College of Engineering Unit:
During the process of remelting a metal with a furnace, different anomalous events can occur, impacting the structure of the resulting metal. These events can cause defects in the metal, and can result in serious problems for customers. For example, if a defective metal was used in an airplane or car, it could put every passenger in danger. Thus, Ampere Scientific wants to make a system that can detect anomalous events inside the crucible in order to prevent defects.
To prevent the production of defective metal, Ampere Scientific has created the VARmetric measurement system. This system is placed around the remelting crucible, and uses around 1,000 sensors. These sensors measure some real-time physical properties inside the crucible, such as current or magnetic flux data. Currently, the VARmetric system can collect many datapoints during the remelting process.
Our team has worked with Ampere to develop a system to analyze the VARmetric data and automatically detect anomalies using different machine learning approaches.
At the beginning of our project we visited Ampere’s facility in Albany. They showed us the workings of a metal remelting crucible and the data we could expect to work with.
Our team then did research and found a variety of anomaly detection algorithms that we could possibly put to use. This includes things like autoregression, the cusum algorithm, scikit-learn’s elliptic envelope, and the rulsif algorithm. We found the best results from the rulsif algorithm.
This project was mostly experimental, and could be improved upon by a team in the future. This could be done through a variety of methods (using new algorithms, tuning hyperparameters for our algorithms, etc). We also dedicated time to visualizing the sensor data and machine learning output using D3 tools to create a web application. This should be useful to Ampere and their customers in using the machine learning algorithm.