AN APPLICATION MACHINE LEARNING AUTOENCODER FOR IMBALANCED DATA CLASSIFICATION

Machine learning research has long focused on the challenge of classifying imbalanced input. The present investigation suggested a stacked denoising autoencoder neural network (SDAE) algorithm that is based on cost-sensitive oversampling. This algorithm combines the use of cost-sensitive learning with a denoising autoencoder neural network, and it points to the overfitting and noise problems that are caused by the oversampling algorithm when synthesising new minority class samples. The suggested approach has the ability to denoise and classify the collected dataset in addition to oversampling minority class samples at the expense of misclassification costs. When compared to the traditional stacked auto-encoder neural network (SAE) and the oversampling auto-encoder neural network without denoising procedure (OSSAE), the newly proposed method improves the classification accuracy of the minority class of unbalanced datasets. This is because the proposed method utilises a denoising procedure to remove noise from the data. In order to improve the dependability of production equipment and identify equipment breakdown events beforehand, prognostics and health management (PHM) are crucial. Machine learning techniques have become widely employed in recent years to identify plant failure occurrences. According to their respective process phases, each sensor's data is often divided into a number of indications. In order to discriminate between various data series patterns for failure detection, the auto-encoder (AE) approach of imbalanced data classification is presented in this study. The auto-encoder is a fault detection method for determining if each sensor's data is normal or abnormal. The time sequence properties are extracted using a deep learning model for multivariate time series classification in order to obtain better results. The suggested AE outperforms existing machine learning classification models, according to the experimental data