Process fault classification based on label propagation and fault relevant ladder network

Abstract Aiming at the problem of process fault data having few labels and the relevance between extracted features and fault labels, This paper presents a process fault classification approach founded on label propagation and a fault - relevant ladder network (LPA-FR-LAN). The Label Propagation Algorithm (LPA) is used to generate labels for unlabeled data, incorporating Maximum Mean Discrepancy (MMD). This method not only relies on the local similarity of the data for label propagation but also considers the global distribution consistency of the data, thereby improving the accuracy of label propagation. A Fault-relevant Feature Extractor (FRFE) is designed for each layer of the ladder network, where a Stacked Autoencoder (SAE) is employed to predict label information. Triplet loss is introduced to enhance the discriminability between different classes. At the same time, a label smoothing mechanism is incorporated to prevent the model from overfitting caused by an excessive focus on a specific class. It is proposed to introduce domain adversarial training in the last layer of the noisy and noiseless encoding process of the ladder network, so that the noise encoder can gradually learn the fault-relevant feature extraction ability of the noiseless encoder during the adversarial process, thereby avoiding the reduction of classification model accuracy caused by errors in label generation. Simulation experiments were carried out on the Tennessee-Eastman (TE) process and three-phase flow(TFF) process, and the results showed that the proposed method can effectively classify process faults.