Deep Residual Principal Component Analysis As Feature Engineering for Industrial Data Analytics
Junhua Zheng,Zeyu Yang,Zhiqiang Ge
DOI: https://doi.org/10.1109/tim.2024.3420267
IF: 5.6
2024-01-01
IEEE Transactions on Instrumentation and Measurement
Abstract:As an efficient feature engineering tool, principal component analysis (PCA) has been widely used for feature extraction in the past decades. While a deep form of PCA has been developed for performance enhancement, the numerical computation problem may cause collapsion to this method, restricting it to a limited model complexity. At the same time, the potential of residual subspace in the deep PCA model has not been well explored, which may have significant impacts on the performance of deep model. In this article, the layer-by-layer information transformation is made upon the residual subspace of the deep model, which we term here as deep residual PCA (DRPCA). Based on this model structure, sufficient model complexity can be easily obtained, and the features in both principal component (PC) subspace and residual subspace can be fully explored. Besides, an additional PCA model is further constructed based on the augmentation of extracted features from different layers of the deep model. Several downstream tasks are carried out by using DRPCA as feature engineering, performances of which are evaluated through two industrial examples.