The accurate prediction of bearing remaining useful life (RUL) is crucial for industrial safety and cost reduction. However, existing deep learning methods often lack interpretability and struggle to integrate local-global and structural-temporal information, limiting their effectiveness. To address this issue, an observable and controllable RUL prediction method is proposed, which visualizes the bearing degradation process through spatio-temporal dynamic graph (STDG), and fully extracts structural information and temporal features to achieve precise RUL prediction. First, canonical variable fluctuation analysis (CVFA) is employed to maximize past-future data correlation, enabling information distillation by analyzing static deviations and dynamic fluctuations. Second, restraint graph attention network (RGAT) is proposed to extract local structural information, generating dynamic factors that feedback into the spatio-temporal static graph (STSG) to construct the STDG for observation and control of experts. Thirdly, to address degradation variability and long sequence dependency, the reversible instance normalization Transformer (RevIN-Former) explores the temporal evolution of the STDG from a global perspective, achieving accurate RUL prediction. Finally, the effectiveness of the proposed method is validated through three accelerated life experiments on bearings, which reduces the RUL prediction error by 88% compared to other methods.

Accurate causal discovery is of great significance for data-driven root cause diagnosis. For multivariate complex industrial processes, traditional methods rarely conduct causal discovery of multiple causes and single effect from the perspective of quantifying causal strength. In this regard, this article proposed an anomaly tracing method based on attention-based postnonlinear (PNL) causal model. The attention mechanism is introduced into the multivariate PNL model to quantitatively calculate the causal contribution of each cause to the effect. To address the issue of distinguishing inherent causal relationships from anomaly propagation paths, a comparative causal diagram analysis method is proposed. It analyzes the changes in attention weights of the cause variables and effect variable under normal and abnormal conditions to determine the anomaly propagation paths. To tackle the problem of multiple root nodes in causal diagram, a root cause scoring method is proposed. The feasibility of the proposed method is demonstrated through simulation and real industrial case study. Compared with existing and ablation methods, the proposed method can provide the root cause more promptly and accurately, as well as identify anomaly propagation path that align with mechanism analysis.

In recent years, data-driven soft sensor modeling methods have been widely used in industrial production, chemistry, and biochemical. In industrial processes, the sampling rates of quality variables are always lower than those of process variables. Meanwhile, the sampling rates among quality variables are also different. However, few multi-input multi-output (MIMO) sensors take this temporal factor into consideration. To solve this problem, a deep-learning (DL) model based on a multitemporal channels convolutional neural network (MC-CNN) is proposed. In the MC-CNN, the network consists of two parts: the shared network used to extract the temporal feature and the parallel prediction network used to predict each quality variable. The modified BP algorithm makes the blank values generated at unsampled moments not participate in the backpropagation (BP) process during training. By predicting multiple quality variables of two industrial cases, the effectiveness of the proposed method is verified.

To solve the problem of incipient fault detection, a fault targeted gated recurrent unit-canonical correlation analysis (CCA) method is proposed. First, this article proposed fault targeted gated recurrent unit (FTGRU) to establish a temporal feature extraction model. The features extracted by FTGRU are more sensitive to the incipient faults, thus increasing the accuracy of the fault detection model. Then, a fault detection model is established by CCA method. In addition, in order to ensure the universality of the detection model, a multilayer fault detection strategy is proposed. At the first layer, the basic CCA model is used. When no fault is detected at this layer, the second layer fault detection method is enabled. In the second layer, the proposed FTGRU-CCA method is used. Finally, the proposed method and detection strategy are validated by two different industrial cases.

Recently, domain adaptation (DA)-based fault diagnosis approaches have been actively studied in chemical processes to build a reliable fault diagnosis model for a new operating mode (i.e., target domain) by making use of labeled data from a historical mode (i.e., source domain). However, this raises privacy concerns, such as data leakage, since industrial data contains sensitive production information. Moreover, preprocessed source and target data used to train an effective target model will result in additional computational costs. Therefore, it is crucial to develop a novel privacy preserving DA-based fault diagnosis approach that can improve the diagnosis performance for a new mode and protect the privacy of a historical mode simultaneously. To this end, fault diagnosis is formulated as thesource-free DAproblem and proposes a temporal attention source-free adaptation (TASFA) algorithm, which only utilizes the pretrained source model and unlabeled target data to learn a diagnosis model. Specifically, for the time-series process, an attention mechanism is designed to capture and leverage the temporal correlations between source and target domains by extracting the most transferable information from the target time series. Empirical results on both the Tennessee Eastman process and the continuous stirred tank reactor demonstrate the effectiveness and efficiency of TASFA.

To monitor the plant-wide process finely, a novel distributed static magnitude-dynamic difference (DSM-DD) method is proposed in this article. First, given the high dimension of the collected data in the plant-wide process, the entire data space is divided into four orthogonal subspaces according to whether the data obey Gaussian distribution and whether it has serial correlation. Second, both the static magnitude and dynamic difference of the data in the four subspaces are used to build the monitoring model. In addition, not only the features within four subspaces are extracted but the correlation between different subspaces is also extracted to construct corresponding statistics. Third, all the statistics with physical significance are put together to form a statistic vector, and the local outlier factor method is used for constructing the synthetic index to determine whether the fault occurs. Finally, the superiority of the DSM-DD method is verified through a typical industrial case.

The dynamic time-varying characteristic has brought great challenges to the plant-wide process monitoring. In this article, a distributed adaptive principal component regression algorithm is proposed for the online indicator monitoring of large-scale dynamic process. First, the distributed data subblocks are constructed according to the process operation units. In each subblock, an adaptive resampling method based on the subblock data and plant-wide data is presented to construct the modeling sample sets, which can extract the process local and global information simultaneously. Afterwards, the indicator-related feature is extracted, and the Bayesian method is used to integrate the subblock monitoring results. Through the collaborative monitoring of the process local and global feature spaces, a refined monitoring decision can be obtained. Finally, a numerical example and Tennessee Eastman process are used to illustrate the effectiveness of the proposed method.

The incipient fault is difficult to detect because of its small amplitude and insignificant impact, however, ignoring such fault may cause irreversible damage to the system. In this article, a hierarchical latent variable extraction and multisegment probability density analysis method is proposed to detect the incipient fault. First, three data subspaces are constructed, which are named dominant, intermediate, and residual spaces, and key latent variables which contain more offline variance or online variation information will be retained. Afterward, the expanded data distribution interval and multiple data segmentsare constructed for the probability density estimation. Based on the improved symmetric divergence index, the distribution distance between the online data and offline modeling data can be evaluated, which has achieved 95.3% and 86.8% average detection rates for the faults in numerical case and Tennessee Eastman process. Finally, a real multiphase flow facility is used to demonstrate the effectiveness of the proposed method.

As industrial technology develops, industrial processes become increasingly large and complex, the traditional methods are difficult to extract features that can represent the condition of the whole process and the effect of fault on quality indicators. Therefore, a novel multiblock decouple convolutional neural network (multiblock DCN) algorithm is proposed. First, key process variables are selected, and process variables are grouped into multiple blocks for the following monitoring. Then, in each block, the proposed DCN constructs a regression model between key process variables and quality indicators, in which the regression model utilizes an improved convolutional neural network as a feature extractor and a decoupling layer as a feature regularizer. Afterward, the monitoring results of each block are integrated into a global monitoring index based on Bayesian theory. After fault detection, variable oblivion contribution plot is presented to locate faulty variables. Finally, two industrial cases are used to demonstrate the effectiveness of multiblock DCN.

Plant-wide processes often have the characteristics of large-scale and multiple operating units. Moreover, due to the closed-loop control, it is possible that the fault never affects product quality. In this article, a novel data-driven method called multisubspace orthogonal canonical correlation analysis (CCA) is proposed, which can not only tell whether the fault occurs but can also judge whether the fault affects the product quality in real time. First, to reduce process analysis complexity and to construct an accurate monitoring model, the original process variable space is divided into four subspaces. Second, the developed orthogonal CCA is conducted on process data and quality data for correlation feature extraction. Then, the quality-related and quality-unrelated features are obtained. Afterward, a total of six monitoring statistics are constructed and integrated to four statistics with physical interpretation via the Bayesian fusion strategy. Finally, the developed method is tested under an industrial case.

In this article, a novel multimode quality-related process monitoring method called multisubspace elastic network (MSEN) is proposed. To make mode partition more precisely, this article develops a novel clustering algorithm based on the neighborhood information and subtractive clustering algorithm. In each single mode, unlike conventional process monitoring models that only focus on whether the fault occurs, a novel elastic network based quality-related process monitoring model is established to judge whether the fault is quality related or not. In addition, to select the most suitable monitoring model for online data, the k-nearest neighbor rule and the voting strategy are applied. Once the fault is detected, the contribution plot method is used in both quality-related and quality-unrelated subspace for fault diagnosis. Finally, the proposed MSEN method is tested under the continuous stirred tank reactor to verify its superiority and advantage.