The global manufacturing sector is undergoing a profound and pervasive transformation. Among the technologies shaping this transformation, collaborative robots have seen rapid and sustained growth in their adoption. As cobots become increasingly widespread in manufacturing environments, ensuring their operational continuity through effective maintenance frameworks becomes a strategic priority. This thesis presents the design, implementation, and experimental validation of a Condition-Based Maintenance (CBM) framework for collaborative robots, which integrates unsupervised anomaly detection, persistent filtering of false positives in the latent space, and incremental classification of fault modes with operator feedback. The framework was developed and validated on the Kawasaki CL108N cobot at the Industry 4.0 Laboratory of Politecnico di Milano. Six multivariate autoencoders, one for each joint, are trained exclusively on nominal kinematic data and learn to reconstruct healthy behavior. Anomalies are flagged through a comparison between the reconstruction MSE and a dynamic per-joint threshold. A vector database stores the latent embeddings of validated false positives and suppresses their recurrence in subsequent sessions. Confirmed anomalies are classified by a Random Forest trained incrementally on statistical features extracted from reconstruction residuals. Validation on experimental datasets highlights variable performance depending on the nature and localization of the simulated fault, with detection rates of up to 25.0%. The classification module achieves 100% accuracy on three fault classes over 34 test windows. The results demonstrate the technical feasibility of a diagnostic system that learns session after session from interaction with the operator, progressively accumulating structured maintenance knowledge.
Il settore manifatturiero globale sta attraversando una trasformazione profonda e pervasiva. Tra le tecnologie che stanno plasmando questo cambiamento, i robot collaborativi si distinguono per una crescita rapida e sostenuta nella loro adozione. Con la crescente diffusione dei cobot negli ambienti di produzione, garantirne la continuità operativa attraverso framework di manutenzione efficaci diventa una priorità strategica. Questa tesi presenta la progettazione, l'implementazione e la validazione sperimentale di un framework Condition-Based Maintenance (CBM) per robot collaborativi, che integra anomaly detection non supervisionata, filtraggio persistente dei falsi positivi nello spazio latente e classificazione incrementale dei modi di guasto con feedback dell'operatore. Il framework è stato sviluppato e validato sul cobot Kawasaki CL108N presso il Laboratorio Industry 4.0 del Politecnico di Milano. Sei autoencoder multivariati, uno per ogni giunto, vengono addestrati esclusivamente su dati cinematici nominali e apprendono a ricostruire il comportamento sano. Le anomalie vengono segnalate tramite un confronto tra l'MSE di ricostruzione e una soglia dinamica per giunto. Un database vettoriale memorizza gli embedding latenti dei falsi positivi validati e ne sopprime la ricomparsa nelle sessioni successive. Le anomalie confermate vengono classificate da un Random Forest addestrato incrementalmente su feature statistiche estratte dai residui di ricostruzione. La validazione su dataset sperimentali evidenzia performance variabili in funzione della natura e della localizzazione del guasto simulato, con tassi di rilevamento fino al 25,0%. Il modulo di classificazione raggiunge un'accuratezza del 100% su tre classi di guasto su 34 finestre di test. I risultati dimostrano la fattibilità tecnica di un sistema diagnostico che apprende sessione dopo sessione dall'interazione con l'operatore, accumulando progressivamente conoscenza manutentiva strutturata.
Condition-based maintenance framework for collaborative robots
Perrone, Carlo
2025/2026
Abstract
The global manufacturing sector is undergoing a profound and pervasive transformation. Among the technologies shaping this transformation, collaborative robots have seen rapid and sustained growth in their adoption. As cobots become increasingly widespread in manufacturing environments, ensuring their operational continuity through effective maintenance frameworks becomes a strategic priority. This thesis presents the design, implementation, and experimental validation of a Condition-Based Maintenance (CBM) framework for collaborative robots, which integrates unsupervised anomaly detection, persistent filtering of false positives in the latent space, and incremental classification of fault modes with operator feedback. The framework was developed and validated on the Kawasaki CL108N cobot at the Industry 4.0 Laboratory of Politecnico di Milano. Six multivariate autoencoders, one for each joint, are trained exclusively on nominal kinematic data and learn to reconstruct healthy behavior. Anomalies are flagged through a comparison between the reconstruction MSE and a dynamic per-joint threshold. A vector database stores the latent embeddings of validated false positives and suppresses their recurrence in subsequent sessions. Confirmed anomalies are classified by a Random Forest trained incrementally on statistical features extracted from reconstruction residuals. Validation on experimental datasets highlights variable performance depending on the nature and localization of the simulated fault, with detection rates of up to 25.0%. The classification module achieves 100% accuracy on three fault classes over 34 test windows. The results demonstrate the technical feasibility of a diagnostic system that learns session after session from interaction with the operator, progressively accumulating structured maintenance knowledge.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_07_Perrone.pdf
non accessibile
Dimensione
2.68 MB
Formato
Adobe PDF
|
2.68 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/261587