Failure Detection Based on Anomaly Detection and Multiple-Layer Perceptron Facing Unbalanced Sample Set

Abstract

As computational performance continues to improve, machine learning is increasingly being used in a variety of areas. Classification problems are one of the most common problems people encounter in their daily lives. However, many classification tasks are confronted with the problem of sample imbalance, which is considered tricky. Although researchers have developed many algorithms for this, problems, such as overfitting, still result in poor classification results in many cases. This paper tries to solve a binary classification with unbalanced sample set applying an idea of combining ready-made anomaly detection and deep learning methods, where anomaly detection algorithms are taken as filters to exclude the effect of samples that are easy to be recognized as the ones from the major category on the final classification done by neural network. This idea is proved more useful on the machine failure detection than using anomaly detection or MLP classifier alone and is believed to be able to serve as a backup or pretest choice in some classification tasks with sample imbalance.