Semi-supervised process fault classification based on convolutional ladder network with local and global feature fusion.

• The dynamic data representation method is applied to semi-supervised fault classification and the differences between it with supervised fault detection are emphasized. • A semi-supervised form of CNN, i.e., SS-LGF-CLN, is proposed for fault classification, which can jointly capture latent manifold features of unlabeled samples and extract discriminative features of labeled samples. • The usefulness of global inter-variable correction features in representing incipient faults is discussed for the first time, while other CNN based fault classification methods focused on local inter-variable correction features. • An overall fault classification accuracy of 91.29% including incipient faults is achieved on the benchmark process. Effective fault classification is of great significance to isolate and eliminate faults in the smart manufacturing process, especially for chemical industry. However, incipient faults and limited labeled samples make it hard to classify faults accurately. This motivates the formulation of a semi-supervised convolutional ladder network with local and global feature fusion. In the algorithm, a convolutional ladder network is developed to capture higher-order correlations from both labeled and unlabeled samples simultaneously to overcome the problem caused by limited labeled samples, skipped connections are embedded within which to make a balance between supervised and unsupervised feature learning for further identifying faults. To improve the classification performance on incipient faults, a local and global feature fusion strategy is proposed to enhance the representation of incipient faults. Furthermore, a semi-supervised dynamic data representation strategy is introduced to jointly deal with labeled and unlabeled process samples, which enables the proposed method to handle process dynamics by characterizing temporal information of process variables. Experiments on the Tennessee Eastman process show that the proposed method is effective for process fault classification when labeled samples are limited compared to the state-of-the-art algorithms. [ABSTRACT FROM AUTHOR]