SDVDiag: multimodal AI finds root causes in software-defined vehicles 2.4x better

As cars become software-defined, failures propagate through service dependencies, making root cause analysis difficult—symptoms often appear multiple causal hops away from the actual defect. Existing tools rely on a single data modality (e.g., just logs or just metrics) and require manual, offline operation, which is insufficient for continuous driving. To address this, Weiß et al. introduce SDVDiag, a pipeline that combines log-based and metric-based service representations into a shared embedding space before constructing causal graphs. An anomaly-driven trigger automatically initiates diagnostics, turning a batch tool into a real-time online system.

Evaluated on an Autonomous Valet Parking testbed, SDVDiag produced sparser causal graphs than a metrics-only baseline (134 vs 182 edges on average). With human-feedback refinement, it outperformed the baseline by 2.4-fold after just 60 queries. In an end-to-end fault injection test, the system correctly recovered a root cause located two causal hops upstream of the observable symptom—demonstrating its ability to trace complex failure chains in modern vehicles.