A Spectrogram-Based Approach for Tire Type Classification Using Pass-By Noise

This study presents a non-intrusive approach for tire type classification using pass-by noise recorded from a single roadside microphone. Unlike camera-based or sensor-based systems, the proposed method enables contactless monitoring under real traffic conditions without requiring any modifications to the vehicle. The approach is based on the assumption that tire–road interaction produces characteristic acoustic patterns that can be used to distinguish between different tire types. The recorded signals are first filtered using a 400–5000 Hz bandpass filter to focus on the most informative frequency range. They are then segmented using fixed window sizes with overlapping intervals and transformed into Log-Mel spectrograms to represent their time–frequency structure. The effects of window size, overlap, and preprocessing parameters are systematically investigated. Data augmentation is applied using a polyphase-based downsampling strategy to increase variability. The framework is evaluated on the TyRoN Tyre Road Noise dataset using a Leave-One-Group-Out validation strategy to evaluate performance across different operational conditions. The best results are obtained with a hybrid approach combining ResNet-based feature extraction and an Extra Trees classifier, achieving up to 96.88% accuracy across six tire classes