Open semester project and master thesis opportunities for students:
This project is part of a collaboration between the IMOS lab and Matterhorn Gotthard Bahn, a railway company operating in the Swiss Alps. The student will work on developing computer vision algorithms for automated visual inspection of retaining walls around railway tracks. Retaining (or supporting) walls are crucial infrastructure elements responsible for maintaining the structural integrity of terrains around railway tracks and ensure safe operation. They are subject to wear and damages including cracks, concrete cancer (i.e., alkali–silica reaction), displacements, erosion and water infiltration. Images of retaining walls have already been collected and labels are available. The goal will be to design algorithms to estimate the condition of a wall, with a focus on robustness, transfer learning, and explainability (XAI).
Keywords: Railway, Structural Health Monitoring, Computer Vision, Machine Learning, Transfer Learning, Domain Adaptation, XAI
Detailed description on SiROP
You can apply via SiROP or directly contact Florent Forest.
The student will investigate the use of physics-enhanced graph neural networks for electrical distribution system state estimation. The focus will be on sim-to-real transfer from conventional state estimation, defining the physics-enhanced model, and on detecting anomalies based on a designed anomaly score.
Keywords: Multi-variate Time Series, Anomaly Detection, Graph Neural Networks, Complex Industrial Systems, Machine Learning, Artificial Intelligence, Sim-to-real Transfer, Neural State Estimation, Neural Simulation
You can apply via SiROP or directly contact Raffael Theiler
Anomaly detection in industrial systems is a crucial task, as it enables the identification of anomalies from normal operating conditions. The increasing availability of system condition monitoring data has recently led to a boost in the use of data-driven approaches for anomaly detection.
The complexity of the systems poses specific challenges for achieving accurate and robust fault detection. First, faulty data is rare in real industrial datasets. On the one hand, failures occurring in critical systems, such as power or railway systems, are unusual. On the other hand, it takes a long time for a system to degrade to a final failure. Consequently, in many cases, no faults are observed during the training phase and models are trained in an unsupervised manner. Despite the numerous previous works in unsupervised domain adaptation for PHM, fault detection remains relatively under-studied. Besides, some methods like TTA for other applications can be directly adapted for the fault diagnostic task but not for unsupervised fault detection. Unsupervised anomaly detection is based on a model that only fits the normal data samples. Thus, anomalies are out-of-distribution and can be identified by large prediction errors. Directly applying current TTA methods during test time would cause the model to fit the faulty data, making it hard to detect anomalies from the model’s prediction.
In this project, the student will explore various test-time domain adaptation methods and develop methods for out-of-distribution anomaly detection under domain shift. The ultimate goal of this project is to achieve robust anomaly detection on novel domains. The potential application cases are shown in the following links:
https://www.mvtec.com/company/research/datasets/mvtec-ad
https://www.epfl.ch/labs/cvlab/data/road-anomaly/
Keywords: anomaly detection, domain adaptation, deep learning, test-time, out-of-distribution generalization
You can apply via SiROP or directly contact Han Sun
In this work the student will attempt to adopt the deep operator learning techniques to prepare a surrogate model for shape optimization process.
Keywords: operator learning, DeepONet, shape optimization, level set function
You can apply via SiROP or directly contact Sergei Garmaev.
In this project the student will attempt to embed relevant physics priors in graph neural networks (GNNs) with the goal of developing an inverse modeling framework for the prediction of the underlying material model of a system, given the labeled input (load/excitation) and the output (mechanical response) data.
Keywords: Graph neural networks, Physics prior, finite element method
You can apply via SiROP or directly contact Vinay Sharma.
The student will investigate the use of Graph Neural Networks for anomaly detection in time series data of industrial systems. The focus will be on developing a novel approach that leverages GNNs to infer graph structures from heterogeneous sensor signals and detect anomalies based on a designed anomaly score. The proposed model will be evaluated on benchmark datasets and compared to state-of-the-art methods.
It is not required that the student has worked with GNNs before, but the student should have basic prior knowledge in deep learning and python programming.
You can apply via SiROP or directly contact Mengjie Zhao.
The primary goal of this project is to develop an RL-based solution based on graph grammar for automated robot design, and if possible, combine it with controller design or optimization. To achieve this, the project will involve the following tasks:
- Conduct a comprehensive literature review on robot design optimization and control optimization.
- Develop an RL-based framework for automated robot design and control optimization
- Implement and evaluate the proposed method on a simulated robot platform.
- Document the findings and contribute to research publications on the topic of RL and robotics
You can apply via SiROP or directly contact Yuan Tian.
The goal of this project is to investigate domain adaptation techniques for different dynamics in reinforcement learning, with a specific focus on industrial systems. The project will involve the following tasks:
- Conduct a comprehensive literature review on domain adaptation in reinforcement learning, identifying the key differences and challenges compared to other areas such as computer vision.
- Develop a theoretical framework for domain adaptation in RL, focusing on the adaptation to different dynamics and compared with other methods.
- Implement and evaluate the proposed method on general RL benchmark, and if possible, simulated environment of industrial systems.
- Document the findings and contribute to research publications on the topic of domain adaptation in reinforcement learning.
You can apply via SiROP or directly contact Yuan Tian.
In the context of this Master’s thesis, conducted in cooperation with Kistler (https://www.kistler.com/) and supervised by the IMOS lab (https://www.epfl.ch/labs/imos/) at EPFL, the student will have the opportunity to work on solving the challenges related to both the heterogeneity of sensor data and the crucial temporal relationships between sensors in bridge systems. The thesis will explore strategies to effectively handle the diverse sensor data generated by different sensor types, such as accelerometers and strain gauges, which possess distinct signal characteristics. Additionally, the thesis will delve into capturing the crucial temporal relationships by investigating the incorporation of temporal encoders, specifically those found in spatio-temporal Graph Neural Networks (GNNs). These encoders will be used to model and account for the dynamic dependencies within and between sensors. By developing and implementing innovative solutions, the student aims to enhance the accuracy and effectiveness of GNN-based analysis in monitoring bridge structural behavior.
An important aspect of the thesis is sensor anomaly detection, as it plays a vital role in ensuring the reliability and safety of bridge structures. The student will explore techniques within the GNN framework to identify and distinguish abnormal sensor behavior, enabling early detection of potential structural issues. By integrating anomaly detection capabilities into the GNN-based analysis, the thesis aims to improve overall monitoring and maintenance strategies for bridge structural health. Throughout the thesis, the student will have access to both benchmark datasets for method development and real industrial data provided by Kistler. This unique opportunity allows for the evaluation of the proposed model in real-world scenarios and provides valuable insights into the practical application of GNN-based Bridge SHM.
You can apply via SiROP or directly contact Mengjie Zhao.
The main tasks of this research project involve exploring graph construction and graph learning methods in the context of district heating networks. The project will start with a comprehensive literature review to gain insights into the existing methods and techniques used in graph signal processing and their application to district heating networks. This will provide a solid foundation for understanding the underlying principles and approaches. Based on the literature review, the project will propose a novel physics-informed graph construction method tailored specifically for district heating networks. This method will incorporate domain-specific knowledge and physics-based constraints to accurately capture the relationships between different nodes in the network. The proposed graph construction method will be evaluated and compared with existing approaches to assess its effectiveness and advantages.
Next, the project will focus on applying graph signal processing techniques to address common inverse problems encountered in district heating networks, such as denoising, imputation, and interpolation. These tasks are essential for improving the quality of the data, filling in missing values, and enhancing the accuracy of predictions or estimations.
It is not required that the student has worked with GSP before, but the student should have prior knowledge in signal processing, python programming, linear algebra, and mathematical optimization.
You can apply via SiROP or directly contact Keivan Faghih.
We open the opportunity for a student to participate in one of the EPFL ENAC projects within the Master’s thesis. The project aims to develop a methodological framework for enabling reuse of steel structures based on physics-based and machine learning models.
During the project, the student will develop a physics-informed deep learning model in order to estimate the remaining load capacity of steel structures. The possible algorithms to apply include graph neural networks [1] and neural operators [2].
The training data for deep learning algorithms is obtained by simulating a steel members numerically with Multiple-Point Constraint models. The data includes point cloud coordinates, the corresponding stress and strain disctributions and macroscopic behavior of steel members experiencing local buckling.
If you are interested and highly motivated to pursue studies in this field, we encourage you to apply via SiROP or email: “[email protected]“. The student is expected to be have a solid background in python and deep learning methods.
Expected duration of the project is 5 to 6 months.
Conducted in close collaboration with Kistler, an industry leader, and the IMOS lab at EPFL, this Master’s thesis provides an opportunity to work on cutting-edge challenges in Bridge SHM. Your work will involve creating strategies to handle diverse sensor data from various types, such as accelerometers and strain gauges, each with distinct signal characteristics. You’ll also delve into capturing temporal relationships by incorporating temporal encoders found in spatio-temporal GNNs. A significant portion of your work will be dedicated to sensor anomaly detection, a critical aspect in ensuring the safety and reliability of bridge structures. By identifying abnormal sensor behavior using GNN techniques, you’ll be contributing to the early detection of potential structural issues and the overall improvement of bridge monitoring and maintenance strategies. You will have access to both benchmark datasets and real-world data from Kistler, allowing for a comprehensive evaluation of the proposed model in realistic scenarios. This unique opportunity offers invaluable insights into the practical application of GNNs in Bridge SHM.
You can apply via SiROP or directly contact Mengjie Zhao.
Time-series data is increasingly prevalent across various domains, including finance, healthcare, and environmental monitoring. The ability to extract meaningful information from time-series data is crucial for prediction, classification, and anomaly detection. This project focuses on exploring different time-series representations and their impact on machine learning tasks.
Objectives
-
Benchmarking Time-Series Representations: Investigate the performance of various time-series representations including Fast Fourier Transform (FFT), Recurrence Plot (RP), Gramian Angular Field (GAF), and Markov Transition Field (MTF) in the context of machine learning tasks.
-
Design of Fusion Method: Develop a novel method to fuse information from these representations, aiming to leverage their combined strengths.
-
Evaluation on Diverse Datasets: Assess the effectiveness of the proposed method on a variety of datasets pertaining to Time-Series Classification, Forecasting, and Anomaly Detection.
Methodology
-
Literature Review: Conduct a comprehensive review of existing literature on time-series representation methods and their applications.
-
Implementation of Representations: Implement and fine-tune FFT, RP, GAF, and MTF, ensuring a fair and consistent basis for comparison.
-
Benchmarking: Evaluate each representation method across multiple datasets, using standard metrics relevant to classification, forecasting, and anomaly detection.
-
Fusion Method Development: Design a fusion algorithm to integrate the different representations, potentially using techniques like feature concatenation, ensemble methods, or deep learning architectures.
-
Experimental Evaluation: Test the fusion method across the same datasets used for benchmarking, comparing its performance against the individual representation methods.
You can apply via SiROP or directly contact Hao Dong.