Spis treści

Zaproszenie na obronę pracy doktorskiej


PRZEWODNICZĄCY i RADA DYSCYPLINY AUTOMATYKA, ELEKTRONIKA i ELEKTROTECHNIKA AKADEMII GÓRNICZO-HUTNICZEJ im. ST. STASZICA W KRAKOWIE
zapraszają na
publiczną dyskusję nad rozprawą doktorską

mgr inż. Anna Stief
COMBINING DATA FROM DISPARATE SOURCES FOR CONDITION MONITORING PURPOSES
Termin: 7 października 2019 roku o godz. 11:00
Miejsce: pawilon B-1, sala 4
Al. Mickiewicza 30, 30-059 Kraków
PROMOTOR: dr. hab. inż. Jerzy Baranowski, prof. AGH - Akademia Górniczo-Hutnicza
PROMOTOR POMOCNICZY: dr. inż. James R. Ottewill - ABB Corporate Research Center Poland
RECENZENCI: dr. hab. inż. Jacek Piskorowski, prof. ZUT, Zachodniopomorski Uniwersytet Technologiczny w Szczecinie
dr. hab. inż. Mariusz Pelc, prof. PO, Politechnika Opolska
Z rozprawą doktorską i opiniami recenzentów można się zapoznać
w Czytelni Biblioteki Głównej AGH, al. Mickiewicza 30

Streszczenie

Łączenie danych z różnych źródeł w celu monitorowania stanu urządzeń przemysłowych

mgr inż. Anna Stief

Promotor: dr. hab. inż. Jerzy Baranowski

Promotor pomocniczy: dr. inż. James R. Ottewill

Dysciplina: Automatyka i Robotyka

Procesy przemysłowe i maszyny generują ogromną ilość danych z wielu różnych źródeł, które potencjalnie mogą być wartościowe dla celów monitoringu i diagnostyki. Celem tej pracy jest zbadanie, w jaki sposób różne dane dostępne w warunkach przemysłowych mogą umożliwić bardziej wiarygodną i solidną ocenę stanu systemu. Badania obejmują dobór i selekcję cech sygnałów, ponieważ jest to jeden z pierwszych kroków w kierunku dokładnego wykrywania błędów i diagnozy. Metody selekcji cech są badane z perspektywy ich przydatności do monitorowania stanu i problemów z fuzją danych. Metoda ReliefF została przeanalizowana i rozszerzona o mechanizmy kompensacji redundancji cech. W celu wyboru cech utworzono metodę hybrydową wykorzystującą ReliefF. W pracy zbadano również nowe algorytmy łączenia danych z różnych źródeł zarejestrowanych online i offline. Opracowano ogólną dwustopniową strukturę Bayesowską, która składa się z fuzji na poziomie cech i fuzji wyników fuzji na poziomie decyzji. Fuzja na poziomie cech jest implementowana naiwnymi klasyfikatorami Bayesowskimi. W przypadku fuzji bayesowskiej na poziomie cech można użyć funkcji wiarygodności opartych na progach, funkcji wiarygodności Gaussa, estymatorów jądrowych lub nowo opracowanej techniki interpolowanych estymatorów jądrowych w zależności od danych monitorowania stanu i systemu. Fuzja na poziomie decyzyjnym jest przeprowadzana z użyciem formuły naiwnego klasyfikatora Bayesowskiego przy użyciu macierzy konfuzji. Ponadto proponuje się dwie metody, uwzględnienia zależności cech od warunków pracy przy użyciu dwustopniowej struktury Bayesowskiej, która jest ważnym zadaniem monitorowania stanu. Nowe metody zwalidowano różnych zastosowaniach dla dwóch studiów przypadku zawierających heterogeniczne dane na temat silników indukcyjnych i wielofazowych instalacji przepływowej. Wyniki potwierdziły, że metody poprawiły wydajność diagnostyki, tworząc solidne, modułowe i skalowalne struktury monitorowania.

Abstract

Title: Combining data from disparate sources for condition monitoring purposes

Industrial processes and machinery generate a vast amount of data from a variety of disparate sources which may potentially be valuable for monitoring purposes. The goal of this thesis is to investigate how disparate data available in an industrial setting may enable more reliable and robust condition assessment. Feature design and selection is investigated as it is one of the first steps towards accurate fault detection and diagnosis. Feature selection methods are reviewed from the perspective of their applicability for condition monitoring and data fusion problems. The ReliefF method, which has been found to be a suitable fit for condition monitoring applications, is further studied and extended to cope with feature redundancy. A ReliefF-based hybrid method is created for feature selection. The thesis also investigates new algorithms to fuse data from disparate sources recorded online, offline, and periodically for equipment condition monitoring. A generic two-stage Bayesian framework is developed, which is composed of a feature-level fusion and a decision-level fusion of the feature-level fusion results. Feature-level fusion is implemented with Naive Bayes classifiers. Thresholds-based likelihood functions, Gaussian likelihood functions, Kernel Density Estimation or a newly developed Interpolated Kernel Density Estimation technique may be used for the feature-level Bayesian fusion depending on the condition monitoring data and system. Decision-level fusion is conducted with a Naive Bayes formulation using confusion matrices. Furthermore, two methods are proposed to account for the operating condition dependency of features when using the two-stage Bayesian framework, which is a typical condition monitoring challenge. The new methods are validated through multiple applications on two case studies containing heterogeneous data obtained from induction motors and a multiphase flow facility. The results confirm that the methods improve the diagnostics performance, while creating a robust, modular and scalable monitoring framework.

Autoreferat

mgr inż. Anna Stief

Promotor: dr hab. inż. Jerzy Baranowski, prof. (AGH)
Dyscyplina: Automatyka i Robotyka

The drive for increased performance stretches operation boundaries. This leads to greater risk of component failures, therefore Condition-based maintenance (CBM) is becoming ever more important. CBM can potentially improve operational safety and reliability, however, monitoring systems have to be developed carefully in a way that false and missed alarm rates remain low.

Different sensors may be used to monitor the health state of a system with different sensor types, as one sensor might be more adept at detecting one fault or operation mode than another sensor. One feature derived from a signal recorded from a particular sensor might be capable of detecting one type of fault, while a different feature calculated from the same source might be more successful at detecting a different type of fault. Hence, feature design and selection is one of the first steps towards successful fault detection and diagnosis.

Industrial processes and machinery can now generate a vast amount of data from a variety of disparate sources, each of which may potentially be valuable for CBM. Condition monitoring approaches which fuse data from multiple sensors have the potential to diagnose faults with reduced false and missed alarm rates. Data may not only take various formats. Alarm and event logs, maintenance logs, design data, connectivity, and topology information may also be used, besides sensor data, as the input of the monitoring system. New process and condition monitoring techniques need to be developed to tackle the challenges of heterogeneous data and combine them in a way which leverages their strengths and suppresses their limitations. Hence, the goal of the thesis is to develop novel methods to combine data from disparate sources recorded online, offline, and periodically in an automated way for equipment condition monitoring.

Work thesis:

Incorporating data from a greater number of diverse sources can enable a more reliable and robust condition assessment. Condition monitoring approaches which fuse data from multiple sensors and sources have the potential to diagnose faults more accurately than using only a single source of information. Disparate data can contain complementary information regarding the health state of the monitored system, therefore their fusion can improve the results of fault diagnostics.

The work objectives based on the work thesis are as follows:

Based on the work objectives, new methods have been proposed in this work to account for various problems which arise when monitoring assets using disparate data. Feature design and selection is investigated as it is one of the first steps towards accurate fault detection and diagnosis. Feature selection methods are reviewed from the perspective of their applicability for condition monitoring and data fusion problems. The ReliefF method, which has been found to be a suitable fit for condition monitoring applications, is further studied and extended to cope with feature redundancy. A ReliefF-based hybrid method is created for feature selection.

The thesis also investigates new algorithms to fuse data from disparate sources recorded online, offline, and periodically for equipment condition monitoring. A generic two-stage Bayesian framework is developed, which is composed of a feature-level fusion and a decision-level fusion of the feature-level fusion results. Feature-level fusion is implemented with Naive Bayes classifiers. Thresholds-based likelihood functions, Gaussian likelihood functions, Kernel Density Estimation or a newly developed Interpolated Kernel Density Estimation technique may be used for the feature-level Bayesian fusion depending on the condition monitoring data and system. Decision-level fusion is conducted with a Naive Bayes formulation using confusion matrices. Furthermore, two methods are proposed to account for the operating condition dependency of features when using the two-stage Bayesian framework, which is a typical condition monitoring challenge.

The newly proposed methods are validated on two case studies, one for monitoring a commonly used component of industrial processes and one for monitoring an entire process plant. The two case studies were selected to show that the proposed methods are both applicable for component level and plant level monitoring. A component of a system may be monitored by several sensors, while a complex process plant may be monitored using various monitoring systems. Both case studies contain heterogeneous data, which are as follows:

The new methods are validated through multiple applications on this two case studies containing heterogeneous data obtained from induction motors and a multiphase flow facility. The results confirm that the methods improve the diagnostics performance, while creating a robust, modular and scalable monitoring framework.

Rozprawa

Pełny tekst rozprawy

Recenzje

Ważniejsze publikacje autora rozprawy

1. Anna Stief, James R. Ottewill, Michal Orkisz, and Jerzy Baranowski. Two stage data fusion of acoustic, electric and vibration signals for diagnosing faults in induction motors. Elektronika ir Elektrotechnika, 23(6):19–24, 2017. DOI: 10.5755/j01.eie.23.6.19690

2. Anna Stief, James R. Ottewill, and Jerzy Baranowski. ReliefF-based feature ranking and feature selection for monitoring induction motors. In 2018 23rd International Conference on Methods & Models in Automation & Robotics (MMAR). IEEE, 2018. DOI: 10.1109/MMAR.2018.8486097

3. Anna Stief, James R. Ottewill, Ruomu Tan, and Yi Cao. Process and alarm data integration under a two-stage Bayesian framework for fault diagnostics. IFAC-PapersOnLine, 51(24):1220–1226, 2018. DOI: 10.1016/j.ifacol.2018.09.696

4. Anna Stief, Ruomu Tan, Yi Cao, and James R. Ottewill. Analytics of heterogeneous process data: Multiphase flow facility case study. IFAC-PapersOnLine, 51(18):363–368, 2018. DOI: 10.1016/j.ifacol.2018.09.327

5. Anna Stief, James R. Ottewill, Jerzy Baranowski, and Michal Orkisz. A PCA - two stage Bayesian sensor fusion approach for diagnosing electrical and mechanical faults in induction motors. IEEE Transactions on Industrial Electronics, 2019. DOI: 10.1109/TIE.2019.2891453

6. Anna Stief, James R. Ottewill, and Jerzy Baranowski. Investigation of the diagnostic properties of sensors and features in a multiphase flow facility case study. IFAC-PapersOnLine, 52(1): 772-777, 2019. DOI: 10.1016/j.ifacol.2019.06.155

7. Matthieu Lucke, Xueyu Mei, Anna Stief, Moncef Chioua, and Nina F. Thornhill. Variable selection for fault detection and identification based on mutual information of multi-valued alarm series. IFAC-PapersOnLine, 52(1): 673-678, 2019.. DOI: 10.1016/j.ifacol.2019.06.140

8. Anna Stief, Ruomu Tan, Yi Cao, James R. Ottewill, Jerzy Baranowski, and Nina F. Thornhill. A heterogeneous benchmark dataset for data analytics: Multiphase flow facility case study. Journal of Process Control, 79: 41–55, 2019. DOI: 10.1016/j.jprocont.2019.04.009