This thesis presents a comprehensive study focused on developing predictive maintenance strategies for tool-holders using advanced vibration analysis and machine learning techniques. The primary objective of the research is to design a system capable of classifying the health condition of tool-holders and estimating their Remaining Useful Life (RUL), ultimately enhancing maintenance processes, minimizing downtime, and improving operational efficiency in machining environments. The study begins with the design of two comprehensive experimental campaigns. Experimental Campaign 1 aims to classify the health status of tool-holders (Healthy or Damaged) based on vibration data collected under various conditions. Vibration signals are captured and analyzed across different tool-holder models, CNC machines, and rotational speeds. Statistical analysis, including factorial ANOVA, confirms the influence of tool-holder design, health state, and rotational speed on vibration signals. These findings underscore the potential of vibration analysis in distinguishing between Healthy and Damaged tool-holders. Experimental Campaign 2 investigates the estimation of Remaining Useful Life (RUL) under controlled operating conditions using an endurance test bench. Vibration signals are monitored as tool-holders are subjected to both axial and radial loads, revealing wear patterns that support the development of an RUL prediction model. Feature extraction and selection are performed to address the high dimensionality of vibration data. A comprehensive set of time and frequency-domain features is evaluated using quantitative metrics, such as Kullback-Leibler divergence, Jensen-Shannon divergence, and the Gini index. From an initial set of thirteen features, six key features—Root Mean Square, Average Amplitude, Peak-to-Peak, Mean Spectrum Amplitude, Gravity Frequency, and Mean Square Frequency—are identified for their strong discriminative power, forming the basis for health classification. These selected features provide a balanced representation of temporal and spectral characteristics essential for subsequent machine learning applications. To classify the health status of tool-holders, a Stacked Neural Network (StNN) is developed, combining a Denoising Autoencoder (DAE) with a Multi-Layer Perceptron (MLP). The DAE preprocesses vibration signals, improving the model's robustness and accuracy in real-world applications where the data integrity may be compromised by noise. This StNN achieves high classification accuracy under both clean and noisy data conditions, significantly outperforming the MLP alone in noise-affected scenarios. This robustness to noise demonstrates the StNN’s suitability for real-time health monitoring in industrial contexts. Finally, the thesis explores a novel method for predicting the RUL of tool-holder bearings by integrating a Sparse Autoencoder (SAE) with an Exponential Degradation Model. The SAE serves as an anomaly detection tool, identifying early signs of degradation in tool-holder bearings through reconstruction error of vibration features. The Exponential Degradation Model fits the progression of these errors to estimate the rate of degradation and predict the time to failure. This method is validated on two experimental tool-holders, demonstrating its ability to accurately predict RUL and offering significant potential for improving predictive maintenance strategies in industrial settings. This research presents significant advancements in the application of vibration analysis and machine learning for predictive maintenance, providing a practical framework for real-time monitoring and decision-making in machining operations. Through the integration of advanced statistical techniques, feature selection, and neural networks, the study underscores the effectiveness of this approach in optimizing maintenance, reducing unexpected failures, and extending tool-holder lifetimes.
Questa tesi presenta uno studio approfondito volto allo sviluppo di strategie di manutenzione predittiva per i portautensili, utilizzando tecniche avanzate di analisi delle vibrazioni e apprendimento automatico. L'obiettivo principale della ricerca è progettare un sistema capace di classificare lo stato di salute dei portautensili e stimare la loro Vita Utile Residua (RUL), migliorando così i processi di manutenzione, riducendo i tempi di inattività e ottimizzando l'efficienza operativa negli ambienti di lavorazione meccanica. Lo studio inizia con la progettazione di due campagne sperimentali approfondite. La Campagna Sperimentale 1 mira a classificare lo stato di salute dei portautensili (Sano o Danneggiato) basandosi sui dati di vibrazione raccolti in diverse condizioni operative. I segnali di vibrazione vengono acquisiti e analizzati su vari modelli di portautensili, macchine CNC e velocità di rotazione. L'analisi statistica, tra cui l'ANOVA fattoriale, conferma l'influenza del design del portautensile, dello stato di salute e della velocità di rotazione sui segnali di vibrazione. Questi risultati evidenziano il potenziale dell'analisi delle vibrazioni per distinguere tra portautensili sani e danneggiati. La Campagna Sperimentale 2 si concentra sulla stima della Vita Utile Residua (RUL) in condizioni operative controllate mediante un banco di prova per test di durata. I segnali di vibrazione vengono monitorati mentre i portautensili sono sottoposti a carichi assiali e radiali, rivelando schemi di usura utili per lo sviluppo di un modello predittivo della RUL. Per affrontare l’elevata dimensionalità dei dati di vibrazione, vengono applicate tecniche di estrazione e selezione delle caratteristiche. Un set completo di caratteristiche nel dominio del tempo e della frequenza viene valutato utilizzando metriche quantitative, tra cui la divergenza di Kullback-Leibler, la divergenza di Jensen-Shannon e l’indice di Gini. Da un set iniziale di tredici caratteristiche, ne vengono selezionate sei fondamentali—Valore Quadratico Medio (RMS), Ampiezza Media, Picco-Picco, Ampiezza Media dello Spettro, Frequenza di Gravità e Frequenza Media Quadratica—per la loro elevata capacità discriminante. Queste caratteristiche forniscono una rappresentazione bilanciata delle informazioni temporali e spettrali, essenziali per le successive applicazioni di apprendimento automatico. Per classificare lo stato di salute dei portautensili, viene sviluppata una Stacked Neural Network (StNN), che combina un Denoising Autoencoder (DAE) con un Multi-Layer Perceptrono (MLP). Il DAE pre-elabora i segnali di vibrazione, migliorando la robustezza e l’accuratezza del modello in applicazioni reali, dove l'integrità dei dati può essere compromessa dal rumore. Questa StNN raggiunge un'elevata accuratezza nella classificazione sia in condizioni di dati puliti che rumorosi, superando significativamente il solo MLP negli scenari affetti da rumore. Questa robustezza al rumore dimostra l’idoneità della StNN per il monitoraggio in tempo reale della salute dei portautensili in contesti industriali. Infine, la tesi esplora un metodo innovativo per prevedere la Vita Utile Residua (RUL) dei cuscinetti dei portautensili, integrando un Sparse Autoencoder (SAE) con un Modello di Degradazione Esponenziale. Il SAE funge da strumento di rilevamento delle anomalie, individuando i primi segni di degrado nei cuscinetti attraverso l'errore di ricostruzione delle caratteristiche di vibrazione. Il Modello di Degradazione Esponenziale adatta la progressione di questi errori per stimare il tasso di degradazione e prevedere il tempo di guasto. Questo metodo è stato validato su due portautensili sperimentali, dimostrando la sua capacità di prevedere con precisione la RUL e offrendo un grande potenziale per migliorare le strategie di manutenzione predittiva in ambito industriale.
Predictive Maintenance Strategies for Tool-Holders Based on Machine Learning and Vibration Analysis: A Two-Stage Experimental Approach
DIPACE, GIUSEPPE
2025
Abstract
This thesis presents a comprehensive study focused on developing predictive maintenance strategies for tool-holders using advanced vibration analysis and machine learning techniques. The primary objective of the research is to design a system capable of classifying the health condition of tool-holders and estimating their Remaining Useful Life (RUL), ultimately enhancing maintenance processes, minimizing downtime, and improving operational efficiency in machining environments. The study begins with the design of two comprehensive experimental campaigns. Experimental Campaign 1 aims to classify the health status of tool-holders (Healthy or Damaged) based on vibration data collected under various conditions. Vibration signals are captured and analyzed across different tool-holder models, CNC machines, and rotational speeds. Statistical analysis, including factorial ANOVA, confirms the influence of tool-holder design, health state, and rotational speed on vibration signals. These findings underscore the potential of vibration analysis in distinguishing between Healthy and Damaged tool-holders. Experimental Campaign 2 investigates the estimation of Remaining Useful Life (RUL) under controlled operating conditions using an endurance test bench. Vibration signals are monitored as tool-holders are subjected to both axial and radial loads, revealing wear patterns that support the development of an RUL prediction model. Feature extraction and selection are performed to address the high dimensionality of vibration data. A comprehensive set of time and frequency-domain features is evaluated using quantitative metrics, such as Kullback-Leibler divergence, Jensen-Shannon divergence, and the Gini index. From an initial set of thirteen features, six key features—Root Mean Square, Average Amplitude, Peak-to-Peak, Mean Spectrum Amplitude, Gravity Frequency, and Mean Square Frequency—are identified for their strong discriminative power, forming the basis for health classification. These selected features provide a balanced representation of temporal and spectral characteristics essential for subsequent machine learning applications. To classify the health status of tool-holders, a Stacked Neural Network (StNN) is developed, combining a Denoising Autoencoder (DAE) with a Multi-Layer Perceptron (MLP). The DAE preprocesses vibration signals, improving the model's robustness and accuracy in real-world applications where the data integrity may be compromised by noise. This StNN achieves high classification accuracy under both clean and noisy data conditions, significantly outperforming the MLP alone in noise-affected scenarios. This robustness to noise demonstrates the StNN’s suitability for real-time health monitoring in industrial contexts. Finally, the thesis explores a novel method for predicting the RUL of tool-holder bearings by integrating a Sparse Autoencoder (SAE) with an Exponential Degradation Model. The SAE serves as an anomaly detection tool, identifying early signs of degradation in tool-holder bearings through reconstruction error of vibration features. The Exponential Degradation Model fits the progression of these errors to estimate the rate of degradation and predict the time to failure. This method is validated on two experimental tool-holders, demonstrating its ability to accurately predict RUL and offering significant potential for improving predictive maintenance strategies in industrial settings. This research presents significant advancements in the application of vibration analysis and machine learning for predictive maintenance, providing a practical framework for real-time monitoring and decision-making in machining operations. Through the integration of advanced statistical techniques, feature selection, and neural networks, the study underscores the effectiveness of this approach in optimizing maintenance, reducing unexpected failures, and extending tool-holder lifetimes.| File | Dimensione | Formato | |
|---|---|---|---|
|
PhDThesis_Giuseppe_Dipace.pdf
accesso aperto
Descrizione: PhDThesis_Giuseppe_Dipace.pdf
Tipologia:
Tesi di dottorato
Licenza:
Creative commons
Dimensione
95.81 MB
Formato
Adobe PDF
|
95.81 MB | Adobe PDF | Visualizza/Apri |
I documenti in SFERA sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


