information (33) is taken into account in the step (104) of generating the second meta-information (20). [7] Method (100) according to one of the preceding claims, wherein the first metainformation (10) is also generated using the second input data (24). [8] System (60) for diagnosing at least one safety function for controlling a safety architecture (55) in a vehicle (50), comprising: - a non-KI-based security module (61) configured to generate first meta information (10) for a first non-KI-based security function (12), wherein the first meta information (10) contains a first confidence value (16) representing a non-KI-based diagnostic result for the first non-KI-based security function (12); - an AI-based security module (63) which is designed to generate a second piece of meta information (20) for a second AI-based security function (22) which contains at least one machine learning model (30), wherein the second piece of meta information (20) contains a second confidence value (26) which represents an AI-based diagnostic result which is generated by means of an AI-based diagnostic module (63) on the basis of at least one AI-based diagnostic function (64) for the second AI-based security function (22); - a fusion module (65) which is designed to generate an overall confidence (40) for the two security functions (12, 22) by fusing the two security functions (12, 22), wherein the first non-AI-based security function (12) and the second AI-based security function (22) are weighted on the basis of the respectively generated meta information (10, 20); and a - a control module (66) which is designed to generate a corresponding safety-related instruction (53) for controlling the system architecture (55) of the vehicle (50) taking into account the generated overall confidence (40). [9] A computer program containing machine-readable instructions which, when executed on one or more computers and / or computer instances, cause the computer or computer instances to control the method according to any one of claims 1 to 7 and / or the system according to claim 8. [10] Machine-readable data carrier and / or download product with the computer program according to claim 9. [11] One or more computers and / or computer instances with the computer program according to claim 9, and / or with the machine-readable data carrier and / or download product according to claim 10.

Images