A Low-Latency Machine Learning Architecture for Scalable Vehicle Recognition in Networked Mobility Ecosystems

Abstract

The rapid advancement of intelligent transportation systems and automotive monitoring has increased the demand for automated and reliable detection of vehicle conditions using acoustic signals. In-vehicle audio data contains rich information about mechanical states and faults such as braking anomalies, engine idle variations, startup failures, and combined fault conditions. However, the primary challenge lies in the unstructured and high-dimensional nature of audio signals, where subtle variations correspond to different operational behaviors. Earlier approaches rely on manual feature extraction techniques and basic Machine Learning (ML) models, which fail to capture complex temporal dependencies and hierarchical relationships between primary vehicle states and fine-grained fault categories, resulting in reduced accuracy and poor generalization. To address these limitations, this work proposes a transformer-driven hierarchical audio classification framework. The system utilizes Waveform Language Model (WavLM) to extract deep and contextual acoustic representations from raw audio signals. These features are used for dual-level classification, where the first level predicts the main class (Y1) and the second level identifies the sub-class (Y2). Multiple ensemble models, including Categorical Boosting Classifier (CBC), Histogram-Based Gradient Boosting (HGB), Extra Trees Classifier (ETC), and a proposed Tree-Based Generalized Additive Model (TGAM), are employed to enhance prediction performance. The TGAM model further improves interpretability by generating simplified rule-based insights from tree structures. The proposed system achieves improved accuracy, robustness, and interpretability, making it suitable for real-time vehicle diagnostics and intelligent monitoring applications, thereby contributing to safer and more efficient transportation systems.