A comparative study of anomaly detection methods for gross error detection problems

Abstract

The chemical industry requires highly accurate and reliable measurements to ensure smooth operation and effective monitoring of processing facilities. However, measured data inevitably contains errors from various sources. Traditionally in flow systems, data reconciliation through mass balancing is applied to reduce error by estimating balanced flows. However, this approach can only handle random errors. For non-random errors (called gross errors, GEs) which are caused by measurement bias, instrument failures, or process leaks, among others, this approach would return incorrect results. In recent years, many gross error detection (GED) methods have been proposed by the research community. It is recognised that the basic principle of GED is a special case of the detection of outliers (or anomalies) in data analytics. With the developments of Machine Learning (ML) research, patterns in the data can be discovered to provide effective detection of anomalous instances. In this paper, we present a comprehensive study of the application of ML-based Anomaly Detection methods (ADMs) in the GED context on a number of synthetic datasets and compare the results with several established GED approaches. We also perform data transformation on the measurement data and compare its associated results to the original results, as well as investigate the effects of training size on the detection performance. One class Support Vector Machine outperformed other ADMs and five selected statistical tests for GED on Accuracy, F1 Score, and Overall Power while Interquartile Range (IQR) method obtained the best selectivity outcome among the top 6 AMDs and the five statistical tests. The results indicate that ADMs can potentially be applied to GED problems.