Improved Fault Detection Using Kantorovich Distance and Neighborhood Preserving Embedding method

Traditional data-driven fault detection methods based on Euclidean distance usually lack the measurements of data distribution, thus they are not sensitive to small faults in the processes. Kantorovich distance measures the distance between two distributions, and can directly reflect the slight differences between two distributions. In this paper, Kantorovich distance is combined with neighborhood preserving embedding (NPE) algorithm to achieve the fault detection. Firstly, the training dataset is divided into two subsets. Two NPE models are established separately for the two subsets, and the residuals are obtained. Secondly, the threshold is calculated based on the Kantorovich distance matrix by employing a moving window on the two subsets. Thirdly, the monitoring index is determined for test data using Kantorovich distance. The effectiveness of the proposed fault detection method is verified by a numerical example.