Prognostics and Health Management (PHM) is a field of research and application aiming at detecting abnormal conditions in industrial systems, diagnosing their causes and the degradation level of the components and predicting their remaining useful life, with the objective of increasing the system safety, reliability and availability and reducing the cost of maintenance. In the era of Internet of Things, the rapid growth of the amount of data collected during the operation of industrial systems makes possible the development of more accurate and efficient PHM methods based on the advantages of artificial intelligence. However, the data typically available for PHM are characterized by the following missing data: 1) the scarcity of signal measurements collected from industrial systems in abnormal conditions; This is due to the fact that faults are rare, especially for safety related systems. 2) the lack of information about the true system state in correspondence of the collected signal measurements; this depends on the large cost and intensive labor of domain experts needed for retrieving this information. 3) the lack of measurements continuously collected during the operation of the industrial systems. This is due to the fact that data are often collected at irregular time steps, e.g. in correspondence of the occurrence of triggering events at the system level, and to the difficulties of managing and storing the large amount of monitoring data. In this context, this PhD work develops a novel PHM framework for PHM in case of missing information, which is one of the most limiting issues toward the real applications of prognostics and health management to industrial systems. The proposed framework is based on the use of Deep Learning (DL) and Reservoir Computing (RC). Deep Learning methods have been considered given their ability of learning the complex distributions of the data collected from industrial equipment, whereas RC methods allow capturing the long-term dynamics of the degradation and failure processes. Specifically, with respect to 1) a method for fault detection based on Generative Adversarial Networks (GANs), Auto Encoder (AE) and adaboost ensemble learning has been proposed. The main novelty is the development of an ensemble of Encoder-aided GANs, which is shown able to identify the boundary of the complex, high dimensional, non-smooth, manifold distributions of the healthy data. The proposed method has allowed improving the detection performance with respect to state of the art methods. With respect to 2), a method based on Reservoir Computing (RC), Conceptor and spectral clustering has been developed for degradation level classification. The novelty is the use of Conceptors for the extraction of degradation indicators. The obtained representation of the long-term degradation dynamics has allowed improving the diagnostic performance with respect to state of the art methods, which are typically limited by the use of sliding time windows of fixed lengths. With respect to 3), a Remaining Useful Life (RUL) prediction model based on Echo State Networks (ESNs) and bagging ensemble has been developed. The novelties are the reconstruction of missing degradation information by updating the reservoir state without providing any input and the estimation of the uncertainty affecting the RUL predictions, which is obtained by using an ensemble of ESNs. The reconstruction of missing information in long-term degradation dynamics has allowed obtaining more precise RUL predictions than that obtained using state of the art prognostic methods. The proposed framework is validated using data collected from electrolytic capacitors, equipment of high-speed trains and bearings used in different industrial areas.
La prognostica e la gestione della salute (PHM) è un campo di ricerca e applicazione che mira a rilevare condizioni anomale nei sistemi industriali, a diagnosticare le loro cause e il livello di degrado dei componenti e a prevedere la loro vita utile residua, con l'obiettivo di aumentare la sicurezza, l'affidabilità e la disponibilità del sistema e ridurre il costo della manutenzione. Nell'era dell'Internet of Things, la rapida crescita della quantità di dati raccolti durante il funzionamento dei sistemi industriali rende possibile lo sviluppo di metodi PHM più accurati ed efficienti basati sui vantaggi dell'intelligenza artificiale. Tuttavia, i dati tipicamente disponibili per PHM sono caratterizzati da: 1) la scarsità di misure di segnali raccolti da sistemi industriali in condizioni anormali; ciò è dovuto al fatto che i guasti sono rari, specialmente per i sistemi legati alla sicurezza. 2) la mancanza di informazioni sul vero stato del sistema in corrispondenza delle misure di segnale raccolte; questo dipende dal grande costo e dal lavoro intensivo degli esperti di dominio necessari per recuperare queste informazioni. 3) la mancanza di misure raccolte in modo continuo. Ciò è dovuto ai dati raccolti a intervalli di tempo irregolari, ad esempio in corrispondenza del verificarsi di eventi scatenanti nel sistema, per la difficoltà di gestire e memorizzare i dati monitorati prodotti da un gran numero di componenti industriali online. In questo contesto, questo lavoro di dottorato sviluppa un nuovo quadro PHM per PHM in caso di informazioni mancanti, che è uno dei problemi più limitanti verso le applicazioni reali di prognostica e gestione della salute ai sistemi industriali. Il quadro proposto si basa sull'uso di Deep Learning (DL) e Reservoir Computing (RC). I metodi di Deep Learning sono stati considerati data la loro capacità di apprendere le complesse distribuzioni dei dati raccolti da apparecchiature industriali, mentre i metodi RC permettono di catturare le dinamiche a lungo termine dei processi di degrado e guasto. In particolare, per quanto riguarda 1) è stato proposto un metodo per il rilevamento dei guasti basato su Generative Adversarial Networks (GANs), Auto Encoder (AE) e adaboost ensemble learning. La principale novità è lo sviluppo di un ensemble di GANs assistite da Encoder, che si è dimostrato in grado di identificare il confine delle distribuzioni complesse, ad alta dimensione, non lisce, manifold dei dati sani. Il metodo proposto ha permesso di migliorare le prestazioni di rilevamento rispetto ai metodi allo stato dell'arte. Rispetto a 2), è stato sviluppato un metodo basato su Reservoir Computing (RC), Conceptor e clustering spettrale per la classificazione del livello di degradazione. La novità è l'uso di Conceptor per l'estrazione di indicatori di degrado. La rappresentazione ottenuta delle dinamiche di degrado a lungo termine ha permesso di migliorare le prestazioni diagnostiche rispetto ai metodi allo stato dell'arte, che sono tipicamente limitati dall'uso di finestre temporali scorrevoli di lunghezza fissa. Rispetto a 3), è stato sviluppato un modello di previsione della vita utile rimanente (RUL) basato su reti di stati eco (ESN) e ensemble bagging. Le novità sono la ricostruzione delle informazioni mancanti sulla degradazione aggiornando lo stato del serbatoio senza fornire alcun input e la stima dell'incertezza che influenza le previsioni RUL, ottenuta utilizzando un ensemble di ESN. La ricostruzione delle informazioni mancanti nelle dinamiche di degrado a lungo termine ha permesso di ottenere previsioni RUL più precise di quelle ottenute utilizzando i metodi prognostici allo stato dell'arte. Il quadro proposto viene convalidato utilizzando dati raccolti da condensatori elettrolitici, apparecchiature di treni ad alta velocità e cuscinetti utilizzati in diverse aree industriali.
Deep learning and reservoir computing for prognostics and health management with missing data and information
Mingjing, Xu
2021
Abstract
Prognostics and Health Management (PHM) is a field of research and application aiming at detecting abnormal conditions in industrial systems, diagnosing their causes and the degradation level of the components and predicting their remaining useful life, with the objective of increasing the system safety, reliability and availability and reducing the cost of maintenance. In the era of Internet of Things, the rapid growth of the amount of data collected during the operation of industrial systems makes possible the development of more accurate and efficient PHM methods based on the advantages of artificial intelligence. However, the data typically available for PHM are characterized by the following missing data: 1) the scarcity of signal measurements collected from industrial systems in abnormal conditions; This is due to the fact that faults are rare, especially for safety related systems. 2) the lack of information about the true system state in correspondence of the collected signal measurements; this depends on the large cost and intensive labor of domain experts needed for retrieving this information. 3) the lack of measurements continuously collected during the operation of the industrial systems. This is due to the fact that data are often collected at irregular time steps, e.g. in correspondence of the occurrence of triggering events at the system level, and to the difficulties of managing and storing the large amount of monitoring data. In this context, this PhD work develops a novel PHM framework for PHM in case of missing information, which is one of the most limiting issues toward the real applications of prognostics and health management to industrial systems. The proposed framework is based on the use of Deep Learning (DL) and Reservoir Computing (RC). Deep Learning methods have been considered given their ability of learning the complex distributions of the data collected from industrial equipment, whereas RC methods allow capturing the long-term dynamics of the degradation and failure processes. Specifically, with respect to 1) a method for fault detection based on Generative Adversarial Networks (GANs), Auto Encoder (AE) and adaboost ensemble learning has been proposed. The main novelty is the development of an ensemble of Encoder-aided GANs, which is shown able to identify the boundary of the complex, high dimensional, non-smooth, manifold distributions of the healthy data. The proposed method has allowed improving the detection performance with respect to state of the art methods. With respect to 2), a method based on Reservoir Computing (RC), Conceptor and spectral clustering has been developed for degradation level classification. The novelty is the use of Conceptors for the extraction of degradation indicators. The obtained representation of the long-term degradation dynamics has allowed improving the diagnostic performance with respect to state of the art methods, which are typically limited by the use of sliding time windows of fixed lengths. With respect to 3), a Remaining Useful Life (RUL) prediction model based on Echo State Networks (ESNs) and bagging ensemble has been developed. The novelties are the reconstruction of missing degradation information by updating the reservoir state without providing any input and the estimation of the uncertainty affecting the RUL predictions, which is obtained by using an ensemble of ESNs. The reconstruction of missing information in long-term degradation dynamics has allowed obtaining more precise RUL predictions than that obtained using state of the art prognostic methods. The proposed framework is validated using data collected from electrolytic capacitors, equipment of high-speed trains and bearings used in different industrial areas.File | Dimensione | Formato | |
---|---|---|---|
response letter to reviewers.pdf
Open Access dal 18/03/2024
Dimensione
462.97 kB
Formato
Adobe PDF
|
462.97 kB | Adobe PDF | Visualizza/Apri |
thesis_final.pdf
Open Access dal 18/03/2024
Dimensione
48.63 MB
Formato
Adobe PDF
|
48.63 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/205013
URN:NBN:IT:POLIMI-205013