Modern industrial systems often operate under complex dynamics and strict reliability constraints, demanding a timely and precise fault diagnosis with efficient root cause analysis to ensure operational safety and minimize downtime. However, the inherent uncertainties and complexities of industrial processes present significant challenges for conventional diagnostic approaches. Specifically, even minor anomalies can escalate into critical incidents, while process uncertainties frequently induce distribution shifts, leading to novel fault types that complicate fault detection and diagnosis. To address these challenges, this article proposes a novel industrial flow topology-induced semi-heterogeneous graph perception network (IFT-SHGPN) model for class-incremental fault diagnosis of complex industrial processes. By embedding the physical topology of industrial processes into a semi-heterogeneous graph perception network (SHGPN) and incorporating gradient-weighted class activation mapping (Grad-CAM), the proposed approach demonstrates strong capability in effective class-incremental fault recognition and interpretable root cause analysis. Rigorous experiments on the Tennessee Eastman process (TEP) and a multiphase flow facility process under various operation conditions showcase the superiority of IFT-SHGPN over existing methods. The proposed approach achieves high diagnostic accuracy for both historical and emerging fault categories while enabling efficient root cause identification with low computational overhead, making it particularly suitable for deployment in resource-constrained industrial environments.

Modern industrial processes often exhibit non-Gaussian, multimodal, and time-varying behaviors due to fluctuating operating conditions, raw material variations, and external disturbances, such as sensor failures and noise. These complexities pose significant challenges for robust fault detection and identification, critical for ensuring process stability and product quality. To tackle these challenges, this article proposes an innovative robust fault identification framework that leverages a variational Bayesian student’s$t$-mixture model (VBstMM) with a lifelong learning mechanism (LLM), named VBStMM-LLM. VBStMM-LLM is particularly effective for multimodal, time-varying industrial processes due to two core innovations: first, a robust mixture modeling approach grounded in the Student’s$t$-distribution with variational Bayesian inference, which accurately captures complex process dynamics while mitigating the impact of noise and outliers; and second, the innovational integration of LLM that enables adaptive model updates, preserving historical knowledge, and preventing catastrophic forgetting. Extensive experiments on the Tennessee Eastman benchmark process and the ventilation exhaust fans utilized in a continuous casting line at a Chinese steel smelting facility demonstrate the framework’s superior performance. For the three operational modes of the TE process, VBStMM-LLM achieved fault identification accuracy of 94.5%, 93.6%, and 97.7%, respectively. Remarkably, for the fault diagnosis of VEF task, it achieved a perfect rate of 100%. These results highlight VBStMM-LLM’s exceptional robustness, accuracy, and adaptability, underscoring its potential for widespread industrial applications.

This article proposes a robust end-to-end deep learning-induced fault recognition scheme by stacking multiple sparse-denoising autoencoders with a Softmax classifier, called stacked spare-denoising autoencoder (SSDAE)-Softmax, for the fault identification of complex industrial processes (CIPs). Specifically, sparse denoising autoencoder (SDAE) is established by integrating a sparse AE (SAE) with a denoising AE (DAE) for the low-dimensional but intrinsic feature representation of the CIP monitoring data (CIPMD) with possible noise contamination. SSDAE-Softmax is established by stacking multiple SDAEs with a layerwise pretraining procedure, and a Softmax classifier with a global fine-tuning strategy. Furthermore, SSDAE-Softmax hyperparameters are optimized by a relatively new global optimization algorithm, referred to as the state transition algorithm (STA). Benefiting from the deep learning-based feature representation scheme with the STA-based hyperparameter optimization, the underlying intrinsic characteristics of CIPMD can be learned automatically and adaptively for accurate fault identification. A numeric simulation system, the benchmark Tennessee Eastman process (TEP), and a real industrial process, that is, the continuous casting process (CCP) from a top steel plant of China, are used to validate the performance of the proposed method. Experimental results show that the proposed SSDAE-Softmax model can effectively identify various process faults, and has stronger robustness and adaptability against the noise interference in CIPMD for the process monitoring of CIPs.

Cardiovascular diseases are the leading cause of mortality, and accurate segmentation of ventricular regions incardiac magnetic resonance images (MRIs) is crucial for diagnosing and treating these diseases. However, fully automated and accurate right ventricle (RV) segmentation remains challenging due to the irregular cavities with ambiguous boundaries and mutably crescentic structures with relatively small targets of the RV regions in MRIs. In this article, a triple-path segmentation model, called FMMsWC, is proposed by introducing two novel image feature encoding modules, i.e., the feature multiplexing (FM) and multiscale weighted convolution (MsWC) modules, for the RV segmentation in MRIs. Considerable validation and comparative experiments were conducted on two benchmark datasets, i.e., the MICCAI2017 Automated Cardiac Diagnosis Challenge (ACDC), and the Multi-Centre, Multi-Vendor & Multi-Disease Cardiac Image Segmentation Challenge (M&MS) datasets. The FMMsWC outperforms state-of-the-art approaches, and its performance can approach that of the manual segmentation results by clinical experts, facilitating accurate cardiac index measurement for the rapid assessment of cardiac function and aiding diagnosis and treatment of cardiovascular diseases, which has great potential for clinical applications.

Abstract The accurate segmentation of blood vessels plays a crucial role in screening, diagnosis and treatment of multiple diseases. However, current automated segmentation approaches do not pay enough attention to the vascular topology errors (such as mistaking vessel‐breakpoints), resulting in considerable scattered vessel‐fragments in segmentation results. This article proposes a retinal vessel segmentation model using multi‐scale context gating and breakpoint attention mechanism, called MCG&BA‐Net. Specifically, it obtains a feature map containing contextual information of vessels through an introduced multi‐scale context module, and then filters the redundant features and noises by a gated structure to highlight target features. Furthermore, a kind of breakpoint attention module is proposed, which can locate and focus on potential breakpoint areas, thereby facilitating accurate segmentation results of tree‐like fine vessels. Extensive confirmatory and comparative experiments have been conducted on five public datasets, including three benchmark datasets, that is, DRIVE, CHASDB1 and SATRE, and two clinical datasets, that is, fundusimage1000 and RFMID. The AUC scores on the benchmark datasets are 0.9878, 0.9923 and 0.9942, respectively. Among them, the AUC score on CHADEDB1 and STARE outperforms the state‐of‐the‐art results. In addition, experimental results on the two clinical datasets demonstrate strong generalization capability of the propose method, indicating high clinical application values.

Due to technical or economic limitations, timely measuring quality-relevant key performance indicators (KPIs) of complex industrial processes (CIPs), especially the chemical composition-related indexes, is intractable. Process monitoring image sequences (PMISs) usually involve significant information about the operation states and KPIs. Thus, soft sensor-based online KPI inference by incorporating process monitoring variables (TPMVs) and PMISs is more promising. However, the extremely inconsistent sampling rates with different expression forms and concerning aspects between PMISs and TPMVs lead to a great challenge in the soft sensor modeling by combining PMISs and TPMVs. In this article, a self-attention dual-channel deep network (SADCDN)-based soft sensor model for the end-to-end online KPI detection/prediction is proposed. Specifically, one channel adopts the gated recurrent unit (GRU) network to extract intrinsic time-series features in TPMVs, and simultaneously the other channel introduces a novel frame-dilated convolution fusion neural network (FDCFNN) to extract intrinsic spatiotemporal features from PMISs to address the sampling inconsistence between PMISs and TPMVs. Successively, dual-channel network features with different concerning aspects are weighted and fused based on an introduced self-attention mechanism to bridge the gap of sampling rates and concerning aspects between PMISs and TPMVs for the soft sensor modeling. Practical application results on two real industrial processes, the bauxite flotation process and the sintering process of a cement rotary kiln, have demonstrated the effectiveness and superiority of the proposed dual-channel model, laying a foundation for the process optimization of CIPs.

Froth color can be referred to as a direct and instant indicator to the key flotation production index, for example, concentrate grade. However, it is intractable to measure the froth color robustly due to the adverse interference of time-varying and uncontrollable multisource illuminations in the flotation process monitoring. In this article, we proposed an illumination-invariant froth color measuring method by solving a structure-preserved image-to-image color translation task via an introduced Wasserstein distance-based structure-preserving CycleGAN, called WDSPCGAN. WDSPCGAN is comprised of two generative adversarial networks (GANs), which have their own discriminators but share two generators, using an improved U-net-like full convolution network to conduct the spatial structure-preserved color translation. By an adversarial game training of the two GANs, WDSPCGAN can map the color domain of froth images under any illumination to that of the referencing illumination, while maintaining the structure and texture invariance. The proposed method is validated on two public benchmark color constancy datasets and applied to an industrial bauxite flotation process. The experimental results show that WDSPCGAN can achieve illumination-invariant color features of froth images under various unknown lighting conditions while keeping their structures and textures unchanged. In addition, WDSPCGAN can be updated online to ensure its adaptability to any operational conditions. Hence, it has the potential for being popularized to the online monitoring of the flotation concentrate grade.

This article presents a simple yet powerful online flotation process working condition (FPWC) discrimination approach based on the sparse representation of froth images. It learns a local Gabor pattern-based discriminative dictionary with a linear classification model simultaneously for the FPWC identification by solving a sparsity-constrained optimization problem. The proposed method tends to achieve similar and distinct sparse codes of froth images for the same and different FPWCs, respectively, facilitating the accurate FPWC identification. To ensure the adaptability of the FPWC discrimination model, an incremental learning-based online model updating procedure is further derived to monitor the dynamically changing characteristics of FPWCs based on an introduced sparsity discrimination index. The proposed method was validated on an industrial preferential lead-flotation subcircuit process. The prototype monitoring system with extensive confirmatory and comparative experiments shows the effectiveness and superiority of the proposed method, which lays a foundation for the optimal control of industrial flotation processes.

This paper presents a froth image statistical modeling-based online flotation process operation-state identification method by introducing a biologically inspired Gabor wavelet transform in accordance with the physiological findings in the biological vision system. It derived the latent probabilistic density models of these biologically inspired Gabor filtering responses (GFRs) based on a versatile intermediate probability modeling frame, Gaussian scale mixture model. It has demonstrated that both the real and the imaginary representation of GFR obey a Laplace distribution. Accordingly, the amplitude representation of GFR obeys a Gamma distribution. Whereas the phase representation of GFR is an important yet frequently ignored aspect in Gabor-based signal analysis; it is demonstrated to be a periodic distribution and can be expressed by a von Mises-like distribution model. Successively, a local spline regression (LSR)-based classifier that the maps scattered statistical feature points of froth images directly to the operation-state labels smoothly is introduced for the operation-state recognition. Extensive confirmatory and comparative experiments on an industrial-scale bauxite flotation process demonstrate the effectiveness and superiority of the proposed method. Performance effects on different parameter settings, e.g., parameters of Gabor kernel and dimensionalities of multivariate statistical models, are further discussed.

To achieve fine segmentation of complex natural images, people often resort to an interactive segmentation paradigm, since fully automatic methods often fail to obtain a result consistent with the ground truth. However, when the foreground and background share some similar areas in color, the fine segmentation result of conventional interactive methods usually relies on the increase of manual labels. This paper presents a novel interactive image segmentation method via a regression-based ensemble model with semi-supervised learning. The task is formulated as a non-linear problem integrating two complementary spline regressors and strengthening the robustness of each regressor via semi-supervised learning. First, two spline regressors with a complementary nature are constructed based on multivariate adaptive regression splines (MARS) and smooth thin plate spline regression (TPSR). Then, a regressor boosting method based on a clustering hypothesis and semi-supervised learning is proposed to assist the training of MARS and TPSR by using the region segmentation information contained in unlabeled pixels. Next, a support vector regression (SVR) based decision fusion model is adopted to integrate the results of MARS and TPSR. Finally, the GraphCut is introduced and combined with the SVR ensemble results to achieve image segmentation. Extensive experimental results on benchmark datasets of BSDS500 and Pascal VOC have demonstrated the effectiveness of our method, and the comparison with experiment results has validated that the proposed method is comparable with the state-of-the-art methods for interactive natural image segmentation.