PHM in practice
Background and motivation
In the field of Prognostics and Health Management (PHM), a lot of methods, techniques and (AI) algorithms are available, and many more are currently being developed. However, when applying these techniques in (industrial) practice, a lot of hurdles are faced, meaning that in many occasions the original enthusiasm changes into frustration, and implementations are not completed. Typical challenges are the lack of (relevant) data (or missing, inaccurate, non-labeled data), limited amount of real failures due to preventive maintenance (which makes training of algorithms and validation difficult), missing guidance in the selection of proper parameters / suitable components / subsystems for PHM, etc.
In this set of tutorials, these practical issues are addressed from different viewpoints. They will make both researchers and potential users aware of the challenges and limitations, but will also provide potential solutions or ways to get around. This will hopefully inspire people from other fields to tackle their specific issues in a similar way.
Format and planning
The tutorial program during PHME21 contains three sessions. After a short introduction, the speakers will cover their topic in an online presentation. However, ways to interact with the speakers will be provided, e.g. for discussion or asking questions.
To allow attendants to maximize the benefits of the tutorials, some preparatory material is provided (see below tutorial abstracts): links to relevant papers, books or websites will allow participants to assess the topic and already gain some basic understanding on that. Further, suggestions are given for presentations / papers in the regular sessions of PHME21 that align with specific topics in the tutorials, and can thus be followed for further details on those topics.
Content
The content of the three tutorials is explained below.
Tutorial 1
Practical issues and challenges in Predictive Maintenance
Prof. dr. ir. Tiedo Tinga – professor at the University of Twente and Netherlands Defence Academy
In this tutorial the practical issues and challenges of applying Predictive Maintenance (or Prognostics) in industrial practice will be discussed. After a short introduction of Predictive Maintenance and its two underlying approaches, i.e. data-driven and (physical) model-based, the main challenges will be identified. These challenges, e.g. the lack of relevant data, limited number of failures, component vs. system level methods and the selection of suitable methods, will be discussed one-by-one, using examples and case studies from practice to illustrate them. Also directions for solving these challenges will be provided. Examples from our research, with applications in the military field, but also in wind energy, rail and infrastructure will be used to clarify the challenges and solutions.
Preparatory material / further reading:
Some of our research papers addressing the challenges discussed (i.e. limited (failure) data, selection of models / approaches):
- Dynamic Maintenance based on Functional Usage Profiles
Tinga, Wubben, Tiddens, Wortmann & Gaalman| JQME, 2020 - Exploring predictive maintenance applications in industry
Tiddens, Braaksma & Tinga | JQME, 2020 - Physical Model-Based Prognostics and Health Monitoring to Enable Predictive Maintenance
Tinga & Loendersloot | Predictive Maintenance in Dynamic Systems, 2019 - Selecting Suitable Candidates for Predictive Maintenance
Tiddens, Braaksma & Tinga | IJPHM, 2018 - Predictive maintenance of maritime systems: models and challenges
Tinga | ESREL, 2017 - The influence of practical factors on the benefits of condition-based maintenance over time-based maintenance
De Jonge, Tinga & Teunter | RESS, 2017
Relevant papers / sessions in PHME21:
| 2839 | Data Selection Criteria for the Application of Predictive Maintenance to Centrifugal Pumps |
| 2806 | Embedding diagnosability of complex industrial systems into the design process using a model-based methodology |
| 2810 | Qualifying Evaluations from Human Operators: Integrating Sensor Data with Natural Language Logs |
| 2814 | The Impact of Data Quality on Maintenance Work Order Analysis: A Case Study in Historical HVAC Maintenance Work Orders |
| 2840 | Robust Model-Based Fault Detection Using Monte-Carlo Methods and Highest Density Regions |
| 2868 | An adaptive framework for remaining useful life predictions of aircraft systems |
Tutorial 2
Challenges in data science application in healthcare
Huiqi (Yvonne) Lu, Ph.D, MIEEE, FHEA – Research Fellow at the Computational Health Informatics Lab, University of Oxford
Dr. Samaneh Kouchaki – Lecturer in Machine Learning for Healthcare, Centre for Vision, Speech, and Signal Processing, University of Surrey
Ever-increasing qualities and quantities of clinical data are routinely collected concerning all aspects of patient care. Electronic health records offer a rich source of clinical information. Developments in wearable sensors, smart home technologies, and the Internet of Things provide the health industry with opportunities to monitor patients either in hospital or in home settings without significant disruption to their everyday activities. Innovations arising from machine learning approaches can facilitate rapid clinical intervention, transform a hospital-only treatment pathway into a cost-effective, home-based, combined alternative, and thus improve the overall quality of patient healthcare. However, analysing real-time collected data poses several challenges as the data can have substantial artefacts, might be highly imbalanced and incomplete, and might contain high variations. Moreover, data labelling can be expensive and time-consuming if there is an insufficient number of high-quality labels. This tutorial will introduce several machine learning techniques to tackle these issues and to provide robust solutions to address such challenges.
Preparatory material / further reading:
- Rapid Triage for COVID-19 using Routine Clinical Data for Patients Attending Hospital: Development and Prospective Validation of an Artificial Intelligence Screening TestSoltan, Kouchaki, Zhu, Kiyasseh, Taylor, Hussain, Peto, Brent, Eyre and Clifton | Lancet Digital Health, 2021
- Hospital Admission Location Prediction via Deep Interpretable Networks for the Year-Round Improvement of Emergency Patient CareEl-Bouri, Eyre, Watkinson, Zhu, Clifton | IEEE Journal of Biomedical and Health Informatics , 2021
- Machine Learning for Clinical Outcome Prediction
Shamout, Zhu and Clifton | IEEE Reviews in Biomedical Engineering, 2021 - PlethAugment: GAN-Based PPG Augmentation for Medical Diagnosis in Low-Resource Settings
Kiyasseh, Abebe Tadesse, Nhan, Tan, Thwaites, Zhu and Clifton | IEEE Journal of Biomedical and Health Informatics, 2020 - A Hamiltonian Monte Carlo Model for Imputation and Augmentation of Healthcare Data
Pourshahrokhi, Narges, et al. | arXiv:2103.02349, 2021 - Semi-supervised Learning for Identifying the Likelihood of Agitation in People with Dementia
Rezvani, Kouchaki, Nilforooshan, Sharp & Barnaghi | arXiv:2105.10398, 2021
Relevant papers / sessions in PHME21:
| 2858 | Diagnosing the Stage of COVID-19 using Machine Learning on Breath Sounds |
Tutorial 3
Domain Adaptation for Fault Diagnosis with Deep Learning
Mr. Qin Wang – PhD student at the Chair of Intelligent Maintenance Systems, ETH Zürich (ETHZ)
Thanks to the digitization of industrial assets and the large amount data coming with it , deep-learning-based fault diagnosis models nowadays are showing promising results for all kinds of fault diagnosis applications. Despite of the great potential of these methods, their ability to generalize on new machines and new working conditions are currently limited because of their tendency to overfit to the training set. This domain shift between the training and testing data can lead to poor test performance. In this tutorial, we will introduce one promising solution to this problem: domain adaptation (DA). DA methods can alleviate the distribution difference between the source and target domains, thus improve the performance on the target. We will first introduce the basic concepts and traditional methods for DA. We will then carefully justify the applicability of these methods in realistic fault diagnosis settings. In addition, we will show some common challenges we face when adopting DA methods to the fault diagnosis context. Finally, we will give perspectives and outlook of DA for fault diagnosis in the deep learning era.
The final part of this tutorial will be a hands-on session, where the participants can work in a jupyter notebook along with the lecturer, demonstrating some of the techniques discussed.
Preparatory material / further reading:
- Unsupervised domain adaptation by backpropagation
Ganin, Yaroslav, and Victor Lempitsky | ICML, 2015 - Revisiting batch normalization for practical domain adaptation
Li, Yanghao, et al. | ICLRW, 2017 - Unsupervised domain adaptation for semantic segmentation via class-balanced self-training
Zou, Yang, et al. | Proceedings of the European conference on computer vision (ECCV), 2018 - Cross-domain fault diagnosis of rolling element bearings using deep generative neural networks
Li, Xiang, Wei Zhang, and Qian Ding | IEEE Transactions on Industrial Electronics, 2018 - Domain adaptive transfer learning for fault diagnosis
Wang, Qin, Gabriel Michau, and Olga Fink | PHM-Paris IEEE, 2019 - Missing-class-robust domain adaptation by unilateral alignment
Wang, Qin, Gabriel Michau, and Olga Fink | IEEE Transactions on Industrial Electronics, 2020 - Domain Adaptive Semantic Segmentation with Self-Supervised Depth Estimation
Wang, Qin, et al. | arxiv, 2021
Relevant papers / sessions in PHME21:
| 2835 | Transfer Learning Approaches for Wind Turbine Fault Detection using Deep Learning |
| 2852 | Feature based deep unsupervised domain adaptation for improved working condition generalization in bearing fault detection with phase current sensor signals |
| 2893 | Domain Adaptations for Guided Wave SHM of Composites: Towards Fleet Monitoring |
| 2952 | A Probabilistic Similarity Based Modeling Approach for Turbomachine Fault Prediction |