Signal Processing for Vehicle Vibration and Ride Harshness Evaluation: Classical and Deep Learning Techniques

This thesis explores the analysis and evaluation of vehicle vibrational comfort, combining traditional methodologies with advanced machine learning techniques to address challenges in the field. Beginning with the concept of ride harshness and its distinction from harsh ride, the work examines the effects of continuous vibrations and sharp accelerations on passenger discomfort. International standards, such as ISO 2631, are introduced to provide a structured framework for quantifying whole-body vibration exposure and assessing comfort levels through indices like the overall ride value and seat effective amplitude transmissibility. A comprehensive review of signal processing methods is presented, detailing statistical measures, frequency-domain transformations, and filtering techniques. Time-frequency analyses, including Fourier and Wavelet transformations, are explored to enhance the resolution and representation of nonstationary signals. These techniques are applied to real-world test data from a Toyota collaboration, which includes detailed procedures for signal acquisition, preprocessing, and KPI generation. The thesis transitions to a simple deep learning-based approach for classifying vehicle comfort, to map acceleration signals to binary comfort scores, in order to compare two by two the signals from the same sensors of different cars. Key processes, including data normalization and cross-validation, are detailed. Initial results highlight the potential of deep learning to surpass traditional methods in accuracy and reliability, though further research and more extensive datasets are required for full validation. Overall, this work bridges classical engineering techniques and modern machine learning to offer a robust framework for assessing and improving vehicular comfort, with promising implications for the automotive industry.