Optimal Level-Crossing Prediction for Electronic Prognostics
Lithium-ion batteries are widely used in industrial, business, and aerospace fields because they are rechargeable and have other desirable features, such as high specific energy and environmental friendliness. In these fields, it becomes critical to monitor battery health and make predictions of the battery life. In this work, we investigate and evaluate the performance of approaches for monitoring battery health using regression and classification. By modeling the physical behavior of the Lithium-ion battery, we can accurately predict the state of discharge, which can be used to determine the condition of the battery. We also build an optimal alarm system based on a sensitivity analysis using ROC curves and computing false alarm and missed detection rates. In classification, the project uses nearest neighbor classification and logistic regression. We draw ROC curves using different fixed times and choose the time whose curve has the largest area under the curve. We also predicted the time at which the battery becomes discharged, and we use K-nearest neighbors, bagging, random forest, gradient boosting, as well as our novel modified nearest neighbor method and a linear interpolation method. Bootstrap method is used to create 95% confidence intervals.
The project was sponsored by Rodney Martin and Chetan Kulkarni from NASA Ames Research Center, executed by three students Haiwen Chen, Jingyi Liu, and Jiadi Wang, under the advisement of Professor Joe Guinness.