TLDR: SDVDiag is a new platform that automates the diagnosis of issues in connected and software-defined vehicles. It collects real-time data, builds dynamic dependency and causal graphs, and uses machine learning to continuously adapt its anomaly detection and root-cause analysis models. Evaluated in a 5G test environment, it successfully identifies and ranks the most likely causes of system failures, significantly reducing troubleshooting time and improving vehicle reliability.
Connected and software-defined vehicles (SDVs) are transforming the automotive industry, bringing advanced features like predictive maintenance and autonomous driving. However, their increasing complexity, reliance on intricate software stacks, and distributed nature across in-vehicle, edge, and cloud systems make diagnosing failures a significant challenge. When a vehicle function degrades or fails, pinpointing the exact root cause can be time-consuming and prone to human error, leading to prolonged downtimes and impacting user trust.
To address this critical issue, researchers have developed SDVDiag, an innovative and extensible platform designed for the automated diagnosis of connected vehicle functions. This platform streamlines the entire diagnostic process, from initial data collection to identifying the underlying root causes of system incidents.
How SDVDiag Works
SDVDiag operates through a modular architecture, integrating several key subsystems to provide comprehensive diagnostic capabilities:
Data Aggregation and Storage: This foundational layer continuously collects real-time data from all relevant sources, including vehicles, edge nodes, and cloud services. It gathers crucial information such as traces (documenting communication flow between components) and metrics (providing insights into system status and performance, like CPU load or memory usage).
Graph Creation: SDVDiag builds and maintains a dynamic, comprehensive overview of the system using graph-based data models. It constructs two primary types of graphs:
- Dependency Graphs: These are derived from live communication flows, showing how different services and instances interact. They reflect the dynamic behavior of the system, unlike static, pre-defined relationships.
- Causal Graphs: These go a step further by inferring cause-and-effect relationships between components. SDVDiag integrates causal discovery models, such as the Amortized Causal Discovery (ACD) framework, to quantify how changes or anomalies in one component influence others. These causal insights are then combined with the dependency graph to create an extended causal graph, where metric nodes are linked to their corresponding service instances.
Learning Environment: To ensure the platform remains adaptive and accurate despite frequent software updates or changes in system resources, SDVDiag includes a dedicated learning environment. This environment continuously updates the models used for causality mining and anomaly detection. It features a “Causal Model Encoder” that learns system dynamics from time-series data and an “Anomaly Detection Model Selector” that automatically chooses and trains the most suitable anomaly detection model for individual data streams. This continuous learning ensures robust performance and helps account for “concept drift” over time.
Incident Analysis: When an anomaly is detected (either automatically or triggered manually), SDVDiag creates a snapshot of the system’s current state. This snapshot includes the extended causal graph and all detected anomalies within a specific timeframe. The platform then performs root-cause analysis, often using algorithms like first-order random walks. This process involves traversing the graph along causal relationships, weighting paths by the strength of causal connections. The result is a ranked list of the most probable root causes, guiding system engineers to the source of the problem.
Evaluation and Impact
The effectiveness of SDVDiag was evaluated in a real-world 5G test fleet environment at the University of Stuttgart, utilizing a smart charging application. The platform successfully monitored the Kubernetes cluster and deployed vehicle services, collecting traces for dependency graphs and metrics for causal relationship inference. The Causal Model Encoder and Anomaly Detection Model Selector were trained on historical data and then deployed live.
During the evaluation, SDVDiag demonstrated its ability to reliably detect anomalies. The learning loop significantly improved detection accuracy over time, reducing false positives. In a simulated scenario involving increased CPU usage in a single charging station service, SDVDiag accurately isolated this as the root cause, even amidst complex system interactions and load balancing mechanisms. This highlights the platform’s potential to gain new insights and reduce downtime by identifying problems and their causes at an early stage.
Also Read:
- Optimizing Network Routing with AI: A Simulation-Driven Approach
- Bridging Intent and Action: Integrating Reasoning and Action Models for Automated Service Composition
Conclusion
SDVDiag represents a significant step forward in diagnosing complex software-defined vehicle environments. By combining distributed tracing and causality mining, it offers a modular and scalable solution for real-time diagnostics, anomaly detection, and precise root-cause analysis. Its adaptive learning capabilities ensure robust performance even with frequent system changes, making it a valuable tool for enhancing the reliability of connected vehicle systems. For more detailed information, you can refer to the full research paper: SDVDiag: A Modular Platform for the Diagnosis of Connected Vehicle Functions.
