A Smart Predictive Maintenance for Industrial Robots using Unsupervised Machine Learning

Abstract

In Industry 5.0, robotic downtime causes high operational costs and reduced productivity. While supervised machine learning offers advantages for Predictive Maintenance (PdM), its deployment is heavily restricted by the lack of labeled real-world failure data. This study investigates unsupervised learning for clustering unlabeled telemetry patterns and isolating operational states using the multi-axis UR3 CobotOps dataset. The methodology involved preprocessing the sensor data through normalization and feature selection, followed by dimensionality reduction using Principal Component Analysis (PCA) and a Variational Autoencoder (VAE). Clustering was then performed using K-Means, Gaussian Mixture Models (GMM), DBSCAN, and VAE+K-Means, with performance evaluated using the Silhouette Score, Calinski-Harabasz Index, and Davies-Bouldin Index (DBI). Unlike traditional reconstruction-error approaches, this study directly analyzes latent representations to compare how linear and non-linear feature spaces capture multi-axis telemetry distributions and expose operational boundaries. Results show that K-Means generated dense spherical partitions, achieving a Silhouette Score of 0.60 and a DBI of 0.66. Conversely, GMM obtained a Silhouette Score of 0.26, indicating that the sensor streams do not follow a Gaussian distribution, while DBSCAN produced the highest DBI (0.74). The VAE+K-Means pipeline achieved a Silhouette Score of 0.52 and a DBI of 0.72, demonstrating its ability to untangle high-dimensional feature relationships. Using an adaptive statistical threshold applied to cluster displacements, the PCA manifold identified anomalous Clusters 3, 4, and 6, whereas the VAE latent space isolated anomalies into Clusters 4 and 8. These regions aligned with the withheld target variables, Robot_ProtectiveStop and grip_lost, demonstrating that unsupervised latent manifolds can transform raw sensor streams into actionable zero-label PdM insights.