With the rapid development of the large-scale industry process, multiple operation units and dynamic time-varying feature bring new challenges to the effective indicator monitoring. In this paper, a novel adaptive local-global principal component regression method is introduced for the plant-wide process monitoring. Firstly, the process is decomposed into multi-subblocks and the adaptive modeling sample set is constructed in each subblock, which can retain the process online operation feature. Afterwards, the KPI-related local and global monitoring models are constructed concurrently, and the indicator state can be refined identified through the collaborative monitoring of the local and global subspaces. In order to reduce the computation burden during the adaptive monitoring, a dual-control-limit model update strategy is presented. Finally, a designed numerical example and Tennessee Eastman process are used to demonstrate the superiority of the proposed method.

With the rapid development of the large-scale industrial process, multiple operation units and time-varying features bring new challenges to the effective indicator monitoring. In this paper, a novel adaptive local-global principal component regression method is introduced for the plant-wide process monitoring. Firstly, the process is decomposed into multiple subblocks, and an adaptive modeling sample set is constructed for each subblock, which can retain the process online operation feature. Afterwards, the KPI-related local and global monitoring models are constructed concurrently, and the indicator state can be finely identified through the collaborative monitoring of the local and global subspaces. To reduce the computation burden in adaptive monitoring, a dual-control-limit model update strategy is presented. Finally, a designed numerical example and the Tennessee Eastman process are used to demonstrate the superiority of the proposed method.

Ensuring the generalizability of fault diagnosis models is critical for maintaining the long-term safety of industrial systems operating under diverse conditions. This study presents a novel method, termed the Generalizable Class-Consistent Network (GCCNet), designed to enhance diagnostic robustness under previously unseen operating conditions. Specifically, GCCNet incorporates a mutual information based feature disentanglement mechanism to extract task-relevant representations. To further promote feature invariance, auxiliary samples are constructed using same-class fault data under different excitation intensities, and a class-consistency regularization is applied during training to enforce consistent predictions. This guides the network to purify task-relevant features into transferable and robust representations. Extensive experiments conducted on the Tennessee Eastman process and industrial dataset validate the effectiveness and generalization ability of the proposed method.

Abstract To address the problem of balancing sensitivity and robustness in incipient fault detection (IFD) for large‐scale processes, a distributed monitoring framework based on causal‐oriented kernel regularized canonical variate analysis (CO‐KRCVA) is proposed. The framework divides operational units into sub‐blocks according to the sequence of production. To reveal the information transfer relationships and the underlying mechanisms of dynamic changes among variables, a causal difference matrix (CDM) using transfer entropy estimation is constructed for each sub‐block. Based on this, a CO‐KRCVA model is established, which specifically captures targeted minor variations of process variables to enhance the sensitivity and reliability. Its regularization strategy further overcomes the effects of random noise or interference to reduce false alarms. In addition, the Wasserstein distance (WD)‐based statistic is employed to sensitively reflect distribution shifts from both probabilistic and geometric perspectives. To facilitate global detection, Bayesian inference is used to fuse the results from local models. Finally, the proposed method and monitoring framework are validated for effectiveness on different types of faults in the vinyl acetate monomer (VAM) process and Tennessee Eastman (TE) process. Experimental results show that compared with CVA, the proposed method improves the early fault detection rate by 28.47%, 46.21%, and 12.6%, respectively, for the three fault types in the TE process, while reducing the false alarm rate by 0.52%, 1.44%, and 1.36%.

While unsupervised domain adaptation (UDA) aligns overall distributions, semi-supervised domain adaptation (SSDA) often produces class-level mismatches in chemical fault diagnosis-insufficient labeled target data and remaining domain gaps cause target samples to be linked to incorrect source classes, harming diagnostic performance. To address this, we propose CLSSDA, a contrastive learning-based SSDA framework designed for multivariate sensor data in industrial chemical processes. First, pseudo labels are assigned to high-confidence unlabeled target data to filter unreliable labels for outliers and uncertain samples. Next, sample pairs are constructed based on source data, along with labeled and pseudo-labeled target data for contrastive learning (CL). To improve the efficiency of model training, weights are assigned to positive sample pairs, and valuable negative sample pairs are selectively included. Then, Intra-and Inter-Domain Contrastive Alignment modules are introduced to reduce the disparity between labeled and unlabeled target distributions, while narrowing the source-target distribution gap. Experiments on representative chemical process benchmarks demonstrate that CLSSDA significantly improves fault diagnosis accuracy and robustness under limited labeled data and severe operating-condition shifts.

Accurate sensor fault detection and diagnosis in industrial processes are essential for maintaining system reliability and operational safety. However, under closed-loop control conditions, the presence of inherent nonlinearities, strong multivariate coupling, and subtle fault signatures poses significant challenges to effective monitoring. To address these issues, this article proposes a novel self-supervised orthogonal decoupling autoencoder (SODAE). The core component of the model, the orthogonal decoupling autoencoder (ODAE), achieves robust feature extraction by decoupling interdependent sensor relationships through the optimization of weight matrices on the Stiefel manifold, supported by theoretical analysis. In parallel, a self-supervised learning mechanism is integrated via adversarial training to enhance sensitivity to minor sensor anomalies. The proposed SODAE is evaluated on a simulated benchmark with customized sensor fault scenarios, as well as on two complex real-world water treatment processes. Experimental results demonstrate that SODAE achieves superior performance in both sensor fault detection and diagnosis under challenging industrial conditions.

Industrial processes often operate under multiple operating conditions. Monitoring for unusual modes is a challenge due to the limited availability of samples. In this article, a hierarchical key-historical-information transfer (HkhiT) and subspaces monitoring method is proposed for processes with multiple operating conditions, where historical operating conditions knowledge is transferred hierarchically for limited-sample fault detection. First, a well-conditioned covariance matrix is constructed within the target sample subspace; this procedure mitigates correlation-estimation uncertainty under limited-sample conditions. The principal monitoring subspace then integrates historical correlation information to regularize the target correlation structure. Subsequently, the complementary monitoring subspace extracts latent variables that are sensitive to process disturbances, thereby achieving more comprehensive monitoring of online variations. With limited training data, the decision fusion of the principal and complementary monitoring subspaces provides robust monitoring performance. Finally, the effectiveness of the proposed method is demonstrated through a numerical case and the Tennessee Eastman process.

With the complexity and intelligence of the industrial process, the identification of faults in the actual process plays a crucial role in ensuring production safety. The traditional fault identification strategies have the problem that similar characterized faults are unable to be accurately identified. Motivated by the limitations, a novel sample-optimized adaptive perceptual enhanced graph neural network (SOAP-EGNN) for large-scale process fault identification is proposed. Initially, process mechanism knowledge and process data correlation are injected into the modeling approach through graph neural networks, and the transmission of information based on the enhanced attention mechanism is introduced to describe the quantitative relationships between process variables at a fine-grained level based on the adaptive perception strategy. Subsequently, to achieve better intra-class compactness and inter-class separability in feature representation, our designed sample-optimized feature processing strategy (SOFPS) is applied. Furthermore, to enhance the robustness and generalization capability of the model during training, a label smoothing regularization (LSR) strategy is incorporated. This approach effectively mitigates the risk of overfitting by introducing a degree of uncertainty into the label space, thereby encouraging the model to learn more discriminative and stable features. Ultimately, the efficacy and superiority of the SOAP-EGNN algorithm are thoroughly validated through comprehensive simulation experiments conducted on the Tennessee Eastman process (TEP).

In contemporary industrial processes, factors such as raw material fluctuations and noise interference lead to significant changes in the statistical characteristics of data over time, resulting in nonstationary behavior. This nonstationary tends to obscure fault-related information in industrial systems, posing severe challenges for quality-related fault detection. This article proposes a method named reversible instance normalization temporal autoencoder (TAE) canonical correlation analysis (CCA) for quality related fault detection. First, this method quantifies the interdependencies among variables to select process variables that are highly correlated with quality indicators. Subsequently, a reversible normalization module dynamically adjusts normalization parameters to achieve local stationarization of process data. Furthermore, a TAE is utilized to extract temporal features from processed data. Then, a CCA model is established by integrating quality indicators, enabling efficient monitoring of quality indicator under nonstationary conditions. Finally, the proposed method was tested and validated in two real industrial cases.

Abstract Fault diagnosis plays an important role in process monitoring under closed‐set settings. However, the unknown faults may be misdiagnosed as one of the known classes with overconfidence, leading to unreliable decisions and process unsafety. Besides, to distinguish multiple unknown fault modes, existing methods need to be completely retrained with sufficient measurements under faulty conditions. However, model retraining is time‐consuming and the labelled faulty data is rare in real industrial processes. To solve the above problems, a novel uncertainty‐aware evidential prototype network (UEPN) with few‐shot class‐incremental learning (FSCIL) for open‐set fault diagnosis (OSFD) is proposed. First, the prototype network (PN) based feature extractor is proposed to learn discriminative embeddings for known classes and preserve more embedding space for unknown ones. An uncertainty‐aware evidential classifier is further developed with uncertainty quantification to identify unknown faults and avoid overconfident misclassification. Then, to continuously identify multiple unknown fault modes, the FSCIL procedure is designed, which enables UEPN to be incrementally updated with few‐shot data of an unknown fault class. Finally, the proposed method is evaluated on the Tennessee Eastman process (TEP) and the vinyl acetate monomer (VAM) plant model. The proposed method achieves up to 10.8% higher fault diagnosis rate (FDR), 5.4% lower false positive rate (FPR), and 8.5% higher true positive rate (TPR) than the comparison methods for OSFD task of continuously distinguishing multiple unknown fault modes.

Background: Unsupervised domain adaptation (UDA) has been widely studied for industrial fault diagnosis under varying operating conditions. However, in cross-enterprise scenarios, source data is often not available due to privacy and security constraints. To address this limitation, source-free unsupervised domain adaptation (SFUDA) has been proposed, which adapts the pretrained source model to the target domain without access to any source data. Methods: Methods To facilitate target domain adaptation under the SFUDA setting, we propose a Core-to-Edge (C2E) framework that progressively exploits target samples from high-confidence core instances to uncertain edge ones. Specifically, we design a Unified Confidence Score Estimation scheme that combines prediction probability and neighborhood consistency to obtain more reliable confidence scores and pseudo labels. Based on these scores, we further introduce a Dynamic Class-Balanced Pseudo-Labeling strategy that gradually enlarges the selected set in an easy-to-hard manner while mitigating class imbalance. To further exploit low-confidence target data, we propose an Edge-sample Mining strategy. Specifically, we incorporate a mean-teacher framework to leverage unlabeled edge samples via consistency regularization, thereby promoting more robust target domain adaptation. Significant findings: Extensive experiments on two industrial benchmark datasets demonstrate the effectiveness and superiority of the proposed C2E framework.

Root cause diagnosis in complex industrial processes is challenged by the smearing effect of contribution plots, the nonlinearity and indirect causality issues of Granger causality and transfer entropy, and the separation between contribution analysis and causal inference. To address these limitations, this paper proposes a semi-frozen auto-encoder (SFAE) with orthogonal constraint and integrated causal-contribution graph. Offline, an auto-encoder with orthogonality constraint is trained on normal data to decouple variable contributions and construct a directed causal graph from the trained weights. Online, the encoder is frozen and only the decoder is updated; the decoder weights directly provide variable contributions. The top-three contributing variables are selected, and the subgraph among them extracted from the offline causal graph reveals the fault propagation path and root cause. Unlike existing methods that treat contribution analysis and causal inference separately, our framework unifies them in a single auto-encoder architecture. Experimental results demonstrate superior root cause localization accuracy, robustness to measurement noise, and interpretable fault propagation paths.

Root cause diagnosis in complex industrial processes is challenged by the smearing effect of contribution plots (CPs), the nonlinearity and indirect causality issues of Granger causality (GC) and transfer entropy, and the separation between contribution analysis and causal inference. To address these limitations, this article proposes a semifrozen autoencoder (SFAE) with orthogonal constraint and an integrated causal-contribution graph. Offline, an autoencoder (AE) with orthogonality constraint is trained on normal data to decouple variable contributions and construct a directed causal graph from the trained weights. Online, the encoder is frozen, and only the decoder is updated; the decoder weights directly provide variable contributions. The top-three contributing variables are selected, and the subgraph among them extracted from the offline causal graph reveals the fault propagation path and root cause. Unlike existing methods that treat contribution analysis and causal inference separately, our framework unifies them in a single AE architecture. Experimental results demonstrate superior root cause localization accuracy, robustness to measurement noise, and interpretable fault propagation paths.

Existing data-driven process monitoring methods rely on fixed model parameters after offline training. However, factors such as equipment aging and changes in production strategies often cause process data distribution shifts, leading to high false alarm rates and model mismatch. Although adaptive monitoring methods update models with new data, they generally focus only on current data, which may overwrite previously learned knowledge. To address these issues, this study proposes a novel adaptive fault detection framework named Dynamic Adaptive Prototype-based Continual Fault Detection (DAPCFD). The proposed method alleviates catastrophic forgetting during incremental learning from two aspects. First, a dynamic prototype buffer is constructed to maintain and update representative historical features. Second, elastic weight consolidation is employed to preserve important model parameters during incremental learning. Moreover, an adaptive dynamic threshold updating strategy is developed to further reduce the false alarm rate. A benchmark simulation platform and a Fluid Catalytic Cracking-Fractionator Unit case are employed to illustrate improved robustness and fault detection performance under dynamic operating conditions.

Data-driven fault diagnosis methods often assume that training and testing data follow identical distributions. However, in chemical processes, variations in operating conditions frequently lead to discrepancies between training and testing data, thereby reducing diagnostic accuracy. While domain-adaptive methods attempt to mitigate these discrepancies, they generally require prior knowledge of operating modes not encountered by the model, thus limiting their applicability. In contrast, domain generalization (DG) techniques have gained attention for their ability to generalize models to unseen modes those the model has never experienced without prior knowledge. Most existing DG methods aim to learn domain-invariant features that are unaffected by operating condition variations from historical data. However, this is challenging because process data typically contain both domain-specific information (e.g., steady-state operating points and production ratios) and domain-invariant information, and the former is often equally valuable for fault diagnosis. To address this challenge, we propose a novel method named domain disentanglement network with entropy-driven weight learning strategies (DDNEL) for diagnosing faults of unseen modes in chemical processes. The entropy optimization algorithm was employed to guide DDNEL in extracting domain-invariant features and domain-specific features, respectively. In the inference phase, an entropy-driven feature fusion mechanism is devised to fuse invariant and specific features for diagnosing the faults of unseen modes. Empirical results on the continuous stirred tank reactor (CSTR), the Tennessee Eastman process (TEP), and a real-world industrial case (Cranfield multiphase flow process) demonstrate the effectiveness of the proposed DDNEL approach.

Process monitoring and root cause diagnosis (RCD) are significant for maintaining process safety and ensuring product quality in industrial processes. Existing RCD methods have achieved great success under linear and stationary assumptions, which limits their application in complex industrial processes. In addition, causality analysis of the performance indicator (PI) helps to precisely identify the path of propagation of faults and locate the root cause that leads to performance degradation. However, PI is not online measurable, which makes it difficult to achieve PI-related RCD in time. A novel causality-driven sequence-to-sequence gated recurrent unit (CSGRU) is proposed for PI-related RCD to address these issues. The proposed method is built under the distributed process monitoring framework based on a PI-related process decomposition strategy to first locate the faulty unit. CSGRU is integrated with the concept of Granger causality (GC) to learn nonlinear and dynamic causal dependencies. Sequence-to-sequence multi-task learning is introduced to avoid time-consuming pairwise causal analysis and improves the predictive accuracy even under strong sparsity constraints. The predictive contribution statistic is built to obtain the real-time faulty causal graph, which contains the causal impacts on PI without using online PI data. Finally, through a reverse causal inference from effect to cause, the fault propagation path is identified backward from PI, and the root cause is subsequently located, which leads to the degradation of PI. The effectiveness of the proposed method is validated on two benchmarks, the Tennessee Eastman process (TEP) and the vinyl acetate monomer (VAM) plant model, and a real industrial application on the three-phase flow facility (TPF).

The fault diagnosis method based on stacked autoencoder (SAE) usually needs sufficient labeled data to train an effective diagnosis model, and it is difficult to obtain labeled data in industrial process. To solve the above problems, a semi-supervised fault diagnosis method based on feature alignment-ensemble learning stacked autoencoder (FA-ELSAE) was proposed. First, samples from the same class should follow the same distribution since the method is based on stacked autoencoders. This means that while training the SAE diagnostic model, this restriction is introduced to the model's loss function, maximizing the information of unlabeled data, enhancing the diagnostic model's capacity for generalization, and decreasing its overfitting. Then, ensemble learning is included into the SAE diagnostic model framework to enhance the method of false labeling unlabeled samples, lower the likelihood of false labeling unlabeled samples, enhance the quality of false labels, and enhance the model's diagnostic efficacy. Lastly, an industrial procedure was used for test validation.