An integrated deep learning approach for predictive vehicle maintenance

Abstract

In the automotive sector, vehicle data gathered through On-board Diagnostics (OBD) systems offers continuous insights into vehicle health status and performance. Leveraging this data for predictive maintenance can significantly reduce unplanned failures, enhance safety, and extend vehicle lifespan. This paper proposes a novel hybrid model for Predictive Maintenance (PdM), that integrates Long Short-Term Memory (LSTM) neural networks with K-means clustering to analyze unlabeled time-series data from OBD systems. Our main contribution is to integrate an unsupervised deep learning approach that effectively captures temporal dependencies and clusters operational patterns to predict engine condition with high accuracy, addressing the common challenge of unlabeled vehicle datasets. The model achieves state-of-the-art prediction performance with a 97.5% R 2 score of the selected feature, demonstrating its strong generalization and reliability in different domain applications. Compared to standalone LSTM, Gated Recurrent Units (GRUs) and Recurrent Neural Networks (RNNs) models, our hybrid approach outperforms traditional methods across all tested metrics, marking a significant advancement in predictive maintenance for vehicular systems. This work paves the way for smarter, real-time diagnostics in next-generation vehicles. • Develop a hybrid model combining sequence learning and clustering for predictive maintenance. • Leverage deep learning to detect vehicle faults and optimize maintenance schedules. • Enhance vehicle health predictions through temporal pattern recognition. • Validate model accuracy against traditional sequence-based approaches. • Integrate clustering to reveal hidden patterns in unlabeled vehicle data.