An integrated framework for diagnosing process faults with incomplete features

摘要

Handling missing values and large-dimensional features are crucial requirements for data-driven fault diagnosis systems. However, most intelligent data-driven diagnostic systems are not able to handle missing data. The presence of high-dimensional feature sets can also further complicate the process of fault diagnosis. This paper aims to devise a missing data imputation unit along with a dimensionality reduction unit in the pre-processing module of the diagnostic system. This paper proposes a novel pooling strategy for missing data imputation (PSMI). This strategy can simplify complex patterns of missingness and incrementally update the pool. The pre-processing module receives incomplete observations, PSMI estimates missing values, and, then, the dimensionality reduction unit transforms completed observations onto a lower-dimensional feature space. These transformed observations are then fed as inputs to the fault classification module for decision making and diagnosis. This diagnostic scheme makes use of various state-of-the-art missing data imputation, dimensionality reduction and classification algorithms. This enables a comprehensive comparison and allows to find the best techniques for the sake of diagnosing faults in the Tennessee Eastman process. The obtained results show the effectiveness of the proposed pooling strategy and indicate that principal component analysis imputation and heteroscedastic discriminant analysis approaches outperform other imputation and dimensionality reduction techniques in this diagnostic application.