Method and device for analyzing software errors after software updates

An artificial neuronal network system using RNNs and GNNs for vehicle systems addresses the inefficiencies of manual troubleshooting by systematically analyzing software errors, incorporating context, to efficiently identify and repair causes of malfunctions.

DE102024132166A1Pending Publication Date: 2025-01-16DR ING H C F PORSCHE AG
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
DE102024132166
Authority / Receiving Office
DE · DE
Patent Type
Applications
Current Assignee / Owner
Filing Date
2024-11-05
Publication Date
2025-01-16

AI Technical Summary

Technical Problem

Manual troubleshooting of software errors in complex vehicle systems is time-consuming and inadequate, and conventional diagnostic methods fail to quickly and precisely determine the causes of malfunctions, often ignoring contextual information like environmental conditions and operating states.

Method used

An artificial neuronal network is used to systematically record complex causes and sequence relationships between software changes and malfunctions, incorporating recurrent neural networks (RNNs like LSTM or GRU) and graph neural networks (GNNs) to model temporal sequences and interactions, with asymmetrical Shapley values explaining the predictions, and considering context information.

Benefits of technology

This approach enables efficient and precise error analysis, identifying root causes and reducing development delays by automating the process, enhancing interpretability and reliability of software updates.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000000_0000_ABST
    Figure 00000000_0000_ABST
Patent Text Reader

Abstract

Method for analyzing software errors (13) after software updates (10) of control units, characterized by the following features: Changes to the control units caused by the software updates (10) are documented, The relevant context (11) for the software errors (13) of the software updates (10) is described and The software errors (13) are analyzed in relation to the context (11) using an artificial neural network (12).
Need to check novelty before this filing date? Find Prior Art

Description

[0001] The present invention relates to a method for analyzing software errors after updates. The present invention also relates to a corresponding device, a corresponding computer program, and a corresponding storage medium. State of the art

[0002] In modern vehicle development, software plays a crucial role in controlling and integrating various electronic control units (ECUs). These ECUs are responsible for a wide range of functions, including engine management, vehicle safety, comfort systems, and entertainment electronics. Continuously updating and improving the software of these ECUs is essential to implement new features, optimize performance, and meet safety standards.

[0003] As is well known, software updates on control units are often performed using so-called flash processes. This involves replacing the existing software version with a new one. However, this process can be complex and carries the risk of problems and malfunctions. In particular, unexpected errors can occur after a software update, affecting not only the updated control unit but also other connected control units in the overall system. These malfunctions often manifest as unexplained system errors or as limitations in the functionality of individual components.

[0004] Various approaches are currently used to analyze and troubleshoot software errors. Traditionally, software errors are diagnosed manually by engineers and technicians. They use diagnostic tools and conduct tests to identify the root cause of a problem. Automated testing procedures are also known. These involve running standardized tests after a software update to verify the proper functioning of the control units.

[0005] With the advent of artificial intelligence methods, there have been increasing attempts to automate failure analysis. Machine learning and neural networks, in particular, have been used to identify patterns in large amounts of data and predict potential failures. Well-known applications in this area focus on predictive maintenance and anomaly detection in vehicle operation.

[0006] The use of sequential ("Seq2Seq") models to analyze temporal sequences of events is also well known. For example, feedback neural networks (RNNs) such as LSTMs or GRUs are used to capture dependencies in sequential data. These models have been successfully applied in various domains, such as speech recognition and predictive text processing.

[0007] In addition, models are known that use graph structures to represent communication and interaction between components in networked systems. These so-called graph neural networks (GNNs) make it possible to model and analyze relationships in graphs. Disclosure of the invention

[0008] One problem is that manual troubleshooting in such complex systems is extremely difficult and time-consuming. This process requires a deep understanding of the software architecture and the interactions between the various control units. Furthermore, conventional diagnostic tools and testing methods are often inadequate to quickly and accurately determine the causes of such malfunctions.

[0009] Another problem is that contextual information such as specific environmental conditions or special vehicle operating states are often not sufficiently considered when analyzing software errors. However, these factors can play a significant role in explaining why a particular software update causes problems in a real-world driving situation while showing no abnormalities in standardized test environments.

[0010] The inability to efficiently identify the cause-and-effect relationships between software changes and occurring malfunctions leads to significant delays in the development process. This not only has financial implications but can also delay the market launch of new vehicle models or features. Furthermore, there is a risk that potential safety risks will not be identified in a timely manner.

[0011] The described problem is solved by a method for analyzing software errors after updates, a corresponding device, a corresponding computer program and a corresponding storage medium according to the independent claims.

[0012] The presented approach enables a more efficient and precise analysis of software errors that occur after ECU updates. In contrast to the current state of the art, where troubleshooting is often manual and time-consuming, this approach relies on the use of an artificial neural network to systematically capture the complex cause-and-effect relationships between software changes and occurring malfunctions.

[0013] By documenting the changes caused by software updates and incorporating relevant contextual information such as environmental conditions or operating states, a comprehensive picture of potential influencing factors is created. This enables the neural network to recognize patterns and relationships that might be overlooked in conventional diagnostic procedures. In particular, the network can account for long-term dependencies within the data, which is crucial when analyzing interactions between networked control units.

[0014] The use of recurrent neural networks such as LSTMs or GRUs allows the temporal sequences of software updates and their effects to be modeled. By alternatively representing the communication between the control units as a graph and using graph neural networks, complex relationships and interactions within the overall system can also be captured.

[0015] In addition to the neural networks that model the relationships, the primary focus is on analyzing the source of the error. These networks are trained to detect the identified errors themselves based on the input data. If they can do so with a quality sufficient to meet the requirements (e.g., an accuracy of 90%), then asymmetric Shapley values ​​or other relevant methods are applied to explain the trained prediction models. For this purpose, it is analyzed which information in the inputs (source code, log messages, nodes in the topology graph) has the greatest influence on the prediction of the error. In this way, the root cause can be identified and corrected through software updates. This eliminates the need for manual analysis and tracing of errors and generally increases the interpretability of the overall system and the traceability of any errors.

[0016] Compared to state-of-the-art automated testing methods, which often only cover known error sources and fail to detect unforeseen interactions, this approach offers an adaptive and learning-capable solution. The neural network is trained with historical data from previous software updates and the resulting malfunctions, enabling it to learn from the underlying data streams as well as the code and its changes in the software update, and to make predictions about potential errors.

[0017] Taking contextual information such as environmental conditions and operating states into account makes it possible to better understand and prevent even situation-dependent errors. This represents a significant advance over previous methods, which often do not systematically incorporate such factors. Early detection and localization of potential malfunctions can accelerate development processes and increase the reliability of software updates.

[0018] Further advantageous embodiments of the invention are specified in the dependent patent claims. Short description of the drawings Fig. 1 shows a method according to the invention in the training phase. Fig. Figure 2 shows the procedure in the application phase. Embodiments of the invention

[0019] Fig. Figure 1 illustrates the process during the training phase, which lays the foundation for the subsequent analysis of software errors. Initially, software updates (10) are performed on the control units. These updates include specific changes to the ECU code, which are carefully documented. Documenting these changes enables subsequent context-related analysis and forms the basis for training the neural network (12).

[0020] In parallel, the relevant context (11) is described, which is important for the software updates and their potential impact on the overall system. Context (11) can include information about environmental conditions, operating states of the control units, or other factors that could influence the functionality of the control units. In an advantageous embodiment, the neural network is configured to capture the context itself using additional data.

[0021] After each software update (10), any software errors (13) and functional errors (14) in the control units are automatically detected. Software errors (13) refer to problems that result directly from the updated ECU code, while functional errors (14) affect the functions of the control units or their interactions within the overall system. This automated detection enables timely collection of data on emerging problems.

[0022] The collected data, consisting of the changes to the ECU code caused by the software update (10), the context (11), as well as the detected software errors (13) and functional errors (14), are fed to the neural network (12). The neural network (12) is trained with this data to learn patterns and relationships between the software updates, the context conditions, and the resulting errors. By comparing (15) the input data with the expected outputs, the model weights are adjusted, leading to a continuous improvement in prediction accuracy.

[0023] The neural network (12) can be implemented as a recurrent neural network to effectively model temporal sequences and long-term dependencies. In particular, architectures such as LSTM or GRU can be used to store and process information across multiple time steps. Alternatively or additionally, the communication between the interconnected control units can be represented as a graph. In this case, a graph neural network (GNN) is used to capture and analyze the complex interactions within the overall network.

[0024] Fig.Figure 2 illustrates the process during the application phase, whose input data is the same as in the training phase. Now, however, the previously trained neural network (12) is used to make predictions about potential software errors (13) following new software updates (10) based on the respective context (11). Based on the learned knowledge and the adaptive model weights, the neural network (12) can perform such predictive analyses. It identifies possible software errors (13) and functional errors (14) that could occur as a result of the software update (10). The network can also make statements about which functions or components are affected and what the possible causes of the errors are: In an advantageous embodiment, the network can, for example, highlight any faulty sections in the code when returning the code and, if necessary, suggest possible solutions, since the reason for the conflict and thus the error is known.These predictions make it possible to identify potential problems during the development phase and eliminate them as quickly as possible, thus accelerating the development process and minimizing unforeseen errors and downtime.

[0025] Through the continuous application of this method, the neural network (12) is continuously confronted with new data and can further refine its prediction capabilities. Taking into account context information (11), such as environmental conditions or specific operating states of the control units, helps to increase the accuracy of the analyses and also detect situation-dependent errors. List of reference symbols 10 Software update (ECU code) 11 Context 12 Neural network 13 software errors 14 functional errors 15 Comparison

Claims

[1] Method for analyzing software errors (13) after software updates (10) of control units, characterized by following features: - changes to the control units caused by the software updates (10) are documented, - the context (11) of the software updates (10) relevant to the software errors (13) is described and - the software errors (13) are analyzed in relation to the context (11) using an artificial neural network (12). [2] Method according to claim 1, characterized by following features: - during the software updates (10), any malfunctions (14) of the control units are automatically detected and - the network (12) is trained with the malfunctions (14) resulting from previous software updates (10). [3] Method according to claim 2, characterized by following feature: - the training includes a comparison (15) of the previous software updates (10) with the software errors (13) and functional errors (14). [4] Method according to one of claims 1 to 3, characterized by following feature: - the network (12) has feedback. [5] Method according to claim 4, characterized by at least one of the following characteristics: - the network (12) has a long short-term memory or - the network (12) comprises a GRU. [6] Method according to one of claims 1 to 5, characterized by following features: - the control units communicate with each other, - the communication is displayed as a graph and - the network (12) is a graph neural network. [7] Method according to one of claims 1 to 6, characterized by following features: - the description of the context (11) is based on environmental conditions or - the description of the context (11) is based on the operating states of the control units. [8] Device, characterized by following feature: - the device is designed to carry out a method according to one of claims 1 to 7. [9] Computer program which is designed to carry out all the steps of a method according to one of claims 1 to 7. [10] A machine-readable storage medium having stored thereon a computer program according to claim 9.