The integration of causal reasoning into the diagnostics of Electric Vehicle (EV) systems is increasingly critical for enhancing safety, interpretability, and predictive maintenance. This study systematically evaluates four prominent causal discovery algorithms Peter–Clark Algorithm (PC), Fast Causal Inference (FCI), Greedy Equivalence Search (GES), and DirectLiNGAM using a real-world EV telematics dataset comprising over 13,000 records. A domain-informed ground-truth causal graph was constructed to benchmark performance, and each algorithm was assessed using Structural Hamming Distance (SHD), precision, recall, and F1-score under both full-variable and partial-observability scenarios. The results show that constraint-based methods (PC, FCI) consistently outperform score-based (GES) and function-based (DirectLiNGAM) approaches, demonstrating superior recall, robustness, and stability under noisy and incomplete data conditions. GES exhibited structural inconsistencies owing to its sensitivity to Bayesian Information Criterion (BIC) scoring, whereas DirectLiNGAM struggled with violations of linearity and non-Gaussianity assumptions. Importantly, this study highlights the limitations of relying solely on SHD, as sparse graphs sometimes yielded deceptively favorable scores but poor recall. By bridging the gap between synthetic evaluations and real-world data, this study underscores the value of constraint-based methods for reliable causal modeling in sensor-rich, safety-critical EV applications. The findings contribute practical insights for researchers and engineers seeking interpretable and robust Causal AI solutions in intelligent transportation systems, while also pointing to future directions in temporal modeling and hybrid causal learning.