Road Anomaly Classification Using Vibrational Signals With a Transfer Learning-Enhanced SepConv1D-LSTM Architecture

Accurate road condition monitoring is critical for transportation safety and infrastructure management. Recent advances in Deep Learning and mobile sensing have enabled anomaly detection using smartphone-acquired vibrational signals. This paper proposes a hybrid SepConv1D-LSTM architecture that integrates separable convolutional layers and Long Short-Term Memory (LSTM) units, enhanced through a domain-specific transfer learning strategy. In the adopted approach, each backbone is pretrained on its optimal input data representation domain. We evaluated the proposed model on a publicly available smartphone-based dataset using a rigorous event-level stratified k -fold cross-validation protocol. Extensive experiments were conducted to compare the proposed approach with standalone models, such as SepConv1D, LSTM, GRU, and CNN-2D as well as several state-of-the-art Deep Learning methods across multiple signal representations, such as time, Discrete Wavelet Transform, FFT, and Continuous Wavelet Transform. Results demonstrate that the proposed hybrid SepConv1D-LSTM consistently outperforms existing methods in classification accuracy and robustness, highlighting the potential of hybrid architectures and domain-specific transfer learning for scalable road condition monitoring using low-cost mobile sensors.