Likelihood-driven reinforcement learning synergizing knowledge graph embedding and supervised pre-training for system-level fault localization

System-level fault localization remains a critical challenge in vehicle powertrain health management, hindered by complex fault propagation, data scarcity, and the "one/multiple-phenomenon, one/multiple-cause" problem. While data-driven methods excel at component-level diagnosis, their reliance on large annotated datasets and inability to model system topology limit their effectiveness for system-level applications. This paper proposes SPKL-RL, a likelihood-driven reinforcement learning framework that synergizes supervised pre-training with knowledge graph embedding to address these limitations. The framework formulates fault localization as a likelihood optimization within a Partially Observable Markov Decision Process (POMDP). A supervised pre-training module provides a warm-start policy using historical data, mitigating RL's cold-start issue. A knowledge graph embedding module encodes system topology and fault propagation into interpretable vectors. A likelihood-driven mechanism then guides exploration by integrating information gain and knowledge consistency into the reward function. Experimental results on a real powertrain dataset show SPKL-RL achieves 89.3% diagnostic accuracy across 15 fault types, reduces average diagnostic steps from 38 to 16 (58% efficiency gain), and significantly outperforms baselines, particularly in complex cascading (86.9% accuracy) and coupled fault (80.7% accuracy) scenarios. The framework demonstrates substantial advantages in reducing data dependency, improving generalization, and enhancing interpretability for complex system diagnostics.