Empowered by the abundant data in the Internet of Things, deep learning models, especially transformers, have achieved great success in the field of industrial process fault diagnosis. However, most existing transformer-based detection methods mainly rely on the temporal information of data while just implicitly considering the correlation between process variables. Moreover, the characteristic of ‘black box’ of transformer models often results in poor interpretability. To handle these problems, an interpretable dual-attention transformer (IDAformer) based fault detection method is proposed, which can explicitly capture both the temporal and variable dependencies. Specifically, an adaptive activation function and a gating mechanism are designed to allocate different activation strengths and weights for the time attention and variable attention, which can enhance network performance. Meanwhile, an interpretable attention mechanism and an attention visualization method are introduced to improve the interpretability. Moreover, state-space model is also integrated to guide the learning of network such that it can further enhance network performance. Finally, the effectiveness of the proposed method is verified through a tandem cold rolling process.

Controlling the crown of hot-rolled strip steel is crucial for product quality, but the imbalance between normal and abnormal samples in production data limits the performance of diagnostic models. In this study, we propose an SNN-ST-CatBoost model that combines the sample pruning method of Sample Nearest Neighbor (SNN) and the self-training (ST) strategy. It focuses on maximizing the value of unlabeled samples after active learning, by pruning to reduce redundant samples, and by using self-training to extend unlabeled data, we can maximize the class value balance of labeled samples and optimize the classification ability of the CatBoost algorithm under data imbalance. The experiment is based on real data from a hot-rolling production line. The results show that compared to traditional models (such as XGBoost and LightGBM), SNN-ST-CatBoost significantly improves in terms of mean average precision, G-mean, and recall rate for the abnormal class, demonstrating its efficient recognition capability for minority class samples. Although the accuracy of the model decreases slightly, its overall diagnostic performance is significantly better than that of existing models.

Flatness is a crucial indicator of strip quality that presents a challenge in regulation due to the high-speed process and the nonlinear relationship between flatness and process parameters. Conventional methods for controlling flatness are based on the first principles, empirical models, and predesigned rules, which are less adaptable to changing rolling conditions. To address this limitation, this paper proposed an offline reinforcement learning (RL) based data-driven method for flatness control. Based on the data collected from a factory, the offline RL method can learn the process dynamics from data to generate a control policy. Unlike online RL methods, the proposed method does not require a simulator for training, the policy can be potentially safer and more accurate since a simulator involves simplifications that can introduce bias. To obtain a steady performance, the proposed method incorporated ensemble Q-functions into policy evaluation to address uncertainty estimation. To address distributional shifts, based on Q-values from ensemble Q-functions, behavior cloning was added to policy improvement. Simulation and comparison results showed that the proposed method outperformed the state-of-the-art offline RL methods and achieved the best performance in producing strips with lower flatness.

As a well-known type of attack, the concept of false data injection attack (FDIA) was originated from electric power grids, and expanded to other fields such as healthcare, transportation, etc. Since the injections of FDIA are hidden in the real measurements, the attack is difficult to be detected. In this paper, we propose a universal detection method against FDIA, which does not need any auxiliary probability distribution model. In order to expose the change of sparsity arising from FDIA, this study first focuses on the structure of mapping matrix and makes it redundant. Column redundancy provides a necessary spatial condition and row redundancy provides a necessary measurement condition, for exposing the change of sparsity caused by FDIA. To prevent an attacker from inferring redundant information, the redundant matrix and system states are randomly scrambled by chaotic system. By the aid of the proposed redundant sparse reconstruction (RSR) algorithm, two different reconstructions of system states are then generated to detect FDIA. Compared with the existing detection methods, the proposed scheme can not only detect FDIA, but also remove it in some degree. Finally, the feasibility of the proposed scheme is verified in the simulations.

This article is concerned with the fault detection and isolation (FDI) problem for a class of nonlinear systems described by the T–S fuzzy models. Based on the concept of minimum unobservability subspace and geometric property of factor space, a set of FDI filters where each residual is only affected by one fault and completely decoupled from other faults is designed. Furthermore, in the decoupling space, the$H_{\infty }/H_{-}$performance indexes are provided to enhance the sensitivity of residual to faults and robustness to disturbances. In particular, to solve the nonconvex filter design problem caused by introducing the$H_{-}$index, the Gerschgorin theorem is first used to linearize the corresponding filter design conditions in the outer region of a ball. Then, the FDI filter design problem is converted into a convex optimization one, which is solved via the linear matrix inequality (LMI) control Toolbox, and the advantages and effectiveness of the proposed FDI method are verified through two simulation examples.