Storage device failure prediction using deep neural network under severely imbalanced data

Abstract

In this thesis, we train a neural network to predict the failure rate of a storage device. Since there is a severe imbalance in the dataset, we analyze how a neural network can learn from imbalanced dataset. We review weighted log likelihood loss, and compare it with well-known approach, weighted SVM, in classical machine learning. Then, we analyze it mathematically in distributional perspective, ROC graph perspective, and prediction threshold perspective. Especially for ROC graph perspective, it represents that all points obtained by weighted negative log likelihood loss of various weights cover the entire optimal points in ROC graph, and vice versa. We also review that sweeping weight in weighted log likelihood loss effects equivalently as sweeping threshold in prediction threshold which is less computationally expensive so that neural network can adapt to new imbalance in dataset easily. After reviewing well-known methods, such as data weighted loss, data augmentation, and prediction threshold, we finally apply these to train the neural network that can predict failure rate of a storage device under severe data imbalance. We use several architectures, such as logistic regression, 3 layer multilayer perceptron and convolutional neural network to model failure rate, and check performance of these three with area under reciever operating curve. The result shows that with these reviewed methods it is possible to train a neural network under severe imbalance of data, and additional techniques, such as log-scaled feature, complex architecture of neural network, frame concatenation, dropout, and etc. can improve performance of failure prediction using a neural network.