Program.
| 7-th September |
| 9:00-9:15 | Opening remarks | Prof. Olga Fink | ETH Zürich |
| Session Chair: Yosi Hammer, SKF |
| 9:15-9:35 | Stefan Leistner | DB Netz | ||
| 9:35-9:55 | Prof. Stefan Marschnig | TU Graz | ||
| 9:55-10:15 | Discusion |
| Panel Chair: Prof. Olga Fink |
| 10:15-11:00 | Panel: Unsolved problems / future directions in intelligent operation and maintenance. Prof. Albert Bifet (Telecom ParisTech), Dr. Sandro Saitta (Swiss Data Sci. Center) |
| 11:00-11:30 | Networking Break |
| 11:30-13:00 | Interactive Lunch, Meet our Sponsor |
| Session Chair: Dr. Gaëtan Frusque |
| 13:00-13:20 | Rombach Katharina | ETH Zürich | ||
| 13:20-13:40 | Lena Höpler Julia Hartl |
BMW Group | ||
| 13:40-14:00 | Alexia Marchand François Miralles Oliver Blancke |
Hydro-Québec | ||
| 14:00-14:20 | Discusion |
| 14:20-14:50 | Networking Break |
| Session Chair: Giuliano Bernard, OST |
| 14:50-15:10 | Prof. Luna Lu | Purdue University | ||
| 15:10-15:30 | Sarah Lukens | GE Digital | ||
| 15:30-15:50 | Joseba Echevarria Garcia | Capgemini | ||
| 15:50-16:10 | Discusion |
| 16:10-16:30 | Networking Break |
| 8-th September |
| Session Chair: Dr. Kai Hencken, ABB |
| 9:30-9:50 | Renato Fasciati | Rhätische Bahn | ||
| 9:50-10:10 | Daniel Moraschetti | SBB-CFF-FFS | ||
| 10:10-10:30 | Discusion |
| 10:30-11:00 | Networking Break |
| Session Chair: Prof. Dr. Slawomir Nowaczyk, Halmstad University |
| 11:00-11:20 | Bertrand Decocq | Orange | ||
| 11:20-11:40 | Tian Yuan | ETH Zürich | ||
| 11:40-12:00 | Discusion |
| 12:00-13:30 | Interactive Lunch, Meet our Sponsor |
| Session Chair: Dr. Arrigo Beretta, Sulzer |
| 13:30-13:50 | Dr. Laurent Seyler Jorge Gamarra |
LafargeHolcim | ||
| 13:50-14:10 | Liridon Kastrati | Siemens | ||
| 14:10-14:30 | Neil Eklund | PARC | ||
| 14:30-14:50 | Discusion |
| 14:40-15:20 | Networking Break |
| Panel Chair: Prof. Olga Fink |
| 15:20-16:05 | Panel: Unsolved problems / future directions in intelligent operation and maintenance. Dr. Daniel Nikovski (MERL), Dr. Patrick Bangert (Samsung) |
| 16:05-16:15 | Closing remarks | Prof. Olga Fink | ETH Zürich |
Abstract
Intelligent track maintenance works
Assoc.Prof. Dipl.-Ing. Dr.techn. Stefan Marschnig, TU Graz.
Researcher, infrastructure managers, and maintenance companies invested lots of effort into condition description and monitoring and data-based forecasting of maintenance tasks. Nowadays, this is addressed as predictive maintenance. For many different aspects of railway track maintenance, algorithm-based forecasting is already possible and is applied at different infrastructure managers. In most cases the improvement of the maintenance task itself is assessed on an average basis. Meanwhile, sensor technology improved rapidly and is also widely available due to the reduced costs. Future research should move the maintenance task into focus. For track, there is ongoing work using information gathered directly while tamping the ballast bed for both, optimizing the current tamping process and delivering detailed data of the ballast bed performance. Similar aspects should be investigated for other maintenance tasks.
Abstract
Challenges and Opportunities of Digital Twin at Hydro-Québec for Power Transmission Assets
Alexia Marchand, Francois Miralles, Olivier Blancke, Research Scientists, Hydro-Québec.
Hydro-Québec is a major electrical utility in Québec, Canada. It has the largest power transmission grid in North America and one of the most complex power grids in the world. In order to ensure the resilience of its power grid at all times in the face of assets failure risk, Hydro-Québec has established a strategy for the development of asset digital twins. The goals of digital twins are to detect, diagnose faults and prognose failure in order to prevent unexpected asset failure that could lead to major disruptive loss of load. They will make it possible to anticipate hazards and mitigate them at all times by ensuring good coordination of network control activities and maintenance activities. This presentation will provide an overview of the complex operational context of Hydro-Quebec and will highlight the challenges and opportunities of those assets digital twins. For this, recent developments will then be presented on detection and fault diagnosis algorithms applied to Hydro-Quebec data.
Abstract
Invest in Intelligence – which one and when: A down to “track” topic for SBB Maintenance
Daniel Moraschetti, Leiter Unterhalt Rollmaterial, SBB-CFF.
Rolling Stock maintenance accounts for 1/3 of the costs for the travelling customer and has therefore a significant impact on the customer satisfaction measurement “price / value”.
The challenge of Rolling Stock Maintenance is having to deal with active trains aging from 0-50 years of age. Further customer requirements do not differentiate between old and new trains since the price is the same, but how does that create opportunities for the maintenance is an interesting question.
Based on those parameters we explain where along our value chain and how to invest into intelligence becomes a matter of purpose, financial return and readiness, technological and organizational.
Abstract
Predicting Electric Current of Rail Switches using LSTM Neural Networks
Liridon Kastrati, Data Scientist, Siemens.
Prediction of energy behavior is an important topic, especially in the rail industry. One of more sensitive parts in the rail industry is the rail switch. A downtime of just one switch has already a huge impact on the overall rail network. Several approaches predict the energy consumption of cities and urban regions and focus also on devices. The possibility comes more and more today with high performance sensors which have a high recording rate of 10 ms. However, the electric current and its prediction was not considered in those approaches. Especially for rail switches, to the best of our knowledge, their exists no similar approach. In our approach, we implemented an LSTM neural network to predict electric current and analyzed the results. Our dataset consists of more than 100.000 movements of operational data in a very high frequency of 10 ms per datapoint. Our final LSTM network performs very well on the prediction of the next sequences which was evaluated in collaboration with domain experts.
Abstract
Challenges and Opportunities to establish a pre-warning system world-wide for production
Lena Höpler,
Fachliche SubProduct Ownerin Predictive Maintenance,
Digitalisierung Shopfloor, Maintenance,
BMW Group,
Julia Hartl,
IT SubProduct Ownerin Predictive Maintenance,
BMW Group IT - OtD
The BMW Group as a global automotive OEM is continuously enhancing its internal processes within the production by using innovative industry 4.0 technologies. In order to reduce production downtimes, failures and breakdowns can be forecasted by predictive analytics, so that maintenance activities can be scheduled to non-production time and become even more plannable. Thereby, to achieve the optimization of the technical availability and the minimization of unplanned maintenance activities, a global, standardized predictive maintenance system at the BMW Group has been setup by an interdisciplinary team.
The agile approach includes providing a common, cloud-based and scalable IT architecture, developing big data preprocessing services, training statistical and machine learning models as well as the standardization of processes, developing suitable training methods for employees, and last but not at least rolling out Best Practice solutions to all plants of the BMW Production System. The various machinery used in global automobile manufacturing within the BMW Group as well as the ongoing development in the area of Machine Learning Models and Cloud Platform technologies bring exciting challenges and opportunities to establish beneficial solutions for the BMW Production System.
Abstract
5G Network Resilience : Automation is key
Bertrand Decocq,
Team, Program and Project Manager,
Orange
The fifth generation (5G) of mobile telecommunication network is designed with an ambition to be a network faster and smarter than its predecessor.With the digital transformation, all industry sectors will develop new applications with new requirements regarding telecommunication networks. 5G should be able to answer to these verticals’ requirements. Some of the use cases have very strong requirements in terms of network resilience.
To fulfill its ambitious goal and to maximize the satisfaction of emerging end-users, new technologies are introduced: Network Function Virtualization (NFV), Software Defined Network (SDN), Mobile Edge Computing. 5G network also introduces a concept of network slicing, a dedicated independent virtualized end-to-end network that better fits specific applications requirements.
Never the less, while these new technologies provide convenience, they also bring new challenges as network is becoming more complex than ever before, a real complex system of interdependent systems. To face these challenges, automation is key. Artificial Intelligence will play a crucial role in several steps of the network and services lifecycle management.
This talk addresses the main resilience challenges to anticipate in the 5G network from an end-to-end perspective (device, radio network, core network, service platform) and from a multi-layer perspective (slicing, orchestration, virtualization/containerization and infrastructure). Research areas within Orange Labs, based on Artificial Intelligence dealing with these challenges will also be presented.
Abstract
Technical Language Processing: putting technical data to work in industrial applications
Sarah Lukens,
Data Scientist, Asset Performance Management,
GE Digital
Technical language processing (TLP) are approaches for tailoring tools from mainstream artificial intelligence (AI) to engineering data to satisfy industrial business needs. TLP is use-case driven, meaning that engineering requirements drive every step of the model development process and that performance measures are dependent on the business objective. This talk will provide an overview of different industrial use-cases for TLP in the context of the use-case driving design decisions along the full model pipeline, linking raw data with analytical models, the people involved and the resulting augmented work processes.
Abstract
Valorising big data on the example of predictive maintenance
Dr. Laurent Seyler,
Head of New Technologies,
LafargeHolcim
As the world’s global leader in building solutions, LafargeHolcim is reinventing how the world builds to make it greener and smarter for all. In manufacturing with 270 sites worldwide, the Plants of Tomorrow program aims at making our plants more efficient, circular economy driven and carbon neutral. One of the Plants of Tomorrow pillars is dealing with digitalization and Industry 4.0 topics in cement industries, were predictive maintenance is one of the key focus areas. In a cement plant, more than 10’000 tags are being recorded each minute in a technical information system, some other data are recorded at much higher frequency in the central process control system directly. In the predictive maintenance module, decisions will not be based on tribal know-how and preventive inspections only but will be based on forecasting the assets condition. Data science and big data play a key role and allow for many positive impacts on efficiency and performance in a plant. In the current presentation, the LH approach for predictive maintenance as well as some key learnings will be shared and illustrated on the example of vertical roller mill.
Abstract
Intelligent Maintenance of Bridges and Pavements using Smart Piezoelectric Sensors
Luna Lu,
Professor of the Lyles School of Civil Engineering, School of Materials
Engineering and a faculty at Birck Nanotechnology Center,
Purdue University
Piezoelectric sensors can enable intelligent maintenance of bridges and pavements by reducing their maintenance costs and increasing their lifespan. Lack of direct monitoring of concrete bridges and pavements has led to the prolonged deterioration of the transportation infrastructure. As per the 2021 Infrastructure Report Card by the American Society of Civil Engineers, the US infrastructure has received a grade of C-. Out of ~617,000 bridges in the USA, 42% bridges are already more than 50 years old and 7.5% are structurally deficient. 40% of the roads are reported as poor or mediocre conditions, and the backlog to repair these is $125 billion. To address this critical issue, my group has invented novel piezoelectric sensors can accurately measure the concrete’s strength and Young’s modulus. This enables automated and real-time monitoring of bridges and pavements at regular intervals and at desired times so that their strength and conditions could be monitored and maintained regularly. Wireless capabilities ensure regular data transmission of the infrastructure’s condition to engineers and inspection officials. As such, an intelligent maintenance schedule can be realized based on the data-driven decisions informed by the wireless sensors. The technology has been implemented in several interstate highways and vertical construction projects in the state of Indiana at USA, and it shows great promises to providing cost-effective solutions to achieve intelligent maintenance of civil infrastructure.
Abstract
Contrastive Learning for Fault Detection and Diagnostics
Rombach Katharina,
PhD. Student,
ETH Zürich
Regardless of the specific domain, modern industrial processes are increasingly equipped with Condition Monitoring (CM) Devices, This opens the possibility of implementing data-driven solutions to monitor the health state of the asset. Yet, data recorded from operating industrial assets pose particular challenges to data-driven models and particularly deep leaning approaches: On the one hand, changes in the health state of the asset might cause only minor variations in the CM data. To detect this reliably, the model needs to be very sensitive to any abnormal variations in the data. However, on the other hand, fluctuations in the data can also be caused by changing operating conditions or novel external factors. Ideally, all of the possible future variations of the CM data are represented in the training dataset. Unfortunately, this is not always possible as many of these factors might be unknown at training time or can simply not be controlled. This is a significant issue as a sensitive data-driven model will raise a false alarm if subjected to unknown variations in the data.
To tackle this challenge, we propose to use contrastive learning to learn a low-dimensional representation of the data that corresponds solely to the semantic meaning of the data. By doing that, the model is encouraged to filter variations of the data that do not correspond to the health state. We demonstrate on different case studies, how contrastive learning helps to overcome the models sensitivity to variations in the data that are not related to a change of the asset’s health.
Abstract
Prescriptive Maintenance with Deep Reinforcement Learning
Tian Yuan,
PhD. Student,
ETH Zürich
Recently, there has been an increasing interest in prescriptive analytics for intelligent maintenance. To use the information on the condition of the systems proactively and to achieve optimal and operation and maintenance, prescriptive maintenance is required. Prescriptive maintenance goes beyond just predicting the evolution of the system condition and the remaining useful life, it aims to prescribe dynamically optimal operating parameters based on the current system condition. This enables to control the operation and maintenance proactively. Hence, prescriptive maintenance may reach the highest degree of maturity that involves complex methods to include and reinforce adaptation and optimization capabilities. This line of research has remained relatively unexplored. However, it is urgently needed in industrial applications due to the increasing complexity and increasing requirements of complex industrial assets. Thus, developing prescriptive maintenance further is an essential enabler of intelligent and proactive maintenance
To use the information on the condition of the systems proactively and to achieve optimal and proactive operation, reinforcement learning is a potentially promising methodology. We demonstrate on different case studies how reinforcement learning can help the prescriptive optimal operation maintenance. Also, we discuss the advantages compared to rather conventional approaches, and the present open challenges and potential solutions.
Abstract
Maintenance at the Rhaetian Railway,
Renato Fasciati,
CEO,
Rhätische Bahn
The maintenance of the largest meter-gauge railway in Switzerland is complex. 90% of 384km route network are single tracks, 30% is over 1500 meters and 20% of the network are bridges, tunnels and galleries. In the next few years, a large part of the infrastructure and rolling stock will have to be renewed and repaired. Intelligent and smart maintenance helps to keep costs under control. What are the biggest challenges and which solutions are being implemented or planned? What does the management of the Rhaetian Railway understand by intelligent maintenance and which target variables are decisive for the management of the company?
Abstract
Maintenance strategy at DB Netz,
Stefan Leistner,
Head of Analytics and Assetmanagement,
DB Netz
The German rail network comprises more than 33,000 km tracks, approximately 66,000 turnouts and a large variety of other assets. It is thus the largest network in Europe. While maintaining assets largely based on breakdowns or time today, it is clearly understood that focus will move to more condition-based and predictive maintenance in the next years in order to increase overall asset availability. Condition based maintenance will be of significant importance in the first place in terms of remote control, diagnosing faults before they occur and in a further stage with regard to predictive analyses as well, eventually taking advantage of possibilities analytics and AI can offer. Also a standalone data-based prediction will be a further step onto predictive maintenance.
German infrastructure manager DB Netz has gone first steps towards automation and diagnosis e.g. by equipping nearly half of all switches with diagnostic sensors or by building models and algorithms for disorders of level crossings. It could e.g. be shown that within 16 months, >4,400 occurring disorders of sensor-connected turnouts have been avoided. A first algorithm to predict railroad crossing disorders has been implemented into our prevention programm. Seeing this as a mere start in terms of moving towards condition-based maintenance, DB Netz furthermore assesses the application of a risk-based maintenance strategy over all types of assets. Aim is to derive the appropriate maintenance mix for all assets, balancing quality and cost aspects.
Abstract
Predictive Maintenance of Voltage Grid System,
Johannes Manser,
Head Business Intelligences and Data Analytics,
Axpo Grid AG
Axpo is a larger Swiss energy company and operates a 50/100kV high voltage grid in northern Switzerland. In order to improve the efficiency of the grid business, Axpo started already in 2010 digitalizing the whole process of collecting, evaluating and simulating the condition of the grid system. Since then, many initiatives and projects have been started to improve and further develop the process and steadily move from a periodic towards a more predictive maintenance scheme.
The talk will give an overview of the currently ongoing initiatives and projects, including automated drone image recognition and SCADA and transformer data analysis. The talk will present the challenges and lessons learnt.
Abstract
Combining Physics-based Models & Machine Learning for Improved Life Estimation,
Neil Eklund,
Principal Scientist,
PARC, a Xerox Company
Physics-based models (PBM) and data-driven / machine learning (ML) models have complementary capabilities and limitations, so it is natural to consider hybrid approaches that leverage the best of both worlds. The two types of models can be combined in a variety of ways, e.g., preprocessing via ML to correct and combine multiple heterogeneous, incomplete, noisy, partially inaccurate, and asynchronous data streams from a diesel engine before feeding to a PBM; or, using data from PBM to help guide the architecture, parameter selection, coefficients, and thresholds for ML models that estimate remaining life of bearings. This talk will provide an overview of several different applications of hybrid models for remaining life estimation from commercial aviation, military assets, and industrial manufacturing.
Abstract
VIMS: Virtual IoT Manufacturing System,
Joseba Echevarria Garcia,
AICAM HO Technologies and VIMS Architect
VIMS is an open-ended state of the art platform created by Capgemini Engineering for creating data-driven digital twins for manufacturing for different industries. It has been proven to work on pharmaceutical and aerospace contexts and leverages the full capabilities of the IIoT stack in Azure to deliver key insights into processes and systems. In this talk we will dive into its architecture and design principles and provide information about its key capabilities.
Slides of the Presentations.
Organisation .
Team
Prof. Olga Fink
Head of the Chair of Intelligent Maintenance Systems,
ETH Zürich
Dr. Gaëtan Frusque
Research Associate,
Chair of Intelligent Maintenance Systems,
ETH Zürich
Katharina Rombach
PhD. Student,
Chair of Intelligent Maintenance Systems,
ETH Zürich
Maximilian Benker
Research associate,
Institute for Machine Tools and Industrial Management,
Technical University of Munich
Haiwei Xie
PhD. Student,
Chair of Intelligent Maintenance Systems,
ETH Zürich
Patricia Marty
Dipl. Kulturing. ETH Zürich,
Lehrbeauftragte ZHAW
Sandra Jennifer Schmid
Executive Assistant,
Chair of Intelligent Maintenance Systems,
ETH Zürich
Program Planning
Prof. Olga Fink, ETH Zürich
Patricia Marty, ZHAW
Stakeholders
Organisation
Platinum Sponsoring
Gold Sponsoring
Sponsoring
Please contact us if you would like to become a Platinium, Gold, Silver or Bronze Sponsor of our ConferenceContact.
Conference Venue: Online
Conference Headquarters
Campus Hönggerberg
EPFL
CH-1015 Lausanne
Questions? Please do not hesitate to contact us!
Click on the map for control.