Diagnostics of Road Conditions Using Acceleration Sensor: Machine Learning - LSTM autoencoder and Gaussian Mixture Model

Assessing road conditions is critical in the transportation system to ensure travel safety, mobility and efficiency. In many studies, poor road conditions are detected and analyzed using anomaly detection algorithms, employing supervised and unsupervised learning approaches. The paper presents a novel and low-cost framework using acceleration sensors attached on a vehicle to assess road conditions through machine learning techniques. Two unsupervised machine learning - LSTM autoencoder and GMM-EM models are applied to detect distressed road conditions (i.e., anomaly detection in acceleration data). The sliding window algorithm is utilized along with LSTM autoencoder to filter accelerations and locate patterns of candidate distressed conditions. The magnitude of all accelerations in the vertical, longitudinal, and lateral directions are normalized and transformed into overlapping windows. These windows then serve as input for an LSTM-autoencoder to detect road anomalies. In GMM-EM, all raw acceleration data are used where anomalies are identified when the log probabilities computed from the GMM component are lower than one percentile. Subsequently, all identified anomalies are georeferenced in the ArcGIS software, and an R² value is computed using the identified anomalies and International Roughness Index (IRI). As a result, 77.91% and 82.63 % R² values were computed, indicating a strong relationship between the identified poor road conditions using LSTM autoencoder and IRI values. The R² values are 58.89% and 74.35% while using GMM-EM model.