A Novel Sequence Discriminative Feature Extraction Network and Its Application in Offline Industrial Fault Pattern Clustering

A valid fault pattern clustering method for stored fault data can effectively help offline fault diagnosis, and the clustering results can provide solid data support for the training of the fault diagnosis model. To achieve it, a novel sequence discriminative feature extraction network (SDFEN) is developed for extracting the discriminative features underlying the industrial time series. The proposed SDFEN composes a prediction network and an extraction network, which are trained successively. The prediction network is designed for preliminary extraction of the discriminative features, and a set of contribution rate sequences are calculated for the supervised training. Besides, to extract local information and contract the features from the same class, a local information extraction network is further connected, trained by reconstruction and neighbor prediction. Dynamic time warping helps to select neighbors among time series, while a parallel strategy is designed to reduce computing load. The final clustering result is given by the Gaussian mixture model. The feasibility and effectiveness of the proposed method are verified by experiments on the Tennessee Eastman benchmark process and the multiphase flow facility.

Automated summary

A valid fault pattern clustering method is proposed in this paper to assist offline fault diagnosis and provide data support for training the fault diagnosis model. The novel sequence discriminative feature extraction network (SDFEN) is developed to extract discriminative features from industrial time series. The proposed method is verified to be feasible and effective through experiments on benchmark processes and flow facilities.