Unsupervised Learning for Robust Vehicle Impact Detection: A Statistical Replacement for Rigid Physical Thresholds

Castañeda H.A.; Mas-Azahuanche G.A.; Vasquez A.H.G.; Arteaga-Cortez H.U.; Carpena-Velasquez E.W. · CEUR Workshop Proceedings (2026)

Resumen

This study presents an interpretable unsupervised anomaly-detection framework for prioritizing candidate vehicle-impact events from 5,447 triaxial acceleration records acquired with a low-cost MPU6050 sensor under urban driving conditions. The binary impact label available in the historical dataset was excluded from model fitting and reserved exclusively for external validation, thereby preserving the unsupervised nature of the analysis. The proposed pipeline combines Z-score standardization, Principal Component Analysis (PCA), exploration K-Means and DBSCAN clustering, Gaussian Mixture Model (GMM) density estimation, and a convergent anomaly criterion based on simultaneously low GMM log-likelihood and high PCA reconstruction error. Two principal components retained 75.51% of the total variance. K-Means with k = 2 achieved a Silhouette score of 0.5843, whereas ARI = 0.2706 and AMI = 0.1161 showed limited correspondence between the discovered clusters and the external impact labels. For DBSCAN, ε = 0.1 and min_samples = 10 produced 8 clusters and 1,127 noise points (20.69%), revealing substantial sensitivity to hyperparameter selection. GMM and PCA each identified 273 candidates through percentile-based criteria; their intersection contained 131 statistically robust anomalies, corresponding to 2.405% of the dataset, with a Jaccard index of 0.3157. Because the row-level reference labels and a reproducible magnitude-threshold run could not be recovered from the archived material, external classification metrics cannot be reported and no claim of superiority over a physical-threshold detector is made. Accordingly, PCA+GMM is interpreted as an auditable unsupervised mechanism for prioritizing statistically robust candidate events rather than as a fully validated impact classifier.

Instituciones: Universidad Nacional de Ingeniería; Universidad Nacional Pedro Ruiz Gallo; Universidad Nacional del Callao