Quantum mechanics-based deep learning framework considering near-zero variance data

Abstract

With the development of automation technology, big data is collected during operation processes, and among various machine learning analysis techniques using such data, deep neural network (DNN) has high analysis performance. However, most industrial data has low-variance or near-zero variance data from the refined processes in the collected data itself. This reduces deep learning analysis performance, which is affected by data quality. To overcome this, in this study, the weight learning pattern of an applied DNN is modeled as a stochastic differential equation (SDE) based on quantum mechanics. Through the drift and diffuse terms of quantum mechanics, the patterns of the DNN and data are quickly acquired, and the data with near-zero variance is effectively analyzed simultaneously. To demonstrate the superiority of the proposed framework, DNN analysis was performed using data with near-zero variance issues, and it was proved that the proposed framework is effective in processing near-zero variance data compared with other existing algorithms.