Abstract:
Conventional fleet maintenance relies on fixed schedules and reactive repairs that overlook the real operating context vehicles face—road quality, terrain, traffic, and weather. This thesis presents SmartFleet, a context-aware predictive maintenance framework that fuses spatial and temporal intelligence to anticipate failures and optimize service planning. The approach integrates Geographic Information Systems (GIS) with deep learning by coupling Long Short-Term Memory (LSTM) networks for multivariate time-series signals (usage, service history, and on-board metrics) and a Graph Neural Network (GNN) that encodes spatial context from road networks, elevation, congestion patterns, and climate. An ensemble layer reconciles the LSTM’s temporal forecasts with the GNN’s location-informed risk scores to produce a unified Vehicle Health Index and maintenance lead-time predictions.
The system is implemented end-to-end: data ingestion and cleaning pipelines harmonize heterogeneous logs with GIS layers; a Flask API serves inference and explanations; and a React dashboard provides role-based access to alerts, cost projections, and drill-down diagnostics. Model explainability (e.g., SHAP-based feature attributions) surfaces which environmental factors and usage patterns drive risk, supporting transparent decisions. Evaluated against schedule-based and single-model baselines, SmartFleet reduces false alarms, improves early-warning horizons, and aligns interventions with real vehicle stress profiles, yielding fewer unplanned breakdowns and more efficient parts and labor allocation. Sensitivity analyses demonstrate robustness across routes and seasons, while ablations confirm the incremental value of spatial features.
SmartFleet contributes: (i) a practical method to fuse GIS with deep temporal learning for fleet health prediction; (ii) a deployable LSTM+GNN ensemble with operational tooling and interpretability; and (iii) empirical evidence that context-aware maintenance planning outperforms traditional policies on accuracy and operational usefulness. The results indicate a viable pathway from raw data to proactive, cost-aware decisions in intelligent fleet management.
