Unsupervised Methods Used in Diagnosis for Predictive Maintenance Application to Transportation Vehicles by Road

This paper consolidates and distills a complete unsupervised prognostics framework for transportation-vehicle acoustics. A shared preprocessing pipeline converts raw audio into 20 features and lated the Health Indices (HIs) are created.We have to make an necessary assumptionthat the samples are normal at the start of the recoding and later we have degradation of the model. We comparatively evaluate three branches developed and validated in the thesis: (i) Density-Based Spatial Clustering of Applications with Noise (DBSCAN) on Principal Component Analysis (PCA) features for interpretability; (ii) Autoencoder+Long Short-Term Memory (AE+LSTM) for temporally coherent anomaly scoring; and (iii) score-level multi-view fusion across energy, temporal, spectral, and cepstral groups using mean, max, softmax, and weighted rules. On a healthy baseline and six progressive-use sessions, AE+LSTM produced higher monotonicity (0.619) and trendability (0.654) than DBSCAN (0.108 / 0.277), while weighted fusion yielded the most stable RUL fits across linear, quadratic, and exponential models, corroborating the thesis findings. The paper details system architecture, mathematical formulations, training procedure, and ablations; it concludes with limitations and deployment-oriented future work.