Recent advancements in sensors and network technologies have led to a significant increase in the availability of data, collected during the entire life cycle of industrial components, from the production phase to field operation. This PhD thesis considers time series of measurements of different signal types, such as vibration, temperature, and pressure and other signals, to enhance the reliability of industrial components. Specifically, the research considers the two most critical phases of the component life-cycle: the early-life phase, during which failures are typically due to manufacturing defects caused by low production quality, and the wear-out phase, during which failures are due to component degradation. In this context, a framework is developed for constructing indicators of abnormality behaviour to estimate the production quality and to predict the Remaining Useful Life (RUL) of industrial components. Specifically, two fundamental tasks are considered: a) the assessment of the production quality by monitoring the production process and b) fault prognostics, which aims to reduce failures, by predicting the future evolution of the degradation of the component and its RUL. The framework is based on the use of Machine Learning (ML) for processing the time series data and Particle Filters (PF) for combining data with physics-based models of the degradation process. With respect to task a), the main challenges are: 1) to process multi-dimensional time-series data of raw signal measurements characterized by highly nonlinear dynamic behaviors; 2) the lack of data labeled with the component state (normal/ abnormal conditions), as the ground truth state of the components is typically unknown; and 3) the scarcity of failure data due to the high-quality standards in several manufacturing industries. To address these challenges, the PhD thesis develops novel unsupervised methodology for the detection of the occurrence of abnormal conditions during component production and the estimation of the component quality. It consists of k-fold cross-validation, Long Short-Term Memory (LSTM) autoencoders, and Mahalanobis distance-based abnormality detection. The proposed methodological framework has been applied to three case studies from the semiconductor industry. The obtained results demonstrate the superior performance of the proposed method compared to the state-of-the-art, enable the identification of low-quality components before they start operation, and allow for optimizing decisions about the Burn-In (BI) policy to be applied to the production lots. With respect to task b), the main challenges to be addressed are: 1) the scarcity of data for training data-driven models and the difficulty in generalizing them for applications in similar systems, as most of the data are collected from run-to-failure laboratory tests; and 2) the complexity involved in developing physics-based models of the degradation mechanism. A new framework defines abnormality indicators of components from information extracted from time-series data using a plethora of methods, including clustering methods, also based on the use of the Manhattan distance, and deep learning methods in conjunction with physics-based models. A PF method combines Monte Carlo simulations with a degradation model to estimate RUL and quantify the associated uncertainties. The proposed methodological framework has been applied to three benchmark case studies based on the IMS, PROGNOSTIA and SUT bearing datasets. The accurate RUL estimations obtained by the developed method can contribute to the deployment of predictive maintenance in industry, which can lead to reduce components failures and maintenance costs and increase production availability.
I recenti progressi nelle tecnologie dei sensori e delle reti hanno determinato un significativo aumento della disponibilità di dati raccolti lungo l’intero ciclo di vita dei componenti industriali, dalla fase di produzione fino all’operatività sul campo. La presente tesi di dottorato analizza serie temporali di misure relative a diversi tipi di segnali, quali vibrazioni, temperatura, pressione e altri parametri, con l’obiettivo di migliorare l’affidabilità dei componenti industriali. In particolare, la ricerca si concentra su due fasi critiche del ciclo di vita del componente: la fase iniziale, in cui i guasti sono prevalentemente dovuti a difetti di produzione legati a una bassa qualità manifatturiera, e la fase di usura, in cui i guasti derivano dalla progressiva degradazione del componente. In questo contesto, viene sviluppato un framework per la costruzione di indicatori di comportamento anomalo finalizzati alla stima della qualità di produzione e alla previsione della vita utile residua (Remaining Useful Life, RUL) dei componenti industriali. Nello specifico, vengono affrontati due compiti fondamentali: a) la valutazione della qualità di produzione attraverso il monitoraggio del processo produttivo; b) la prognostica dei guasti, finalizzata a ridurre il verificarsi di malfunzionamenti mediante la previsione dell’evoluzione futura della degradazione del componente e della sua RUL. Il framework proposto si basa sull’applicazione di tecniche di Machine Learning (ML) per l’analisi delle serie temporali e sull’uso di Filtri Particellari (Particle Filters, PF) per la combinazione dei dati con modelli fisico-matematici del processo di degradazione. Per quanto riguarda il compito a), le principali sfide affrontate includono: 1) l’elaborazione di dati multidimensionali derivanti da misurazioni di segnali grezzi caratterizzati da dinamiche altamente non lineari; 2) la mancanza di dati etichettati rispetto allo stato del componente (condizioni normali/anomale), poiché lo stato reale del componente è spesso sconosciuto; 3) la scarsità di dati relativi ai guasti, dovuta agli elevati standard qualitativi dell’industria manifatturiera. Per affrontare tali sfide, la tesi sviluppa una metodologia innovativa non supervisionata per il rilevamento delle anomalie durante la fase di produzione del componente e per la stima della qualità del prodotto. La metodologia si basa su un approccio che combina k-fold cross-validation, autoencoder basati su Long Short-Term Memory (LSTM) e il rilevamento delle anomalie mediante la distanza di Mahalanobis. Il framework metodologico proposto è stato applicato a tre casi di studio nell’industria dei semiconduttori. I risultati ottenuti dimostrano le superiori capacità della metodologia sviluppata rispetto agli approcci allo stato dell’arte, consentendo l’identificazione precoce di componenti di bassa qualità prima della loro messa in esercizio e ottimizzando le decisioni relative alla politica di Burn-In (BI) da applicare ai lotti di produzione. Per quanto concerne il compito b), le principali criticità affrontate sono: 1) la scarsità di dati per l’addestramento di modelli data-driven e la difficoltà di generalizzazione di tali modelli per applicazioni su sistemi simili, poiché la maggior parte dei dati disponibili proviene da test di laboratorio condotti fino al guasto; 2) la complessità nello sviluppo di modelli fisico-matematici in grado di descrivere accuratamente i meccanismi di degradazione. Per risolvere queste problematiche, viene proposto un nuovo framework per la definizione di indicatori di anomalia nei componenti, attraverso l’estrazione di informazioni dalle serie temporali mediante diverse tecniche, tra cui metodi di clustering basati sulla distanza di Manhattan e metodi di deep learning combinati con modelli fisico-matematici. Il metodo basato sui Filtri Particellari sfrutta simulazioni Monte Carlo abbinate a un modello di degradazione per stimare la RUL e quantificare le incertezze associate. Il framework metodologico sviluppato è stato applicato a tre dataset di riferimento nel settore della diagnostica e prognostica dei guasti: IMS, PROGNOSTIA e SUT per cuscinetti volventi. Le accurate stime della RUL ottenute con il metodo proposto favoriscono l’implementazione della manutenzione predittiva in ambito industriale, contribuendo alla riduzione dei guasti dei componenti, all’ottimizzazione dei costi di manutenzione e all’aumento della disponibilità operativa dei sistemi produttivi.
A methodological framework for the prediction of quality and remaining useful life of industrial components and systems
Fatemeh, Hosseinpour
2025
Abstract
Recent advancements in sensors and network technologies have led to a significant increase in the availability of data, collected during the entire life cycle of industrial components, from the production phase to field operation. This PhD thesis considers time series of measurements of different signal types, such as vibration, temperature, and pressure and other signals, to enhance the reliability of industrial components. Specifically, the research considers the two most critical phases of the component life-cycle: the early-life phase, during which failures are typically due to manufacturing defects caused by low production quality, and the wear-out phase, during which failures are due to component degradation. In this context, a framework is developed for constructing indicators of abnormality behaviour to estimate the production quality and to predict the Remaining Useful Life (RUL) of industrial components. Specifically, two fundamental tasks are considered: a) the assessment of the production quality by monitoring the production process and b) fault prognostics, which aims to reduce failures, by predicting the future evolution of the degradation of the component and its RUL. The framework is based on the use of Machine Learning (ML) for processing the time series data and Particle Filters (PF) for combining data with physics-based models of the degradation process. With respect to task a), the main challenges are: 1) to process multi-dimensional time-series data of raw signal measurements characterized by highly nonlinear dynamic behaviors; 2) the lack of data labeled with the component state (normal/ abnormal conditions), as the ground truth state of the components is typically unknown; and 3) the scarcity of failure data due to the high-quality standards in several manufacturing industries. To address these challenges, the PhD thesis develops novel unsupervised methodology for the detection of the occurrence of abnormal conditions during component production and the estimation of the component quality. It consists of k-fold cross-validation, Long Short-Term Memory (LSTM) autoencoders, and Mahalanobis distance-based abnormality detection. The proposed methodological framework has been applied to three case studies from the semiconductor industry. The obtained results demonstrate the superior performance of the proposed method compared to the state-of-the-art, enable the identification of low-quality components before they start operation, and allow for optimizing decisions about the Burn-In (BI) policy to be applied to the production lots. With respect to task b), the main challenges to be addressed are: 1) the scarcity of data for training data-driven models and the difficulty in generalizing them for applications in similar systems, as most of the data are collected from run-to-failure laboratory tests; and 2) the complexity involved in developing physics-based models of the degradation mechanism. A new framework defines abnormality indicators of components from information extracted from time-series data using a plethora of methods, including clustering methods, also based on the use of the Manhattan distance, and deep learning methods in conjunction with physics-based models. A PF method combines Monte Carlo simulations with a degradation model to estimate RUL and quantify the associated uncertainties. The proposed methodological framework has been applied to three benchmark case studies based on the IMS, PROGNOSTIA and SUT bearing datasets. The accurate RUL estimations obtained by the developed method can contribute to the deployment of predictive maintenance in industry, which can lead to reduce components failures and maintenance costs and increase production availability.File | Dimensione | Formato | |
---|---|---|---|
PHD_THESIS_POLIMI_SHARIF_2024.pdf
accesso solo da BNCF e BNCR
Dimensione
8.33 MB
Formato
Adobe PDF
|
8.33 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/297372
URN:NBN:IT:POLIMI-297372