Methodology for Commercial Vehicle Mechanical Systems Maintenance: Data-Driven and Deep-Learning-Based Prediction

Abstract

This paper presents a predictive maintenance (PdM) strategy for commercial vehicles, focusing on the turbocharger—a critical yet often under-monitored component. By combining sensor signals, workshop maintenance logs, and technical specifications, the study demonstrates how data-driven deep-learning techniques can robustly identify pending failures. Specifically, Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM) architectures were employed to capture temporal dependencies and detect patterns that conventional approaches and purely onboard monitoring might overlook. Results on real-world fleet data indicate that BiLSTM achieved higher recall (98.65%) and a lower cost-score than standard LSTM, highlighting its effectiveness in minimizing missed failures. Although BiLSTM incurred slightly higher computational overhead, its superior performance underscores the value of integrating multi-sourced data and advanced sequence models for reliable, actionable PdM in heavy-duty fleets.