Semi-supervised LSTM ladder autoencoder for chemical process fault diagnosis and localization

Physical Graph-Based Spatiotemporal Fusion Approach for Process Fault Diagnosis

The rapid development of big data technology and machine learning has increasingly focused attention on fault diagnosis in complex chemical processes. However, data-driven approaches often overlook the inherent physical correlations within the system and lack a robust mechanism for providing trusted explanations for fault diagnosis. To address this challenge, a graph-based fault diagnosis model framework is proposed along with a dependable fault node diagnosis analysis method. In order to enhance the extraction of chemical process features from a spatial perspective, a graph convolution network (GCN)-based node spatial encoding module is integrated. The construction of the adjacency matrix involves combining a priori knowledge of chemical processes with Pearson correlation, thereby incorporating the physical correlations between nodes. Simultaneously, to capture temporal dependencies in fault data, a spatiotemporal feature fusion module based on the long short-term memory network (LSTM) is employed. In terms of model training, a dual-supervision strategy is adopted to ensure stable convergence of the multiclass fault diagnosis model. For model inference, a multi-model voting strategy is designed to mitigate accuracy degradation resulting from model prediction bias. To tackle the interpretability challenge, a fault diagnosis analysis method based on node masking is designed, effectively identifying critical nodes contributing to system faults. Experimental validation on the Tennessee Eastman process demonstrates the effectiveness of the proposed model, achieving high accuracy in fault diagnosis. The average fault diagnosis rate for all fault types reaches 0.9844, showcasing state-of-the-art performance in fault diagnosis.

Semi-supervised learning for industrial fault detection and diagnosis: A systemic review

The automation of Fault Detection and Diagnosis (FDD) is a central task for many industries today. A myriad of methods are in use, although the most recent leading contenders are data-driven approaches and especially Machine Learning (ML) methods. ML algorithms fall into two main categories: supervised and unsupervised methods, depending on whether or not the instances are labeled with the expected outputs. However, a new approach called Semi-Supervised Learning (SSL) has recently emerged that uses a few labeled instances together with other unlabeled instances for the training process. This new approach can significantly improve the accuracy of conventional ML models for industrial environments where labeled data are scarce. SSL has been tested as a promising solution over the past few years for several FDD problems, although there have been no systemic reviews of this sort of approach up until the present review. In this study, an attempt to organize the existing literature on SSL for FDD using the taxonomy of van Engelen & Hoos is reported. The most and the least frequently used SSL algorithms are identified and considered in terms of different fault detection tasks and their most common dataset structure. Moreover, a set of best practices are proposed in the conclusions of this work for implementation under real industrial conditions, so as to avoid some of the most common faults.

Metric Learning-Guided Semi-Supervised Path-Interaction Fault Diagnosis Method for Extremely Limited Labeled Samples under Variable Working Conditions

The lack of labeled data and variable working conditions brings challenges to the application of intelligent fault diagnosis. Given this, extracting labeled information and learning distribution-invariant representation provides a feasible and promising way. Enlightened by metric learning and semi-supervised architecture, a triplet-guided path-interaction ladder network (Tri-CLAN) is proposed based on the aspects of algorithm structure and feature space. An encoder-decoder structure with path interaction is built to utilize the unlabeled data with fewer parameters, and the network structure is simplified by CNN and an element additive combination activation function. Metric learning is introduced to the feature space of the established algorithm structure, which enables the mining of hard samples from extremely limited labeled data and the learning of working condition-independent representations. The generalization and applicability of Tri-CLAN are proved by experiments, and the contribution of the algorithm structure and the metric learning in the feature space are discussed.