Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Master Internship in Machine Learning for Intelligent Maintenance: Collective Learning For Anomaly Detection
The Chair of Intelligent Maintenance Systems is looking for a motivated intern to work on a collective deep learning methodology. The application is the industrial case where fleets of systems need to be monitored individually, while data and epexrience can be collected on several systems.
Keywords: deep learning; multi-agent-systems; reinforcement learning, predictive maintenance; condition monitoring
Motivation:
The condition of complex engineered systems is tightly monitored with numerous and diverse condition monitoring devices. For such systems, the space of possible faults is large, yet the fault occurrences are rare. Due to this lack of examples, the traditional fault classification approach cannot be applied, and the problem is rather tackled as an unsupervised anomaly detection in condition monitoring time series. In this case, healthy condition data represents the main class, and faults would be the anomalies to find. This problem has solutions in the literature, but it becomes even more challenging when the available data from the main class is not fully representative of all the conditions the system can experience. The anomaly detection task faces the difficulty of distinguishing between a new healthy operating condition and a fault.
A solution to this problem arises when multiple similar machines are available (denoted as ‘fleet’), by combining the available data. This solution is limited to cases where the systems are sufficiently similar such that the data collected on one system would also be representative of others. When the dissimilarity between the systems increases, the distribution mismatch can confuse the anomaly detector and decrease highly its performance. So far, efficient distribution alignment methods relied on the sample labels to anchor the transformation that would make the two distributions match. In our setting, labels are not available.
Motivation: The condition of complex engineered systems is tightly monitored with numerous and diverse condition monitoring devices. For such systems, the space of possible faults is large, yet the fault occurrences are rare. Due to this lack of examples, the traditional fault classification approach cannot be applied, and the problem is rather tackled as an unsupervised anomaly detection in condition monitoring time series. In this case, healthy condition data represents the main class, and faults would be the anomalies to find. This problem has solutions in the literature, but it becomes even more challenging when the available data from the main class is not fully representative of all the conditions the system can experience. The anomaly detection task faces the difficulty of distinguishing between a new healthy operating condition and a fault. A solution to this problem arises when multiple similar machines are available (denoted as ‘fleet’), by combining the available data. This solution is limited to cases where the systems are sufficiently similar such that the data collected on one system would also be representative of others. When the dissimilarity between the systems increases, the distribution mismatch can confuse the anomaly detector and decrease highly its performance. So far, efficient distribution alignment methods relied on the sample labels to anchor the transformation that would make the two distributions match. In our setting, labels are not available.
Task description:
The goal of the project is to develop a framework to solve the problem of collecting data from a fleet and to train an anomaly detector efficient for each individual machine. We propose to model this setup as a multi-agent system, where each agent, linked to a machine, needs to perform intelligent data selection for their own training, collaborative data evaluation for distinguishing between new operating conditions and anomalies, and data alignment between the most relevant systems. By combining and sharing experience, analysis and data across the fleet, the agents should learn the optimal policies robust to the aforementioned difficulties.
We will test the solution on fleets of complex industrial systems such as power plants. These systems operate in very different environments (eg, different parts of the world) but also have evolving environments (different missions, seasonality, etc...). The systems are well maintained and faulty conditions are extremely rare. Yet their accurate and early detection is crucial.
Requirements:
The candidate for this project should have
• Previous experience with machine learning/quantitative tools
• Knowledge in a vector programing language (Python, Matlab, R)
• Knowledge in Deep Learning libraries is a plus (ideally Python-Tensorflow)
• Willingness to work on real application case studies with data collected from complex industrial assets (the data are non-perfect)
Task description: The goal of the project is to develop a framework to solve the problem of collecting data from a fleet and to train an anomaly detector efficient for each individual machine. We propose to model this setup as a multi-agent system, where each agent, linked to a machine, needs to perform intelligent data selection for their own training, collaborative data evaluation for distinguishing between new operating conditions and anomalies, and data alignment between the most relevant systems. By combining and sharing experience, analysis and data across the fleet, the agents should learn the optimal policies robust to the aforementioned difficulties. We will test the solution on fleets of complex industrial systems such as power plants. These systems operate in very different environments (eg, different parts of the world) but also have evolving environments (different missions, seasonality, etc...). The systems are well maintained and faulty conditions are extremely rare. Yet their accurate and early detection is crucial.
Requirements: The candidate for this project should have • Previous experience with machine learning/quantitative tools • Knowledge in a vector programing language (Python, Matlab, R) • Knowledge in Deep Learning libraries is a plus (ideally Python-Tensorflow) • Willingness to work on real application case studies with data collected from complex industrial assets (the data are non-perfect)
Application:
Please send per email in a single PDF that includes a CV with two referees, a motivation letter (how is your profile relevant to the project and how is the project relevant for your career goals) and your transcripts.
Contact:
Dr. Gabriel Michau
ETH Zürich
Chair of Intelligent Maintenance Systems
Dept. of Civil, Environmental and Geomatic Engineering
gmichau[AT]ethz.ch
Application: Please send per email in a single PDF that includes a CV with two referees, a motivation letter (how is your profile relevant to the project and how is the project relevant for your career goals) and your transcripts. Contact: Dr. Gabriel Michau ETH Zürich Chair of Intelligent Maintenance Systems Dept. of Civil, Environmental and Geomatic Engineering gmichau[AT]ethz.ch