A Lightweight 1D-CNN Architecture for Accurate and Efficient Road Type Classification Using Vibrational Signals

This paper addresses the challenge of road type classification using deep learning techniques applied to vibrational signals collected from inertial sensors. Two novel architectures, SepRNet-1D and SepSERNet-1D, are proposed to achieve high classification accuracy while maintaining computational efficiency. The SepRNet-1D architecture is a lightweight 1D-CNN composed of multiple residual blocks, built around a Separable Convolution-1D block, that decomposes the conventional convolution operation into two distinct stages: a depthwise convolution and a pointwise convolution. SepSERNet-1D extends this design by incorporating Squeeze-and-Excitation (SE) modules to enhance feature recalibration and adaptability. Extensive experiments were conducted on a publicly available benchmark dataset, comparing the proposed architectures with state-of-the-art CNN, LSTM, hybrid CNN-LSTM models, several 2D-CNN frameworks and Transformer-based architectures. The evaluations demonstrate the superior classification performance of SepRNet-1D and SepSERNet-1D in terms of Accuracy, Precision, Recall, and F1-score. Computational experiments further highlight the lightweight design of the proposed models, achieving inference times below 4 ms in TensorFlow Lite format on a 13GB RAM desktop CPU. The results underscore the robustness, versatility, and computational efficiency of SepRNet-1D and SepSERNet-1D, making them highly suitable for real-world road condition monitoring applications.