Si trova su / Altri legami
© 2021 American Chemical Society.This paper proposed an adaptive positive semidefinite matrix–based contribution for nonlinear process fault diagnosis. The proposed method is based on the deep learning model autoencoder (AE) and its variants, aiming at exacting nonlinear features in industrial processes. In order to make full use of the extracted feature, an appropriate location in the normal–state area is estimated for local linearization based on the normal–state–feature sequence set built by the normal–state data set. Moreover, an adaptive weight measuring the correlation between hidden layer feature and original variable is built to reduce the redundancy of the variables. Following the basic concept of reconstruction–based contribution (RBC), an adaptive positive semidefinite matrix is constructed for each fault sample to suppress the smearing effect, combining the features extracted by the AE and the built weight. On this basis, a contribution index for each variable is designed to indicate the correlation between the variable and the fault. The feasibility and effectiveness of the proposed method are verified by experiments on a nonlinear numerical case and the Tennessee Eastman benchmark process.
