A common scenario in aluminum electrolysis process is that the collected dataset contains different behavioral policies and some risky policies, such industrial scenario brings new challenges for offline reinforcement learning to learn safety and feasible optimization policy. This article proposes an offline multiobjective reinforcement learning with multicategory policy constraint for operation optimization decision-making (OODM) of the aluminum electrolysis process. The learned optimization policy can surpass the behavior policy while also meet the strict safety requirements of industrial operations. To alleviate the distribution shift problem on multicategory behavioral policy, we present a multicategory policy constraint in the actor network that utilizes the mixture Gaussian variational autoencoder (GMVAE) to implement behavior cloning between the behavioral policy and the learned policy. Based on the actor-critic reinforcement learning architecture, we design two critic networks for multiobjective optimization. Except for the operational performance critic network, an additional safety critic network is introduced to guarantee that the learned policy satisfies the strict safety requirements of industrial operations. We also conduct extensive comparative experiments on the real-world aluminum electrolysis process. Experimental results demonstrate that the proposed method can achieve superior performance against the other offline reinforcement learning algorithms.

The industrial flotation process involves multimodal data and cross-procedural knowledge, presenting significant challenges for knowledge system management and traditional knowledge graph (KG) construction methods. This study develops a large language model (LLM)-driven framework to construct a domain-specific multimodal KG. The Flotation Knowledge Tree ontology is designed as the structural backbone of the KG to organize multi-level operations and interactions while integrating heterogeneous flotation data. To enhance the LLM's domain adaptability, adaptive prompt optimization is proposed to iteratively refine the extraction template with flotation-specific examples and performance feedback, enabling more accurate triple extraction. Additionally, a fuzzy information entropy-driven Retrieval-Augmented Generation (RAG) method is proposed, leveraging fuzzy logic to assign weights to key terms and numerical contexts to preserve causal relationships and semantic integrity under data uncertainty. Furthermore, we established a two-stage LLM-driven pipeline to generate initial triples using a LoRA fine-tuned lightweight LLM with the optimized prompt and fuzzy RAG context, followed by refinement with a large-scale LLM for precision and ontology alignment. Validated triples are mapped onto the Flotation Knowledge Tree to form a structured, high-quality KG with minimal manual intervention. This framework fuses ontology structuring, fuzzy semantic retrieval, and LLM reasoning to enable automated, domain-tailored knowledge assembly for industrial flotation processes. The constructed KG is evaluated using the Flotation Knowledge Tree-based method, achieving a score of 0.92, with a structural score of 0.97 and a semantic score of 0.88, demonstrating robustness and coherence.

Multi-task video classification is a challenging task in video understanding, which is particularly evident in industrial applications. Existing multi-task learning methods enhance task performance by sharing representations of different tasks, which lacks systematic guidance with prior knowledge. In industrial process, abundant prior knowledge is underutilized for video understanding, because the prior knowledge is usually unstructured text which is hard to be embedded for video understanding. To address this limitation, we introduce a novel fuzzy rule graph representation that systematically encodes expert knowledge derived from manual video observations, providing a highly generalized and structured knowledge representation. Correspondingly, a novel multi-channel fuzzy rule graph embedding (MC-FRGE) method is proposed to model relationships among fuzzy rules. Then, a self-supervised two-stream attention autoencoder is constructed for visual feature extraction. Combining the MC-FRGE and the two-stream attention autoencoder, a novel knowledge guided multi-task video classification (KG-MTVC) framework is developed. Finally, experimental results on both the general video understanding dataset HMDB51 and a specialized zinc flotation process video dataset demonstrate the effectiveness of the proposed method. Our implementation is available at https://github.com/Galazxhy/KG-MTVC.git

In order to integrate the advantages of Model Predictive Control (MPC) and artificial neural networks, this paper proposes a control framework based on sequence generation (SGCF). The main idea of SGCF is to first generate an ideal subsequent state trajectory conditioned on the current working condition, and then infer feasible control sequences accordingly. To achieve implicit optimization of the performance metric during sequence generation, different learning weights are assigned to labeled samples. A confidence score based on model interaction is further introduced to evaluate the reliability of each inference and assist training. Theoretical analysis is provided to explain how the sequence generation network learns under performance guidance, promoting better trajectories. The effectiveness of SGCF is validated through experiments on inverted pendulum control and lead-zinc rougher flotation reagent control. Multiple baseline methods are considered, including state feedback control (SFC) models and data-driven MPC approaches. Results demonstrate that SGCF achieves advantages in both control performance and cost, while maintaining favorable inference efficiency. Moreover, SGCF exhibits potential for interpretable reasoning.

In high-stakes process industries, the Basic Process Control System (BPCS) relies on conventional control to enhance productivity, while the Safety Instrumented System (SIS) uses safety functions to maintain safety. Compared to the BPCS, attackers targeting the SIS can modify safety function activation conditions to trigger them prematurely or to evade the activation of the safety function. While various attack detection methods focus on the BPCS, they often overlook the SIS. This can lead to undetected safety breaches, significantly increasing the risk of catastrophic fault. Recent methods face three key limitations that hinder their practical application to SIS. First, both attackers and engineers can exploit the hot update mechanism of SIS to add or modify control logic. However, current methods lack verification for the newly added or modified logic. Second, current methods are unable to assess the rationality of dangerous value ranges. Third, these methods struggle to distinguish between faults and attacks, making it difficult to determine the appropriate time to activate the SIS’s safety function. To overcome these limitations, we propose SecureSIS, a method for securing SIS safety functions by leveraging the safety attributes of the SIS and incorporating information from the BPCS. The core of SecureSIS includes: 1) using the safety attributes of the SIS to verify automatically extracted candidate control logic detection rules; 2) utilizing information from the BPCS to verify automatically extracted candidate value range detection rules; and 3) distinguishing between safety function attacks and industrial process faults with validated rules and integration of process data from BPCS. Our scheme was evaluated using a Tricon SIS controller deployed on a gas pipeline network platform. The results indicate that SecureSIS achieved 97.3% accuracy in detecting data injection attacks and a detection accuracy of 96.0% for control logic modification attacks. Compared with the other representative detection approaches, our scheme has better detection performance.

Accurate identification of multimodal Superheat Degree (SD) plays a critical decision-making role in Aluminum Electrolysis Process (AEP). Because the labeled SD data are scarce and annotation is expensive in real-world AEP, it is a challenge to develop a well-behaved SD identification model. In this paper, a Contrastive Anchors-based Label Propagation (CALP) algorithm is proposed to construct a Deep Semi-Supervised Learning (DSSL) model, called CALP-DSSL, for better identifying SD by utilizing large-scale unlabeled data and limited labeled data. Specifically, to improve the reliability of affinity graph and its affinity matrix, positive anchors and negative anchors are generated by estimating the uncertainty of label predictions, which can guide the correct direction of inferring pseudo-labels. For tackling the unsupervised domain adaptation problems existing in AEP, we propose a Variational Information Domain Adaptation (VIDA) module using the pseudo-labels generated by CALP to fine-tune the deep Variational Information Bottleneck (VIB) network. Finally, the overall CALP-DSSL model is trained by the Mini-Batch Incremental Learning (MBIL) technique in local level. It matches the nearest neighbors based on batch embedded features, which provides more distinct information flow during subsequent label propagation to construct the affinity graph. Benchmark datasets verify the superiority of CALP algorithm. Case study on a real-world AEP shows that CALP-DSSL model improves the accuracy of SD identification over other state-of-art DSSL methods. Our source code is available at https://github.com/wjiecsu/CALP-DSSL.Note to Practitioners—The focus of this paper is to develop a well-behaved SD identification model based on the improved deep semi-Supervised learning (CALP-DSSL) model. In this model, a Contrastive Anchors based-Label Propagation (CALP) algorithm is used to predict pseudo-labels to construct DSSL model for improving the reliability of the generated pseudo-labels. In this way, a Variational Information Domain Adaption (VIDA) module is constructed to solve the domain shift problems existed in the unstable working conditions. Moreover, the overall model is trained by Mini-Batch Incremental Learning (MBIL) strategy to build better underlying representation for supporting effective label propagation.

Edge preserving filter is the basis of many computational photography and image processing. This can be achieved by global optimization method or local filtering method. Generally, the filtering results of global optimization methods are better than that of local filtering methods, and local filtering methods usually run much faster than global optimization methods. In this paper, a globally optimized method called iterative self-guided image filter (isGIF) is extended based on the assumptions of the guided image filter (GIF), which can produce high-quality edge-preserving filtering results by using the input image itself as the guidance image. Some comparisons with other edge-aware filters are presented to show the advantages of our method. Extensive experiments demonstrate that our filter generates images with better visual quality, while reducing/avoiding halo artifacts in the final image, and the running time is competitive.

Deep learning models are increasingly being used as effective techniques for industrial process modeling. However, decisions generated from deep learning models can hardly to interpret and cannot provide convinced results to users. To break the traditional tradeoff between accuracy and interpretability of deep neural network for industrial process modeling, we propose a deep probabilistic network-based knowledge-guided framework, which injects external knowledge into the deep neural network to guide its training process. In this framework, a deep probabilistic regression model (DPRM) is first developed to learn Gaussian latent feature representation yet establish regression relationship. Then, the external knowledge represented by fuzzy rules, which can evaluate the fitness of current network output and characterize constraints of the process, is encoded into a same structured Gaussian latent feature representation. We propose to inject the feature representation of external knowledge into DPRM using Kullback–Leibler divergence between two Gaussian distributions. The proposed knowledge-guided framework is evaluated on the Tennessee Eastman Process and the real-world aluminum electrolysis process. Experimental results highlight that our proposed approach achieves better prediction performance than the compared methods, also interprets the results.

Numerous studies on learning algorithms have been done in the neural networks community, however concerning to training strategy is limited. In this article, to enhance the performance of fuzzy neural network (FNN) in industrial process identification, an FNN based on two-stage training (TSTFNN) is proposed. First, to initialize TSTFNN, we propose a supervised clustering method. It determines the initial centers and number of neurons for TSTFNN through clustering part of samples based on their labels. Then, the gradient-based algorithm is employed to further learning the parameters of TSTFNN. To improve its generalization performance, we developed a two-stage training scheme. In the first stage, a standard 10-fold cross-validation is executed to learn parameters. The samples whose prediction errors exceed a preset threshold are utilized to train the network in the second stage. According to Lyapunov criterion, we show the convergence of TSTFNN if the learning rate is selected in an appropriate range. The effectiveness of TSTFNN is validated by simulations on two benchmark problems. The proposed TSTFNN performs better and outperforms the state-of-the-art. Compared with supervised long short-term memory, root mean square error is decreased by 17%. Then, we apply it to the industrial rougher flotation process for soft sensor modeling. The experimental results show that TSTFNN achieves a satisfactory prediction performance.

Improving the robustness of the soft sensor model of industrial processes is an important yet challenging problem for a large amount of noise interference and missing data in practical industrial data. In this article, an adversarial training-based deep supervised variational autoencoder (Adv-DSVAE) is proposed to enhance the performance of industrial soft sensor models. Specifically, a supervised variational autoencoder (SVAE) is first designed to extract the quality-relevant feature representation. Then, a deep SVAE (DSVAE) model is constructed by stacking the hidden features extracted by SVAE, such that a high-level output-related feature representation can be captured. In this way, the missing data situation can be handled by the probabilistic latent feature representation extracted in DSVAE. To improve the robustness of a DSVAE-based soft sensor model, an adversarial training method is designed, in which adversarial examples are generated by adding perturbations to the last hidden feature of DSVAE, such that the model can perform well on both clean and perturbed feature representations. We further provide theoretical convergence analysis for the proposed Adv-DSVAE to guarantee its successful practical application. The ablation studies confirm that industrial quality prediction using the adversarial training strategy can ensure better robustness. Case studies on both the debutanizer column process and the real-world aluminum electrolysis process validate the superiority of Adv-DSVAE.

The purpose of texture smoothing is to preserve the prominent structure in the image while smoothing the salient texture. However, the existing methods are difficult to achieve a satisfactory balance between filtering out salient textures and preserving weak edge structures and small structures. To this end, we propose a structure-preserving texture smoothing method via scale-aware bilateral total variation. First, the joint bilateral filter is introduced to construct the window bilateral variation, and combined with the window total variation, a regularizer called the bilateral total variation is formed, which accurately quantifies the characteristics of texture and structure, to finely smooth salient textures while preserving weak edge structures and small structures. Subsequently, we proposed a scale-aware scheme to make the proposed regularizer more powerful in preserving small structures and adopted an optimization scheme to convert the original non-convex optimization problem into a least squares regression problem. The effectiveness of the proposed regularizer is verified in the dataset. The experimental results demonstrate the superiority of the proposed method in texture smoothing and other applications compared to other state-of-the-art approaches.

Industrial process data have the characteristics of less label, multimode, high dimension, containing noise, and mixing with outliers, which increase the difficulty of mode identification and anomaly detection in process monitoring using limited labeled data. In this article, to address the effect of these adverse factors, a semi-supervised discriminative projective dictionary pair learning (SSDP-DPL) is proposed for industrial process monitoring. First, a semi-supervised dictionary pair learning framework with a class estimation regularization term is developed to address the problem of lacking labeled data. Based on unlabeled data and their reconstruction errors, the class estimation regularization term is designed to obtain a discriminative extended synthetical dictionary, mining the hidden discriminative information in unlabeled data and reducing the impact of incorrect class estimation. Second, a sparse constraint is added to the analytical dictionary to enhance the robustness of projective dictionary pair. Then, a class discriminative function term is designed to improve the discrimination of dictionary pair, which ensures the intraclass compactness and interclass separation of coding space. Finally, the synthetical dictionary, the analytical dictionary, and the control threshold are obtained by iterating the dictionary pair to a process monitoring model for anomaly detection and mode identification. The effectiveness of the proposed process monitoring method is verified by the continuous stirred tank heater process and Tennessee Eastman benchmark tests. The proposed method is then applied to a real-world aluminum electrolysis industrial process. Experimental results demonstrate the superiority of our SSDP-DPL in contrast to other existing methods.

Industrial processes often operate in the complex dynamic environment. They challenge the cost-effective and reliable operation of industrial processes. This article proposes a multiobjective optimization and control strategy for the iron removal process to be optimally operated in the dynamic environment. In the optimization layer, multiobjective optimization and local optimization problems, minimizing the process cost and maximizing economic index, are developed to obtain the optimal set-points of the outlet ferrous ion concentrations. In the control layer, model predictive control based on the system dynamic model is constructed to control the outlet ferrous ion concentrations to track the set-points. To reduce the influences from dynamic disturbances and improve the tracking performance, we use least squares support vector machine (LSSVM) to establish the correction model. Fuzzy-logic-based compensation is proposed to compensate the set-points according to the priori and posteriori information. Finally, simulations using real-world plant data are carried out to verify the effectiveness of the proposed control strategy. The simulation results demonstrate that the proposed control strategy achieves a satisfactory tracking performance with less process consumption. It improves the iron content in the goethite precipitate, which has more commercial value. The proposed control strategy can make the plant more profitable.

Satisfying the process technical requirements while optimizing the control of oxygen and zinc oxide is the main task for optimal operation of the iron removal process. However, due to the complicated mechanism and varied production conditions, the process is difficult to achieve optimal operation with the simple controller. The manual control, as a result, is extensively used in practice. In this paper, we develop an optimal setting and control (OSC) system for the iron removal process to achieve its technical requirements with minimal process consumptions. The OSC system is composed of a neural network soft-sensor, an optimal setting module, an optimal controller, and a fuzzy-logic-based compensator. First, we define the oxygen reaction efficiency (ORE) to measure the difference between the theoretical oxygen amount and its actual amount. Due to the ORE cannot be measured online and it varies with the production conditions, an adaptive weight radial basis function neural network soft-sensor is designed to estimate it. The adaptive weight adjusting method contributes to improve the adaptability of the soft-sensor, and its convergence is discussed. The optimal setting module provides the set-point of outlet ferrous ion concentration for the optimal operation of every reactor. The steady-state optimal control of oxygen and zinc oxide is then established, and the compensator compensates the control inputs utilizing the feedforward and feedback information. Finally, simulations validate that the ORE is important for the optimal control of the process. Furthermore, industrial experiments are presented to verify the effectiveness and potential of the proposed strategy.

In the iron removal process, which is composed of four cascaded reactors, outlet ferrous ion concentration (OFIC) is an important technical index for each reactor. The descent gradient of OFIC indicates the reduced degree of ferrous ions in each reactor. Finding the optimal descent gradient of OFIC is tightly close to the effective iron removal and the optimal operation of the process. This paper proposes a coordinated optimization strategy for setting the descent gradient of OFIC. First, an optimal setting module is established to determine the initial set-points of the descent gradient. The oxygen utilization ratio (OUR), an important parameter in this module, cannot be measured online. Therefore, a self-adjusting RBF (SARBF) neural network with an adaptive learning rate is developed to estimate the OUR. The convergence of the SARBF neural network is discussed. Then, a coordinated optimization strategy is proposed to adjust the set-points of the descent gradient when the measured OFICs drift away from their desired set-pints. If the final OFIC does not satisfy the process requirements, a compensation mechanism is developed to provide a compensation for the set-points of the descent gradient. Finally, industrial experiments in the largest zinc hydrometallurgy plant validate the effectiveness of the proposed coordinated optimization strategy. Our strategy improves the qualified ratio of the OFIC and the quality of the goethite precipitate. More profit is created to the iron removal process after our strategy is applied.

As part of the zinc hydrometallurgy plant, the iron removal process is a complex system with four cascaded reactors. Tighter process-index control is difficult to achieve due to the complicated, long, and time-varying removal process. The control performance is also affected by the quality of the ore source and external disturbances. Little research is documented in the literature to address these difficulties and manual control is widely used. An innovative hybrid control strategy is developed to control the iron removal process' indices within narrow ranges with minimum cost of additive regents, including oxygen and zinc oxide. This strategy is composed of an optimal setting model, a model-based optimal controller, an integrated prediction model, a fuzzy-logic-based feedforward compensator, and a model feedback adjustor. The optimal setting model automatically optimizes the set-points of the process indices under different production conditions. To achieve the process requirements with minimal cost, the model-based optimal controller is designed. The integrated prediction model is established to provide a more accurate on-line prediction of the process indices by integrating the mechanism prediction model and an error compensation model based on the least-square support vector machine. Based on the predicted process indices, the compensator is developed for the optimal controller. The adjustor provides a parameter adjustment mechanism. Four-week-long industrial experiments in the largest zinc hydrometallurgy plant in China show that the control strategy can not only improve the process-indexes control performance, but also save 6.55% oxygen and 4.61% zinc oxide consumptions, which translates to 222 858 m3oxygen and 1236 t zinc oxide per year (a saving of about $570 000). The hybrid control strategy can be extended to cover other similar processes in the zinc hydrometallurgy and other industries.