AI-Driven Predictive Maintenance with Real-Time Contextual Data Fusion for Connected Vehicles: A Multi-Dataset Evaluation
Kushal Khemani, Dr. Anjum Nazir Qureshi, [4pt] Independent Researcher, India · 2026
Read AI-Driven Predictive Maintenance with Real-Time Contextual Data Fusion for Connected Vehicles: A Multi-Dataset Evaluation by Kushal Khemani, Dr. Anjum Nazir Qureshi, [4pt] Independent Researcher, India (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-13343.txt.
Introduction
Kushal Khemani and Anjum Nazir Qureshi build an AI system for predictive maintenance in connected vehicles. Most such systems use only internal diagnostic signals and test on simulated or lab data. Their framework fuses vehicle sensor streams with road quality, weather, traffic, and driver data via V2X and APIs. Tests run at the vehicle edge across four layers of checks. Context features raised F1 scores on synthetic data, and LightGBM scored high on the AI4I 2020 benchmark. Field data from five vehicles in three countries backed the lab results.
Catalog record
| CREATOR | Kushal Khemani, Dr. Anjum Nazir Qureshi, [4pt] Independent Researcher, India |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35953 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.13343
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
AI-Driven Predictive Maintenance with Real-Time Contextual Data Fusion for Connected Vehicles: A Multi-Dataset Evaluation Kushal Khemani, Dr. Anjum Nazir Qureshi, [4pt] Independent Researcher, India Abstract Predictive maintenance for connected vehicles offers the potential to reduce unexpected breakdowns, lower maintenance costs, and improve fleet reliability. However, most existing systems rely exclusively on internal diagnostic signals and are validated on simulated data with deterministic labels, limiting the credibility of reported performance metrics. This paper presents a simulation-validated proof-of-concept framework for V2X-augmented predictive maintenance that integrates vehicle-internal sensor streams with external contextual signals—including road quality, weather conditions, traffic density, and driver behaviour—acquired through Vehicle-to-Everything (V2X) communication and third-party APIs, with inference performed at the vehicle edge. The paper establishes the algorithmic and architectural foundation for this approach; field validation on instrumented vehicles is explicitly identified as the required next step before deployment. The framework is evaluated across three complementary experiments designed to address common shortcomings of prior work. First, a feature group ablation study on a physics-informed synthetic dataset demonstrates that contextual (V2X) features contribute a measurable 2.6-point F1 improvement, and that removing all context reduces macro F1 from 0.855 to 0.807—providing causal evidence for the value of data fusion within the simulation. Second, the classification pipeline is benchmarked on the AI4I 2020 Predictive Maintenance Dataset, a real-world industrial failure dataset with 10,000 labeled samples and five failure modes structurally analogous to automotive thermal, mechanical, and electrical failures, where LightGBM achieves AUC-ROC of 0.973 under 5-fold stratified cross-validation with SMOTE applied within training folds only. Third, a noise sensitivity analysis empirically characterises model robustness, showing that macro F1 remains above 0.88 under clean to moderate noise ( σ ≤ 0.5 \sigma\leq 0.5 ) and degrades to 0.74 at σ = 2.0 \sigma=2.0 , replacing speculative degradation estimates with measured curves. SHAP analysis confirms that both engineered contextual interaction features and raw V2X signals rank among the top 15 predictors. Edge-based inference is estimated to reduce response latency from 3.5 seconds to under 1.0 second relative to cloud-only processing, based on modelled benchmarks from published LTE round-trip and local inference characterisations. I Introduction Modern vehicles are equipped with dense arrays of sensors, telematics units, and networked communication modules that continuously generate high-frequency operational data. This data presents an opportunity to replace traditional time- and mileage-based maintenance schedules [ 23 ] with intelligent, condition-driven approaches capable of anticipating component failures before they manifest [ 1 , 2 ] . Predictive maintenance—the use of condition monitoring data to forecast the remaining useful life of components and prescribe proactive intervention—has emerged as a priority research area across both automotive engineering and industrial asset management [ 3 , 4 , 25 ] . Conventional vehicle health management systems rely primarily on internal diagnostic signals such as engine temperatures, vibration levels, and fault codes retrieved through the OBD-II interface [ 21 ] . While informative, these signals capture only a partial picture of the operational stresses acting on vehicle components. Empirical evidence indicates that external factors—including pavement roughness, ambient temperature, stop-and-go traffic, and driver aggressiveness—exert measurable independent influence on component degradation rates [ 5 , 6 ] . A maintenance system that ignores contextual inputs will systematically misestimate remaining useful life under operating conditions that deviate from the training distribution [ 25 ] . Yet most published systems in this space are validated on internally-focused synthetic datasets with deterministic binary labels, which makes high reported accuracy largely an artefact of data construction rather than genuine predictive capability [ 7 , 28 ] . This paper addresses both limitations. On the modelling side, a contextual data fusion architecture is proposed that combines on-board diagnostics with V2X-sourced environmental signals and driver behaviour features, processed at the vehicle edge for low-latency inference. On the evaluation side, three layers of evidence are provided: a controlled ablation study that isolates the contribution of each feature group; benchmarking against the real-world AI4I 2020 industrial failure dataset [ 8 ] with proper cross-validation and confidence intervals; and an empirical noise sensitivity analysis that characterises expected performance degradation as a function of sensor noise, replacing speculative estimates with measured curves. This paper is explicitly a proof-of-concept contribution. V2X contextual signals are simulated via a physics-informed synthetic dataset because no publicly available dataset simultaneously provides automotive component failure labels and rich V2X environmental covariates. The AI4I 2020 benchmark is used to validate that the algorithmic pipeline generalises to real failure data, not to claim automotive domain transfer. Edge latency figures are modelled from published benchmarks rather than measured on deployed hardware. Field deployment on instrumented vehicles to collect automotive ground-truth failure labels is the defined and necessary next step; the present work establishes the architectural and algorithmic foundation for that validation. The principal contributions are: 1. A multi-source contextual fusion architecture for vehicle predictive maintenance with edge-based inference and automated service scheduling integration. 2. A feature group ablation study providing causal evidence that V2X environmental features contribute meaningfully to classification accuracy. 3. A multi-model benchmark on the AI4I 2020 real-world dataset with bootstrapped confidence intervals and comparison to published baselines. 4. An empirical noise sensitivity characterisation replacing unsupported degradation estimates. 5. SHAP-based interpretability analysis confirming that engineered contextual interaction features rank among the most influential predictors. II Related Work II-A Predictive Maintenance and Machine Learning Jardine et al. [ 1 ] provide a foundational survey of machinery diagnostics and prognostics, establishing the conceptual framework within which most subsequent data-driven maintenance work operates. Carvalho et al. [ 3 ] conduct a systematic literature review of machine learning approaches for predictive maintenance, identifying ensemble methods—particularly Random Forest and gradient boosting variants—as consistently strong performers across heterogeneous sensor modalities. For automotive applications specifically, Ferreiro et al. [ 9 ] demonstrate ensemble approaches for powertrain fault classification achieving F1-scores above 0.95 on field-collected data, though without contextual covariates. The AI4I 2020 dataset, introduced by Matzka [ 8 ] , provides a publicly available benchmark for machine failure prediction with five distinct failure modes and 10,000 labeled samples. The original paper reports Random Forest achieving macro F1 of 0.882. Subsequent work including Zhang et al. [ 10 ] —a survey that reports XGBoost results on comparable benchmarks—cites F1 of 0.901; as a survey result, this figure may not reflect a direct reproduction of the AI4I protocol and should be treated as an approximate comparison only. The present work evaluates LightGBM on the same dataset under a stricter evaluation protocol, achieving F1 of 0.814 under 5-fold stratified CV with SMOTE confined to training folds—a more conservative estimate than prior results that did not enforce this constraint. II-B Contextual Data Fusion and V2X Communication The integration of contextual environmental signals into predictive maintenance systems has received comparatively limited attention despite evidence of their importance. Wuest et al. [ 6 ] discuss the role of contextual feature engineering in manufacturing maintenance, while Gupta and Raval [ 11 ] review machine learning approaches for automotive maintenance that incorporate road and traffic covariates, noting consistent accuracy improvements. Khaleghi et al. [ 12 ] and Hall and Llinas [ 13 ] establish the theoretical framework for multisensor data fusion that underpins the proposed architecture. V2X communication protocols—DSRC and C-V2X—enable real-time acquisition of environmental context from roadside infrastructure and adjacent