Equipment failures in industrial settings can have significant consequences for safety, production availability and overall operational efficiency. Prognostics and Health Management (PHM) systems assist maintenance operators by detecting abnormal conditions, diagnosing their causes and predicting future failure occurrences. In PHM, Deep Learning (DL)-based methods have emerged for their capability of learning complex relationships in large datasets, while also reducing the need for extensive feature engineering. The deployment of PHM methods based on DL for predictive maintenance in industry is hindered by their unsatisfactory performance in applications characterized by large noise, complex dynamics, variability of the operating conditions and scarcity of degradation data. In this context, the primary objective of the research presented in this PhD thesis work is to enhance the performance of DL methods for PHM by leveraging domain-specific knowledge through the use of Physics-informed Deep Learning (PIDL). Two forms of domain-specific knowledge are considered: empirical expertise of field experts and physics-based laws governing the operation of industrial equipment. These sources of domain-specific knowledge are systematically incorporated within advanced state-of-the-art deep learning techniques, such as Deep Neural Networks (DNNs), Auto-Encoders (AEs), Long Short-Term Memory (LSTM) networks, Transfer Learning (TL) and Domain Adaptation (DA). The contribution of this PhD work is the development of novel PIDL methods enhancing prediction accuracy and stability, by dealing with the uncertainties affecting the domain-specific knowledge and managing different operating conditions. The proposed methods have been successfully verified on data from Electro-Hydraulic Servo Actuators (EHSAs) of turbofan engines and Electro-Mechanical Actuators (EMAs) of aircrafts, regulating valves of nuclear power plants, Insulated-Gate Bipolar Transistors (IGBTs) used in electric vehicles, and international benchmarks. The proposed methods allow obtaining a significant increase of performance with respect to the state-of-the-art methods: Area Under Curve of Receiver Operating Characteristic (AUC-ROC) increased on average of 33.1% in the fault detection tasks, Root Mean Squared Error (RMSE) decreased 29.2% and R^2 increased 4.1% in the fault diagnostic tasks, RMSE decreased 12.1% and Average Steadiness Index (ASI) improved 24.7% in the prognostic tasks.
I guasti ai componenti in ambito industriale possono avere conseguenze significative sulla sicurezza, sulla disponibilità degli impianti produttivi e sull'efficienza operativa complessiva. I sistemi di Prognostica e Gestione della Salute (Prognostics and Health Management, PHM) supportano gli operatori di manutenzione nel rilevare condizioni anomale, diagnosticare le cause dei guasti e prevedere l'insorgenza futura di guasti. Nell’ambito del PHM, i metodi basati su Deep Learning (DL) hanno dimostrato notevoli potenzialità grazie alla loro capacità di apprendere relazioni complesse da grandi quantità di dati, riducendo al contempo la necessità di un’ingegnerizzazione manuale delle caratteristiche. Tuttavia, l’applicazione industriale dei metodi PHM basati su DL per la manutenzione predittiva è spesso limitata da prestazioni insoddisfacenti in contesti caratterizzati da elevato rumore nei dati, dinamiche complesse, variabilità delle condizioni operative e scarsità di dati relativi al degrado. In questo contesto, l’obiettivo principale della ricerca condotta in questa tesi di dottorato è il miglioramento delle prestazioni dei metodi DL per il PHM, attraverso l’integrazione di conoscenze dominio-specifiche mediante l’uso del Deep Learning informato dalla fisica (Physics-Informed Deep Learning, PIDL). Vengono considerate due tipologie di conoscenza dominio-specifica: l’esperienza empirica degli esperti di settore e le leggi fisiche che governano il funzionamento dei componenti industriali. Tali conoscenze sono sistematicamente incorporate all’interno di tecniche avanzate di deep learning allo stato dell’arte, come le Deep Neural Networks (DNN), gli Auto-Encoder (AE), le reti Long Short-Term Memory (LSTM), il Transfer Learning (TL) e l’Adaptation Dominale (DA). Il contributo principale di questa tesi di dottorato è lo sviluppo di nuovi metodi PIDL in grado di migliorare l’accuratezza e la stabilità delle previsioni, affrontando le incertezze associate alla conoscenza dominio-specifica e gestendo efficacemente la variabilità delle condizioni operative. I metodi proposti sono stati validati con successo su dati provenienti da Attuatori Elettro-Idraulici (Electro-Hydraulic Servo Actuators, EHSAs) di motori turbofan, Attuatori Elettro-Meccanici (Electro-Mechanical Actuators, EMAs) di velivoli, valvole di regolazione utilizzate in centrali nucleari, Transistor Bipolari a Gate Isolato (Insulated-Gate Bipolar Transistors, IGBT) impiegati nei veicoli elettrici, nonché su benchmark internazionali. I risultati mostrano un incremento significativo delle prestazioni rispetto ai metodi allo stato dell’arte: nelle attività di rilevamento guasti, l’Area Under Curve della Receiver Operating Characteristic (AUC-ROC) è aumentata in media del 33,1%; nelle attività diagnostiche, l’errore quadratico medio (RMSE) è diminuito del 29,2% e il coefficiente di determinazione (R²) è aumentato del 4,1%; nelle attività prognostiche, il RMSE è diminuito del 12,1% e l’Average Steadiness Index (ASI) è migliorato del 24,7%.
Physics-informed deep learning methods of prognostics and health management for predictive maintenance of industrial equipment
Chenyang, Lai
2025
Abstract
Equipment failures in industrial settings can have significant consequences for safety, production availability and overall operational efficiency. Prognostics and Health Management (PHM) systems assist maintenance operators by detecting abnormal conditions, diagnosing their causes and predicting future failure occurrences. In PHM, Deep Learning (DL)-based methods have emerged for their capability of learning complex relationships in large datasets, while also reducing the need for extensive feature engineering. The deployment of PHM methods based on DL for predictive maintenance in industry is hindered by their unsatisfactory performance in applications characterized by large noise, complex dynamics, variability of the operating conditions and scarcity of degradation data. In this context, the primary objective of the research presented in this PhD thesis work is to enhance the performance of DL methods for PHM by leveraging domain-specific knowledge through the use of Physics-informed Deep Learning (PIDL). Two forms of domain-specific knowledge are considered: empirical expertise of field experts and physics-based laws governing the operation of industrial equipment. These sources of domain-specific knowledge are systematically incorporated within advanced state-of-the-art deep learning techniques, such as Deep Neural Networks (DNNs), Auto-Encoders (AEs), Long Short-Term Memory (LSTM) networks, Transfer Learning (TL) and Domain Adaptation (DA). The contribution of this PhD work is the development of novel PIDL methods enhancing prediction accuracy and stability, by dealing with the uncertainties affecting the domain-specific knowledge and managing different operating conditions. The proposed methods have been successfully verified on data from Electro-Hydraulic Servo Actuators (EHSAs) of turbofan engines and Electro-Mechanical Actuators (EMAs) of aircrafts, regulating valves of nuclear power plants, Insulated-Gate Bipolar Transistors (IGBTs) used in electric vehicles, and international benchmarks. The proposed methods allow obtaining a significant increase of performance with respect to the state-of-the-art methods: Area Under Curve of Receiver Operating Characteristic (AUC-ROC) increased on average of 33.1% in the fault detection tasks, Root Mean Squared Error (RMSE) decreased 29.2% and R^2 increased 4.1% in the fault diagnostic tasks, RMSE decreased 12.1% and Average Steadiness Index (ASI) improved 24.7% in the prognostic tasks.File | Dimensione | Formato | |
---|---|---|---|
PhD Thesis Chenyang Lai_28.6.pdf
embargo fino al 29/06/2028
Dimensione
14.06 MB
Formato
Adobe PDF
|
14.06 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/298389
URN:NBN:IT:POLIMI-298389