Case Study: A Semi-Supervised Methodology for Anomaly Detection and Diagnosis

Abstract

In this paper, a semi-supervised methodology for anomaly detection and diagnosis is proposed. The approach combines techniques of non-parametric statistics, quality control, and deep learning to provide a tool that allows an adequate and online detection of faults in a production system and a diagnosis of the factors associated with the failure. We propose a semi-supervised neural network for detection and a particular control chart called Open Up for the diagnosis. This neural network is composed of the adjustment of an autoencoder followed by a Long Short-Term Memory model (LSTM). Open Up is used in the last stage to identify the variables associated with the anomaly. This proposal achieves a high correct classification rate using real data of a monitoring system in paper manufacturing and simulated data from the Tennessee Eastman Process.