Research Projects
Ongoing
Project Description
The project aims at using the potential of the digitalized 2-stroke engines of maritime vessels and enabling early fault detection, identification of fault mechanisms and prediction of the remaining useful lifetime of components. The project will develop an improved framework for predictive maintenance of heterogeneous fleets based on hybrid adaptive approaches combining physical performance models, probabilistic approaches and deep machine learning algorithms.
The developed set of models and algorithms will proactively support the crew on board for interventions and operation, but also optimize engine operation, reliability and components lifetime and maintenance interventions for a portfolio / fleet of maritime vessels. The results will also provide feedback for improvements in design and development and thereby ensure improved system reliability.
Project Description
Modern tunnel boring machines are capturing a wide range of operating and system condition parameters, including speed, drive torque, cutting wheel contact pressure and the support pressure in the excavation chamber. Typically, this data is used for control purposes.
The goal of this project is to use machine learning to develop new applications from the available data. First, we explore how to use the data to obtain more information on the geology surrounding the machine during its operation. This information should allow the operator to adapt the settings of the machine accordingly. Second, we propose to learn from historic data the sets of operating parameters that lead to efficient tunneling. The aim is to develop a recommendation system for the operators, such that they can always operate in the most efficient regime.
This project particularity is the strong intertwining between first, the geological conditions, second, the consequences on the operation of the machines and on the wear of the tools, and last, the expertise of the operator, who adapt the operation accordingly. In addition, the large uncertainty on the geology profile, only partially inferred from prior explorations, makes the validation of the models extremely challenging.
Project Description
Monitoring of the stator insulation on aged generators is mainly performed by partial discharge measurements. Having an accurate estimation of the degradation of the insulation allows the plant owner to plan predictive maintenance at economically optimum time. Partial discharge (PD) data is however tricky to assess, due to several reasons. The measured data are shown in 3D-matrixes, called phi-q-n diagrams (other versions exist, but all of them are based on accumulated high frequency data during a time frame), and different patterns are sought after in this data. However, the data are overlaid with disturbances mainly from the excitation system as well as the presence of several different PD sources that make the assessment a challenging task. Normally, well-trained persons are used for the assessment which include manual data evaluation, which is both time consuming and expensive. This project aims to first, remove and separate PD sources at the high frequency level by learning the pulse characteristics that are linked to either disturbances or different PD sources, and secondly to automatize the PD assessment by applying pattern recognition coupled with deep learning technologies.
Project Description
Railway wheels are safety critical components, have a significant impact on the performance and are a major cost driver for maintenance. The condition of the wheels has also a significant influence on the infrastructure condition and its maintenance. Furthermore, wheel defects cause noise and vibration emissions.
Due to their criticality, wheels are tightly monitored by different condition monitoring devices: including fixed installations, wayside monitoring devices and in-workshop inspections. While wheel defect detection with wayside monitoring devices belongs to the state of the art, particularly based on strain gauges, prediction of the wheel deterioration and defect evolution in time under varying operating conditions is still an open research question.
Even though wheel-rail interaction has been studied comprehensively for decades, it is still not fully understood under real operating conditions and several influencing parameters are not able to be included in the existing models. Particularly the introduction of high-strength wheel and rail steels has imposed new challenges on the wheel-rail interaction. The lack of understanding and modelling of causal relationships can be observed in the fact that different fleets with similar design and similar operating profiles experience different degradation and defect evolution.
The goal of the proposed research project is to predict the evolution of the wheel condition in time by integrating the information of several heterogeneous data sources including real-time information on wheel condition and influencing parameters of its deterioration. The proposed methodology is based on deep learning algorithms enabling to learn the relevant features and their relationships from the heterogeneous data sources and use the learnt relationships to predict the profile evolution.
Project Description
The amount of data collected per period unit has recently grown by orders of magnitude, potentially resulting in much better information on system condition. System health condition is increasingly captured by different types of monitoring systems, at different time scales, with different types of data, ranging from vibration signals to video streaming. The large volume and variety of condition monitoring data drive traditional model- and knowledge-based approaches too difficult to implement and too demanding in resource usage.
The research project tackles the challenge of effectively using the diverse, high-dimensional, high-velocity condition monitoring data of industrial assets to improve their availability and resilience. Intelligent adaptive hybrid predictive maintenance algorithms will be developed in this research project based on a combination of physical performance models and deep neural networks learning the hidden information, dependencies, and structure in condition monitoring data. The developed framework aims at extracting high-level knowledge representations from high-dimensional condition monitoring signals with a high volume, variety and velocity. The developed algorithms will enable (a) learning fault signatures from heterogeneous condition monitoring data sources, (b) transferring them between different systems of the same fleet and (c) predicting failures based on the learned fault signatures and their predicted evolution in time. The developed approaches will enable a fast and efficient adaptation of predictive maintenance systems to new operating conditions and newly developed systems.
Finalized
Project Description
In this collaboration, the objective is to detect anomalous conditions in a sample of condition monitoring observations that could be the precursor of a future failure for the system under study. In order to do so, the only information available is a database of observations where no failure has occurred. Due to the amount of assets to monitor, this task is unaffordable by humans only and automation by machine learning is required to support the analyst by pointing out the most strange assets and limit its work load. The detection of anomaly in the curve shapes is often an unsupervised problem (since data are not labelled).
This task, referred to as anomaly detection, is particularly challenging in time series. Inherent variable relationships evolve dynamically over time, making it difficult to learn a “reference” behaviour. However, with the recent advances in Neural Networks, in particular with Convolutional Neural Networks, Anomaly Detection in 2D images has progressed dramatically. In this collaboration we explored several ways to convert time series into images, to then perform state-of-the art anomaly detection on images. The study explored the strength and weaknesses of each encoding, depending on the type of anomaly present in the data at testing time. Results demonstrated that the approach is sound and can provide very satisfactory results.
Project Description
Modern trains are equipped with sophisticated condition monitoring devices that are typically generating diagnostic events. Each double-decker train of the S-Bahn Zürich, for example, produces an average of 400-800 diagnostic event messages per day.
The project aims at developing a complex event processing system that enables the extraction of temporal and associative patterns and relationships to support the decision makers.