On Data Windows for Fault Detection with Neural Networks

Abstract

Fault-tolerant industrial systems need effective fault detection, which we can treat as a classification problem based on the observable system properties. Artificial neural networks (ANN) are provenly strong contenders in this field, capable of supervised learning from experience. They operate with large amounts of data, which we need to prepare and use carefully. This paper explores how different temporal data quantities provided to a long short-term memory network affect its performance and fault detection delay. The case study for this research was the Tennessee Eastman Process data. Test results indicate that networks detect faults best when given the full-length time series data, regardless of the data shape during training. On the downside, the longer the time series, the more delayed the recognition of a possible fault. The systematic approach in feeding the ANN with specific amounts of data revealed a potential overfitting problem of an otherwise perfect classifier.