AI-Driven Predictive Maintenance with Real-Time Contextual Data Fusion for Connected Vehicles: A Multi-Dataset Evaluation

Abstract

Most vehicle predictive maintenance systems rely exclusively on internal diagnostic signals and are validated on deterministic synthetic data, limiting the credibility of reported metrics. This paper presents a simulation-validated proof-of-concept framework for V2X-augmented predictive maintenance, integrating on-board sensor streams with external contextual signals -- road quality, weather, traffic density, and driver behaviour -- acquired via V2X communication and third-party APIs, with inference at the vehicle edge. Field validation on instrumented vehicles is identified as the required next step. Three experiments address common shortcomings of prior work. A feature group ablation study shows that V2X contextual features contribute a 2.6-point F1 gain, with full context removal reducing macro F1 from 0.855 to 0.807. On the AI4I 2020 real-world industrial failure dataset (10,000 samples, five failure modes), LightGBM achieves AUC-ROC of 0.973 under 5-fold stratified CV with SMOTE confined to training folds. A noise sensitivity analysis shows macro F1 remains above 0.88 under low noise and degrades to 0.74 under very high noise. SHAP analysis confirms that V2X and engineered interaction features rank among the top 15 predictors. Edge inference is estimated to reduce latency from 3.5s to under 1.0s versus cloud-only processing.

Figures (7)

• Figure 1: Feature group ablation study. (a) Absolute macro F1-score per configuration with 95% error bars. (b) F1 drop relative to the full model when each feature group is removed. Environmental (V2X) removal causes the second largest contextual drop ($-$0.026), confirming meaningful predictive contribution from V2X-sourced features beyond internal mechanical state alone.
• Figure 2: AI4I 2020 real-world benchmark results. (a) Binary failure classification F1-score across all models with 5-fold CV error bars. (b) AUC-ROC comparison. (c) Confusion matrix for LightGBM on the last CV fold. (d) Per-failure-mode F1-scores; Random Failure ($n{=}19$) shows near-random performance as expected. (e) Class distribution showing 3.4% failure rate addressed by SMOTE.
• Figure 3: Multi-model classification benchmark on the synthetic contextual dataset. LightGBM achieves the highest AUC-ROC (0.949) and macro F1 (0.837). Note that results are obtained on synthetic data; the AI4I 2020 benchmark (Section \ref{['sec:results']}) provides real-world validation.
• Figure 4: SHAP feature importance for the LightGBM classifier. (a) Top 15 features by mean $|$SHAP value$|$. (b) Beeswarm plot showing directional feature effects; colour represents normalised feature value (red = high, blue = low). Four contextual/interaction features appear in the top 9, confirming the value of V2X data fusion.
• Figure 5: Model robustness to increasing sensor noise. F1 remains above 0.88 for $\sigma \leq 0.5$ and degrades gracefully at higher noise levels. The shaded band represents $\pm 1$ standard deviation across 5 seeds. AUC-ROC (dashed red) degrades more slowly than F1, suggesting the model retains ranking ability even under heavy noise.