Generating a class integration test order (CITO) is essential to reduce the overhead of test stub construction (the primary cost in integration testing) and to ensure system reliability in complex software systems. Although reinforcement learning (RL) has shown promise in automating CITO generation, existing methods suffer from unstable policy learning and limited robustness against structural perturbations and defect injection. These challenges stem from insufficient reward shaping and the lack of reliable oracles for validation. To address these limitations, we propose LM-CITO, a stability-aware RL framework that integrates Lyapunov-guided reward shaping with semantic validation through metamorphic testing (MT). Specifically, we design a Lyapunov energy function over class dependency graphs to promote monotonic structural convergence during training, and define metamorphic relations (MRs) to verify behavioral consistency under controlled perturbations. Extensive experiments on six real-world systems demonstrate that LM-CITO consistently produces more effective policies, yielding CITOs with significantly reduced stubbing costs compared to baseline models. Furthermore, MT verifies the capability of our MRs to detect defects in 19 injected bug variants, confirming the robustness of LM-CITO under various fault-induced perturbations. These results highlight the synergy of stability guidance and MR-based validation, offering an effective, principled solution for oracle-free RL in software testing.

The performance of actuators is essential for ensuring the safe and reliable cruise of unmanned aerial vehicles (UAVs). However, limited data and constrained computational resources pose significant challenges for accurate and timely fault diagnosis of UAV actuators in practice. To this end, this paper proposes a novel lightweight fault diagnosis method termed ensemble domain adaptation with constructive incremental learning (EDA-CIL). First, a cloud feature extraction strategy is developed to adaptively extract fault-sensitive information from vibration signals using cloud entropy theory. Next, the proposed meta domain adaptation (MDA) is gained with node-based constructive incremental learning, which serves as a base classifier leveraging knowledge from the source domain and few-shot target domain data. Specifically, MDA minimizes discrepancies in marginal and conditional distributions across different domain features during this incremental process, benefiting the lightweight and compact structure for domain adaptation. Moreover, the convergence analysis of MDA is given to guarantee the efficacy of cross-domain performance and network compactness theoretically. Finally, to avoid the negative transfer arising from excessive dependence on a single-source domain, the stable diagnosis performance is obtained via the domain energy-based parallel ensemble learning of multiple source actuators. Extensive experimental results demonstrate that the proposed method achieves good accuracy and the fastest speed on hexacopter UAV actuator diagnosis with limited data. In comparison with the bidirectional LSTM-based multi-source transfer learning method, EDA-CIL improves the diagnostic accuracy by 3.45%, 5.61%, 9.34%, 11.86%, 2.29%, and 12.80% across six actuators, while achieving approximately 18 times faster diagnostic speed.

The paper investigates the problem of distributed robust state estimation over wireless sensor networks for non- Gaussian systems under hybrid cyber-attacks, where both local measurements and exchanged information among neighboring nodes are subject to varying deception attacks. To overcome the limitations of the existing correntropy-based approach, an extended multi-kernel correntropy (MKC) is adopted to evaluate the similarity of different elements between stochastic vectors. Meanwhile, confronting the controversy of solving such maximum-a-posterior-like cost functions during the derivations of the posterior estimate, the half-quadratic optimization approach is employed to transform the maximization problem of non-convex functions into two convex optimization subproblems, and the final solution is obtained through an alternating iterative method. In addition, a novel inverse Kalman-type compensation mechanism is proposed within the MKC framework to compensate for the corrupted local exchanged estimations, which is then fused through the weighted average to yield the final estimation for each sensor of the distributed algorithm. Simulation results demonstrate that the proposed algorithms outperform related work in terms of estimation performance and robustness.

Accurate state of health (SOH) estimation of lithium-ion batteries is essential for ensuring the safe and reliable operation of battery-powered systems. Variations in battery types and operating conditions give rise to distribution discrepancies, for which various domain adaptation strategies have been proposed. However, current domain adaptation approaches typically require access to source domain information, including model parameters, model structure, and even the source data. Considering the increasing awareness of data protection and security restrictions, this work proposes a novel source-free foundation model-enabled SOH estimation framework designed for black-box scenarios, where only the access permission to a pre-trained source model and limited labeled target samples are required. First, degradation-sensitive features based on crucial voltage ranges are extracted from source batteries to construct a foundation SOH estimation model, reducing reliance on full-cycle measurements. Second, a novel intelligent adapter model is proposed to bridge the distribution gap between the source and target domains by leveraging an intermediate reference battery, enabling latent feature alignment without access to the source data or internal details of the source model. Finally, a fine-tuning strategy under limited target labels is employed for model adaptation to the target domain. Extensive experiments are conducted on multiple cells and compared with several representative domain adaptation approaches. The results demonstrate that, compared with the most competitive source-free adaptation baseline, the proposed framework achieves approximately 41% lower average RMSE and a consistently higher average R 2 of 0.9657, validating its effectiveness under label scarcity and privacy constraints.

The modeling of process data is essential for the prediction of product quality in industrial processes. However, the multirate process data lead to a mismatch between the input and output of soft sensing models. In addition, the time delays between quality indices and process variables destroy the actual temporal correspondence within the process data and reduce the accuracy of these models. To address these problems, this paper proposes a multirate modeling method for industrial quality index prediction with time delays (MRM-TD) which integrates two key modules: time delay estimation (TDE) and multirate modeling (MRM). Firstly, in the TDE module, a cooperative evolutionary strategy, which effectively combines copula entropy (CE) and evolutionary algorithm (EA), is designed for optimizing the time delay parameters. The temporal coupling relationship within process data is assessed as CE. To ensure that each individual in the offspring population reflects empirical realities, prior knowledge is employed to guide the generation process. Secondly, in the MRM module, an incremental network with cross sampling rate constraint (CSRN) is proposed. The time-local proximity graph of process variables, which ensures the consistency of the prediction of quality indices close to each other in the time domain, is constructed to make full use of the process variables without quality indices to build accurate soft sensing models. Additionally, a node control strategy is designed based on a constructive manner to guarantee the effectiveness of nodes. Finally, the simulation results of real mineral grinding process dataset indicate that the proposed method performs favorably.

Dense medium separation (DMS) process is one of the most effective clean coal technologies. Layered operational optimization control for DMS process typically entails fast-time-scale density adjustment in basic loop process and slow-time-scale ash content control in operational process. Unfortunately, the multi-rate, time-varying and uncertainty problems of DMS process make it difficult to design layered operational optimization control methods. To address these issues, this paper proposes a multi-rate dual-MPC (Model Predictive Control) layered operational optimization control approach. For the basic loop process, a multi-rate MPC controller is designed by employing lifting technique and reconstructing the output prediction vector. For the operational process, a neural learning-based MPC control strategy is developed, incorporating control error entropy to enable online updates of model parameters under uncertain conditions. Finally, experiments were conducted using real data from the DMS process and an industrial-grade controller to verify the effectiveness of the proposed algorithm.

Accurate state estimation (SE) for systems affected by unknown non-Gaussian (NG) noise is significantly challenging, especially when compounded by hybrid cyberattacks. In this article, we investigate the fixed-interval smoothing problem for NG systems under such attacks. Since most attack detectors are susceptible to failure in these scenarios, we propose the flag-bit-based detection mechanism by expanding the measurement equation with a marking signal dedicated to identifying attacks exclusively. To improve the estimation accuracy of NG systems at risk of undetected attacks, the robust forward filtering and backward smoothing are derived by solving the new cost functions defined based on the proposed generalized statistical measure (GSM) with enhanced flexibility, based on which we obtain the new robust Rauch-Tung-Striebel smoother. The sufficient conditions for the convergence of both forward and backward passes are rigorously established, which provides the theoretical support of the proposed estimator in terms of optimality and uniqueness. Extensive simulations validate the lower false detection rate of the proposed detector and the improved estimation accuracy of the proposed smoother compared to related works under various noise and attack conditions.

Accurate and rapid state-of-health (SOH) monitoring plays an important role in indicating energy information for lithium-ion battery-powered portable mobile devices. To confront their variable working conditions, transfer learning (TL) emerges as a promising technique for leveraging knowledge from data-rich source working conditions, significantly reducing the training data required for SOH monitoring from target working conditions. However, traditional TL-based SOH monitoring is infeasible when applied in portable mobile devices since substantial computational resources are consumed during the TL stage and unexpectedly reduce the working endurance. To address these challenges, this article proposes a lightweight TL-based SOH monitoring approach with constructive incremental transfer learning (CITL). First, taking advantage of the unlabeled data in the target domain, a semisupervised TL mechanism is proposed to minimize the monitoring residual in a constructive way, through iteratively adding network nodes in the CITL. Second, the cross-domain learning ability of node parameters for CITL is comprehensively guaranteed through structural risk minimization, transfer mismatching minimization, and manifold consistency maximization. Moreover, the convergence analysis of the CITL is given, theoretically guaranteeing the efficacy of TL performance and network compactness. Finally, the proposed approach is verified through extensive experiments with a realistic autonomous air vehicles (AAVs) battery dataset collected from dozens of flight missions. Specifically, the CITL outperforms SS-TCA, MMD-LSTM-DA, DDAN, BO-CNN-TL, and AS$^{3}$LSTM, in SOH estimation by 83.73%, 61.15%, 28.24%, 87.70%, and 57.34%, respectively, as evaluated using the index root-mean-square error.

To improve the reliability and interpretability of industrial process monitoring, this article proposes a causal graph spatial-temporal autoencoder (CGSTAE). The network architecture of CGSTAE combines two components: a correlation graph structure learning module based on spatial self-attention mechanism (SSAM) and a spatial-temporal encoder-decoder module utilizing graph convolutional long short-term memory (GCLSTM). The SSAM learns correlation graphs by capturing dynamic relationships between variables, while a novel three-step causal graph structure learning algorithm is introduced to derive a causal graph from these correlation graphs. The algorithm leverages a reverse perspective of causal invariance principle to uncover the invariant causal graph from varying correlations. The spatial-temporal encoder-decoder, built with GCLSTM units, reconstructs time series process data within a sequence-to-sequence framework. The proposed CGSTAE enables effective process monitoring and fault detection through two statistics in the feature space and residual space. Finally, we validate the effectiveness of CGSTAE in process monitoring through the Tennessee Eastman process (TEP) and a real-world air separation process (ASP).

Collaborative modeling in industrial processes with similar production procedures and complex industrial parameters is an effective approach to address issues about online soft measurement. However, it is a challenging task for collaborative modeling at the presence of non-independent and identically distributed (Non-IID) data, which is arising from variations in production materials and equipment. In this paper, we propose a cooperative reinforced resilient federated learning for clients with multiple working conditions to solve the Non-IID problem by implementing a personalized strategy centered on local data storage, which can accurately build personalized models suitable for distributed data on a single client. Inspired by the success of multi-agent reinforcement learning (MARL) in solving complex decision problems, we design a personalized aggregation mechanism guided by MARL and modeling process information. This mechanism optimizes model aggregation decisions based on data distribution and quality, thereby improving the effectiveness of the training process and enhancing the overall performance of personalized models. Meanwhile, we design an exploration and reward strategy for federated scenarios with distribution differences, which utilizes the collaborative modeling characteristics and advantages of federated learning to improve the exploration efficiency of reinforcement learning in solving personalized federated problems. Experimental results conducted on industry and benchmark datasets demonstrate that the proposed model surpasses existing methods in collaborative modeling under Non IID scenarios.

Class Integration Test Order (CITO) generation is essential to minimize testing cost in object-oriented software.Traditional methods based on static dependencies often producesuboptimal results, while recent approaches that incorporatedynamic dependencies typically neglect accurate stubbing costestimation and face scalability challenges. We propose DyCITO+,an extension of DyCITO, which originally modeled CITO generationas a Reinforcement Learning (RL) problem using Qlearning.However, DyCITO relies on tabular methods, and thislimits its scalability. DyCITO+ addresses this by introducingthree Deep Reinforcement Learning (DRL) algorithms: DeepQ-Network (DQN), Proximal Policy Optimization (PPO), andAdvantage Actor-Critic (A2C), to handle the complexity oflarge-scale systems more effectively. DyCITO+ builds on thedynamic dependency analysis mechanism from DyCITO, whichcaptures more accurate runtime relationships, including interfaceimplementation, abstract class inheritance, method overriding,and multilevel inheritance. We evaluated DyCITO+ on eightJava programs of varying sizes. The results show that DyCITO+significantly improves the scalability and effectiveness of CITOgeneration. Among the three DRL methods, A2C consistentlyproduces the lowest overall stubbing complexity, particularly inmedium- and large-scale systems.

The generation of class integration test orders (CITOs) is a pivotal element in integration testing, which focuses on determining the optimal order for integrating classes while testing an object-oriented system. Due to a high number of dependencies and their possible error proneness, some classes are more critical than others in a program. Existing methods for handling these classes only assess risk in terms of their dependencies; they do not consider historical bug information as an additional indicator and mainly work on small programs. To overcome these limitations, this paper introduces Criticality-Driven CITO (CD-CITO) generation, an innovative approach to optimize CITOs by focusing on class criticality. CD-CITO assesses both the importance of a class in terms of its dependencies and the likelihood of defects, based on historical bug data, to determine a criticality score. Then, it reformulates the CITO generation problem as a Reinforcement learning (RL) task and uses the Advantage Actor-Critic (A2C) algorithm to address it. We propose a novel reward calculation strategy to guide the learning agent, balancing stubbing costs with the criticality values of classes to optimize the test order. To extract fault proneness information and assess the approach, the paper uses Defects4J, a data set that contains real bugs and patches of Java programs. The results obtained show that CD-CITO effectively identifies and prioritizes highly critical classes and also minimizes stubbing costs while generating CITOs, which makes it a valuable tool for integration testing.

In this study, the skewed and asymmetric measurement noise is considered and solved in the identification of linear parameter varying (LPV) systems and a new robust global identification framework is established based on the shifted asymmetric Laplace (SAL) measurement distribution. The skewness and tails of the SAL distribution can be adaptively adjusted by the hyperparameters, that means the statistical property of the SAL distribution is governed by the hyperparameters which makes the SAL distribution flexible to resist various types of outliers including the skewed and asymmetric noise. The mathematical formulations of the identification problem are realized by the expectation maximization (EM) algorithm and the maximum likelihood estimates of the parameters are produced. It is realized that both the model parameters and the hyperparameters are extracted directly from the collected identification data. Compared with the existing robust methods, the advantages and disadvantages of the current work are revealed through the designed verification tests performed on the numerical example and the three-tank system, and the main results of this paper are also summarized.Note to Practitioners—The LPV system is widely applied in industrial processes due to its flexible capability of describing the complex nonlinear dynamics. For the probability-based identification of LPV systems, the Gaussian, Laplace and Student’s t distributions are commonly used to describe the output noise. But all of them exhibit symmetric statistical properties which may limit their applications in practical industrial settings, will them keep effective for the skewed and asymmetric measurement noise? Motivated by this question, this paper solves the robust identification of LPV systems with skewed and asymmetric measurement noise and a new robust global identification approach is introduced based on the SAL distribution. In this paper, it is proved that the common Laplace distribution can be seen as a special case of the SAL distribution. That means the proposed method is robust not only for the skewed and asymmetric measurement noise but also for the outliers, which could extend its applications in practical industrial processes. The tests performed on the numerical example and the three-tank system verify the proposed approach.

In this paper, the skewed output noise is considered and we propose a lightweight robust algorithm for nonlinear state-space system identification based on the generalized hyperbolic variance gamma (GHVG) distribution, which enhances the robustness of the proposed algorithm. To facilitate the realization of the proposed algorithm, the hidden variables are introduced to decompose the GHVG distribution into the Gaussian gamma mixture (GGM) distribution, which improves the computational efficiency of the proposed algorithm. The expectation maximization (EM) and the particle smoothing (PS) approaches are combined to solve the hidden variables and unknown states problems, which contributes to derive the estimation formulas of the model parameters and noise parameters simultaneously. To further reduce the computational burden of PS method for estimating the nonlinear states, a novel nearest neighbor idea is used in the identification process which ensures the performance of the proposed algorithm while reducing the number of particles involved in the calculation of the cost function. Finally, the verification results are fairly carried out to demonstrate the effectiveness of the proposed algorithm. Note to Practitioners—The nonlinear state-space model (NSSM) is widely applied in industrial processes due to its ability to characterize the internal dynamics of the system. For probability-based identification of NSSM, the Gaussian or heavy-tailed distributions with symmetric statistical properties are commonly used, which may limit their application in practical industrial settings. Motivated by this problem, this paper proposes a lightweight robust identification algorithm of NSSM based on GHVG distribution. An improved PS method based on nearest neighbor idea is also introduced to reduce the computational efficiency of the algorithm while ensuring its performance. This means that the proposed method is not only robust to the outliers and the skewed output noise, but also greatly reduces the running time of the algorithm, which can extend its application in real industrial processes.

For the output probability density function (PDF) control problem of singularly perturbed systems, a new neural networks-based output PDF shape identification and control is proposed to address the impact of fast and slow time scales. First, an identification scheme for the output probability density function of singularly perturbed systems based on multi-time scale neural networks is proposed, by designing a new weight adaptive update algorithm through the optimal bounded ellipsoid constraint. Second, two control methods are proposed under different uncertainty conditions. By uncertainty approximation assumption, a direct optimization method based on derivation is proposed. Then, to enhance the applicability, a control strategy based on convergence domain and the gradient descent optimization method is proposed to update the control parameters. The closed-loop stability is analyzed by constraining the convergence domain under Lyapunov stability, ensuring the tracking effect of the system state. Simulation results and grinding process confirm the effectiveness of the proposed method.

This article studies the joint design of stealthy actuator and sensor attacks against cyber-physical systems with the aim of keeping the system’s state in an unsafe region. The Kullback-Leibler divergence is adopted as the metric of the joint attacks’ stealthiness. The attacker’s objective is realized by making the system’s ellipsoidal invariant reachable set under stealthy joint attacks belong to the unsafe set. Firstly, the relationship between the actuator attack and the shape of the ellipsoid is analyzed and it can be characterized by a non-convex optimization problem. Parameters of the actuator attack are obtained by solving another convex optimization problem constructed through applying a linear transformation to the original problem. Then, the sensor attack is analytically solved from a non-convex optimization problem to move the center of the ellipsoid to the desired target and increase the controller’s cost. Finally, an example of the flotation industrial process is illustrated to demonstrate effectiveness of the attack. Note to Practitioners—This paper aims to study security of cyber-physical systems from the perspective of attackers, which can help defenders fully understand the behavior of attackers. Existing works have not investigated which kind of stealthy attacks can move the state to the unsafe region. In this article, novel stealthy joint attacks are proposed such that the state of the attacked system is kept in the unsafe region. In detail, the actuator attack is to reshape the system’s ellipsoidal invariant reachable set and the sensor attack is to move the center of the ellipsoid to the desired target. In practical applications, the attacker need to obtain the system’s parameters and eavesdrop the input and output data, then solve the actuator attack from a convex optimization problem and compute the sensor attack by analytically solving a non-convex optimization problem. The attacks’ effectiveness is verified through the flotation industrial process. In the future, we will further investigate the design of stealthy attack strategies for nonlinear systems.

Effective expert selection is an important guarantee for academic evaluation. Recently, graph structure learning (GSL) have been used to model the complex relationships between candidates and experts in both professional fields and avoidance strategies. Most existing GSL models aim to learn the structure of homogeneous graph. However, they cannot learn the structure of heterogeneous graph (HG) because they ignore the complex relation attributes in HG. To fill this gap, we propose a relation enhanced heterogeneous graph structure learning (RE-HGSL) model that refines HG structure and learns node and relation representations simultaneously for expert selection in academic evaluation. The main idea of RE-HGSL is refining the HG structure by reconstructing edges through the similarity of relation triplet$<$head node, relation, tail node$>$. Specifically, we first set relation representation vector for each kind of relation to capture multiple relation attributes and learn node and relation representations to preserve multiple semantic information. Then, we construct multiple feature graphs where the edges are constructed according to the similarity of the relation triplet. Finally, the refined graph structure is generated by aggregating the original graph and the feature graphs, which maintains not only the part of original vital structure but also the refined edges.

In the scenario of class-incremental learning (CIL), deep neural networks have to adapt their model parameters to non-stationary data distributions, e.g., the emergence of new classes over time. To mitigate the catastrophic forgetting phenomenon, typical CIL methods either cumulatively store exemplars of old classes for retraining model parameters from scratch or progressively expand model size as new classes arrive, which, however, compromises their practical value due to little attention paid toparameter efficiency. In this paper, we contribute a novel solution, effective control of the parameters of a well-trained model, by the synergy between two complementary learning subnetworks. Specifically, we integrate one plastic feature extractor and one analytical feed-forward classifier into a unified framework amenable to streaming data. In each CIL session, it achieves non-overwritten parameter updates in a cost-effective manner, neither revisiting old task data nor extending previously learned networks; Instead, it accommodates new tasks by attaching a tiny set of declarative parameters to its backbone, in which only one matrix per task or one vector per class is kept for knowledge retention. Experimental results on a variety of task sequences demonstrate that our method achieves competitive results against state-of-the-art CIL approaches, especially in accuracy gain, knowledge transfer, training efficiency, and task-order robustness. Furthermore, a graceful forgetting implementation on previously learned trivial tasks is empirically investigated to make its non-growing backbone (i.e., a model with limited network capacity) suffice to train on more incoming tasks.

This article, from the attacker’s standpoint, develops a model-free stealthy attack that can steer the system state to the predefined target value and evade detection, without prior knowledge of the system dynamics. A constrained Markov decision process (CMDP) is first modeled to characterize the objective of the stealthy attack. On the basis of the established CMDP, an actor–critic reinforcement learning algorithm is proposed to train the attacker’s policy. Furthermore, by introducing a Lyapunov function constructed from the action value function to the algorithm, convergence of the attacked system’s state to the target is theoretically guaranteed. Differing from existing model-free stealthy attacks which are only suitable for linear systems, the proposed approach guarantees the applicability to nonlinear systems. A linear numerical example and a nonlinear example of flotation industrial system are provided to validate the effectiveness of our proposed stealthy attack.

The increasing frequency of school shootings in the United States has been raised as a critical concern. Active shooters kill innocent students and educators in schools. These incidents highlight the urgent need for effective strategies to minimize casualties. This study aims to address the challenge of simulating and assessing potential mitigation measures by developing a multi-agent simulation model. Our model is designed to estimate casualty rates and evacuation efficiency during active shooter scenarios within school buildings. The simulation evaluates the impact of a gun detection system on safety outcomes. By simulating school shooting incidents with and without this system, we observe a significant improvement in evacuation rates, which increased from 16.6% to 66.6%. Furthermore, the Gun Detection System reduced the average casualty rate from 24.0% to 12.2% within a period of six minutes, based on a simulated environment with 100 students. We conducted a total of 48 simulations across three different floor layouts, varying the number of students and time intervals to assess the system’s adaptability. We anticipate that the research will provide a starting point for demonstrating that a gunshot detection system can significantly improve both evacuation rates and casualty reduction.

Class-incremental learning (CIL) aims to train a model to learn new classes from non-stationary data streams without forgetting old ones. In this paper, we propose a new kind of connectionist model by tailoring neural unit dynamics that adapt the behavior of neural networks for CIL. In each training session, it introduces a supervisory mechanism to guide network expansion whose growth size is compactly commensurate with the intrinsic complexity of a newly arriving task. This constructs a near-minimal network while allowing the model to expand its capacity when cannot sufficiently hold new classes. At inference time, it automatically reactivates the required neural units to retrieve knowledge and leaves the remaining inactivated to prevent interference. We name our model AutoActivator, which is effective and scalable. To gain insights into the neural unit dynamics, we theoretically analyze the model’s convergence property via a universal approximation theorem on learning sequential mappings, which is under-explored in the CIL community. Experiments show that our method achieves strong CIL performance in rehearsal-free and minimal-expansion settings with different backbones.

Accurate estimation of the Lithium-ion batteries (LiBs) state of health (SOH) is essential to ensure the safe and reliable operation of battery-powered devices. Most of the current data-driven SOH estimation models are designed under fixed ambient temperatures, overlooking the high sensitivity of LiBs to changing ambient temperatures. To bridge this gap, a novel method is proposed with transfer learning (TL) to model and estimate the SOH at varying ambient temperatures. First, canonical variate analysis is employed to capture the temporal dynamics in the time-series data of LiBs and extract temporal features. Second, at the reference temperature, an interpretable SOH estimation model with a long short-term memory network is utilized to learn the regression relationship between the temporal features and the labeled capacities. Third, at the new temperature, a small amount of data is projected into the key feature domain at the reference temperature to obtain the common temporal features. Afterward, TL is employed to take advantage of the existing process knowledge through keeping all functional layers of the previous well-trained SOH estimation model. Finally, the efficacy of the proposed method is verified on the NASA dataset with two discrete temperatures, 24°C and 44°C. Evaluated by the index of root mean squared error, the proposed method is capable of improving prediction accuracy by 94.18% when the model is transferred from 44°C to 24°C.

Industrial streaming data exhibits the concept drift characteristic due to the time-variant operating conditions, which degrades the performance of models established by traditional offline batch learning. Moreover, the widespread outliers and the correlations between data variables in industrial data streams can have a devastating impact on modeling. Therefore, this paper presents a novel online sequential sparse robust neural networks with random weights (OSSR-NNRW) for imperfect industrial streaming data to achieve highly reliable online modeling of time-variant dynamic systems. First, sparse partial least squares regression is used to replace least squares estimation for network output weights calculation, which not only can effectively solve the multicollinearity problem caused by correlations, but also enable variable selection to improve the performance and interpretability. Second, we introduce the online sequential learning strategy with forgetting factor to realize adaptive updating of model parameters, thus enhancing the online learning ability and overcoming the time-variant dynamics of industrial systems. More importantly, in order to strengthen the robustness of the model, Schweppe generalized M-estimation is adopted to determine the modeling weights by the model residual size and the distance information of input vectors in the high-dimensional space to resolve the prevalent existence of outliers in the input and output samples. Finally, data experiments on two industrial systems have validated the effectiveness, advancement, and practicality of the proposed method. Note to Practitioners—In the process industry, the product quality relies on the timely and accurate measurement of key production indicators. However, owing to the time-varying characteristics of industrial processes and the limitations of measurement devices, conventional batch learning-based data-driven models are difficult to apply to imperfect industrial data stream scenarios. To this end, the OSSR-NNRW is proposed for online robust modeling of complex time-variant dynamic systems by combining sparse robust modeling and online learning strategy in a unified framework of neural networks with random weights. The OSSR-NNRW enables online learning based on industrial data streams while resolving correlations between data variables and mitigating the negative effects of outliers from both input and output samples on the modeling process. Experimental results using two typical process industry datasets show that the proposed OSSR-NNRW has high estimation accuracy and can be easy to implement in industrial processes.

In this article, we aim to offer an interpretable learning paradigm for incremental random weight neural networks (IRWNNs). IRWNNs have become a hot research direction of neural network algorithms due to their ease of deployment and fast learning speed. However, existing IRWNNs have difficulty explaining how hidden nodes (parameters) affect the convergence of network residuals. To address this gap, this article proposes an interpretable construction algorithm (ICA). Specifically, we first conduct a spatial geometric analysis of the network construction process and establish the spatial geometric relationship between the network residuals and hidden parameters to visualize the influence of hidden parameters on the convergence of the network residuals. Second, based on the spatial geometric relationship and node pool strategy, an interpretable control strategy with spatial geometry information is established to obtain hidden parameters conducive to the convergence of network residuals. In addition, to facilitate ICA to handle complex tasks of big data, this article proposes a lightweight ICA with low complexity, namely ICA+. Finally, it is proved theoretically that the ICA and ICA+ proposed in this article have universal approximation properties. The experimental results on two real-world datasets and seven benchmark datasets demonstrate the advantages of the proposed ICA and ICA+ in terms of fast learning, good generalization, and compactness of network structure.

Incremental random weight networks (IRWNs) are increasingly being employed for modeling industrial processes. However, since most of the existing IRWNs do not consider the relationships between multiple targets, they have limited capability of conducting the task of multitarget modeling encountered in many industrial processes. In this article, a novel multitarget IRWN integrating$L_{2,1}$-norm regularization, termed as MTIRWN, is proposed to learn potential intertarget correlations for improved modeling performance. First, the feature layer is inserted between the input layer and hidden layer, based on this, the input data are fed into a more proper feature space using the feature mapping approach. Then, to benefit from the relationships among targets, an objective function using$L_{2,1}$-norm regularization is used, and a new inequality constraint that can explore the influence relationships between the newly added hidden node and generalization performance is constructed by the Greville's method. Then, alternating direction method of multiplier (ADMM) is employed to optimize the output weights. Finally, an extensive evaluation is untaken on six real-word datasets and a real multitarget modeling problem, which indicates that the proposed algorithm can effectively solve the multitarget modeling problems.

This article presents a case study on predicting the particle size of copper ore grinding with stochastic configuration networks (SCNs). A set of temporal data was collected from the copper ore grinding process for building the predictive model. The well-known least absolute shrinkage and selection operator (LASSO) is employed for temporal feature selection. Then, an SCN model is constructed to estimate the residual between the output of the LASSO model and the product particle size. The weights of the linear model and the output weights of the SCN model are optimized simultaneously using the elastic-net regularization method to improve the prediction performance. An analysis on the selected input variables is given, which is helpful to better understanding the output results. Comparative experiments are carried out to demonstrate the effectiveness and merits of the proposed modeling techniques in terms of the model's accuracy, reliability, and interpretability.

Incremental random weight networks (IRWNs) face the issues of weak generalization and complicated network structure. There is an important reason: the learning parameters of IRWNs are determined in a random fashion without guidance, which may increase many redundant hidden nodes, and thereby leading to inferior performance. To resolve this issue, a novel IRWN with compact constraint that guides the assignment of random learning parameters (CCIRWN) is developed in this brief. Using the iteration method of Greville, a compact constraint that simultaneously assures the quality of generated hidden nodes and the convergence of the CCIRWN is built to perform learning parameter configuration. Meanwhile, the output weights of the CCIRWN are assessed analytically. Two types of learning methods for constructing the CCIRWN are proposed. Finally, the performance evaluation of the proposed CCIRWN is undertaken on the 1-D nonlinear function approximation, several real-world datasets, and data-driven estimation based on the industrial data. Numerical and industrial examples indicate that the proposed CCIRWN with compact structure can achieve favorable generalization ability.

The organic Rankine cycle (ORC) is an effective application for converting low-grade heat sources into power and is crucial for environmentally friendly production and energy recovery. However, the inherent complexity of the mechanism, its strong and unidentified nonlinearity, and the presence of control constraints severely impair the design of its optimal controller. To solve these issues, this study provides a novel event-triggered (ET) constrained optimal control approach for the ORC systems based on a safe reinforcement learning technique to find the optimal control law. Instead of employing the usual non-quadratic integral form to solve the control-limited optimal control problems, a constraint handling strategy based on a relaxed weighted barrier function (BF) technique is proposed. By adding the BF terms to the original value function, a modified value iteration algorithm is developed to make the control input solutions that tend to violate the constraints be pushed back and maintained in their safe sets. In addition, the ET mechanism proposed in this article is critically required for the ORC systems, and it can significantly reduce the computational load. The combination of these two techniques allows the ORC systems to achieve set-point tracking control and satisfy the control restrictions. The proposed approach is conducted based on a heuristic dynamic programming framework with three neural networks (NNs) involved. The safety and convergence of the proposed approach and the stability of the closed-loop system are analyzed. Simulation results and comparisons are presented to demonstrate its effectiveness.

It is well common for industrial processes to employ a hierarchical control structure involving a basic loop process and an operation loop process with two timescales. However, the control system suffers from another multirate challenge where control and sampling rates may differ even within a single loop. Additionally, the underlying complex mechanism of the operation loop further complicates the accurate modeling of its dynamics, especially in the presence of external unknown disturbances. This gives rise to the difficulty in obtaining desired control performance. To overcome these problems, this article develops a novel operational optimal tracking control method for a class of multirate systems subject to unknown disturbances. To this end, a lifting technique is integrated with a general model predictive controller for the basic loop process, aimed at handling the asynchronism phenomenon and achieving loop setpoint tracking control. Furthermore, a nonlinear disturbance observer is used for estimating the unknown external disturbance of the operation loop process. In this way, offset-free tracking control of the system, along with loop setpoints optimization, can be achieved using the policy iteration reinforcement learning algorithm. The convergence of the proposed method is analyzed and tangible improvements are verified by simulations.

Detecting insulators and their defects is a key task in real-time power grid surveillance with the rapid development of national smart grid. Traditional surveillance usually relies on maintenance personnel, leading to the issues of inefficiency and unsafety. Thus, with the prosperity of deep learning, we proposed a detection algorithm, named E-VarifocalNet, which is an enhanced version of the basic VarifocalNet method. The proposed E-VarifocalNet is specifically designed for detecting insulators and their defects. We developed a classification loss based on varifocal loss and the number of samples to solve the imbalance problem in object detection. Furthermore, a regression loss based on GIoU loss and Wasserstein distance is designed to gain higher flexibility in the representation of bounding boxes. Additionally, we applied a feature pyramid network based on dilated convolution and heatmap to build global and local semantic relations among pixels so as to enhance the detection accuracy on salient areas. Our dataset containing 2,100 images and 5,217 object instances was collected through real-time drones and an open data platform. Our E-VarifocalNet gets the highest mAP and a low model complexity on our dataset among state-of the-art object detectors, indicating the potential of our algorithm in real-time power grid surveillance applications.

In this article, the$H_{\infty }$control problem is investigated for a class of two-time-scale cyber-physical systems (TTSCPSs). In order to reduce the network bandwidth occupation and lighten the computation burden, an asynchronous dynamic event-triggered protocol (ADETP) is designed to arrange the signal transmissions from sensors to the composite controllers, in which the triggering sequences of fast and slow system components are determined separately. A novel composite controller based on the proposed ADETP is designed dependent on the singular perturbation parameter (SPP) such that the closed-loop TTSCPS is asymptotically stable with meeting a required$H_{\infty }$performance index when the SPP is no more than a given upper bound. Gain matrices of the desired composite controller are parameterized in terms of the solutions to certain matrix inequalities that are readily solvable. Finally, simulation results of a nuclear reactor are presented to verify the effectiveness of the proposed approach.

The subject of attack detection for industrial cyber-physical systems (IPCSs) is covered in this paper, which addresses threats from transient covert attacks (TCAs), also referred to as the second version of replay attacks with a specific frequency and short duration. A comprehensive model of the TCAs is built using the active instant and period of the attacks, as well as the dynamics of a virtual system to replicate IPCS function and produce attack signals. Though watermarking-based detection algorithms have been shown to be effective in detecting TCAs, the induced system performance loss is too significant, and as such, our primary goal is to minimize system performance degradation while maintaining the detection rate. Because the active periods of TCAs are substantially shorter than their sleep ones, or even because they are practically always silent, it makes sense that reducing superfluous watermarking will facilitate system performance. So, using an “event-triggered” strategy, a unique recursive watermarking-based detection algorithm is proposed. Here, the trigger modes of watermarking are divided into three types: forced, high probability, and low probability. The design principles are proven via algorithms and criteria, and a theoretical analysis of the detection rate and the system performance loss is also supplied. The advantages of the suggested algorithms are finally demonstrated by numerical simulations of a quadruple-water-tank system and experiments with a permanent magnet synchronous motor on the dSpace platform.

This article addresses the replay attack detection issue for a class of industrial process operation optimization (IPOO) cyber-physical systems (CPSs). In contrast to conventional CPSs with a single-loop structure, the IPOO CPSs employ a dual-layer network environment consisting of a wireless network for the setpoint optimization loop and a controller area network (CAN) bus for the device control loop. Consequently, the issue of attack detection is more complicated for the IPOO CPSs, and this motivates our current research. First, a unified model of the IPOO CPS is built, with PI controller controlling the physical plants and linear quadratic Gaussian (LQG) controller managing setpoint optimization. Then, a novel joint watermarking detection mechanism is established with the PI controller watermarking, LQG controller watermarking, and a watermarking compensator. The proposed watermarking compensator with an augmented Kalman filter is utilized to efficiently eliminate false alarms brought on by information interactions of the coupled cyber-layers, enabling the accurate detection and location of replay attacks. Furthermore, a linear relationship is established between watermarking parameters and system performance loss. Finally, simulations with a quadruple water tank system are conducted to verify the effectiveness of the proposed algorithm.

We consider the robust state estimation over sensor networks with non-Gaussian noise, which is often encountered in many applications. Motivated by the fact that the elliptical distribution is the natural extension of the Gaussian distribution and includes a variety of non-Gaussian distributions with heavy-tailed characteristics, we here adopt the elliptical distribution to model the heavy-tailed process and measurement noise. The general state evolution model is used to replace the process equation and the elliptical distribution is denoted as a Gaussian mixture form. Based on that, the posterior estimation of the system state together with the parameters of the process and measurement noises can be inferred by the variational Bayes method. Moreover, the corresponding distributed estimation algorithm is then provided, which enables distributed implementation. The target tracking over sensor networks is used to show the estimation accuracy of the proposed algorithms.

Dynamic memristor (DM)-cellular neural networks (CNNs), which replace a linear resistor with flux-controlled memristor in the architecture of each cell of traditional CNNs, have attracted researchers’ attention. Compared with common neural networks, the DM-CNNs have an outstanding merit: when a steady state is reached, all voltages, currents, and power consumption of DM-CNNs disappeared, in the meantime, the memristor can store the computation results by serving as nonvolatile memories. The previous study on stability of DM-CNNs rarely considered time delay, while delay is quite common and highly impacts the stability of the system. Thus, taking the time delay effect into consideration, we extend the original system to DM-D(delay)CNNs model. By using the Lyapunov method and the matrix theory, some new sufficient conditions for the global asymptotic stability and global exponential stability with a known convergence rate of DM-DCNNs are obtained. These criteria generalized some known conclusions and are easily verified. Moreover, we find DM-DCNNs have$3^{n}$equilibrium points (EPs) and$2^{n}$of them are locally asymptotically stable. These results are obtained via a given constitutive relation of memristor and the appropriate division of state space. Combine with these theoretical results, the applications of DM-DCNNs can be extended to other fields, such as associative memory, and its advantage can be used in a better way. Finally, numerical simulations are offered to illustrate the effectiveness of our theoretical results.

In large-scale process industries, one of the great challenges is to achieve optimum operation of systems with multi-time-scale property and partially unknown models. To this end, this article proposes a novel multi-rate layered operational optimal control (OOC) method, which employs lifting technique to unify the relatively fast dual-rate of basic loop layer and relatively slow single-rate of operational layer. Besides, by integrating model-based predictive control of basic loop layer with data-based actor-critic reinforcement learning (RL) of operational layer, it overcomes the difficulty of building the operational process dynamic model. The convergence of the proposed method is proved, and dense medium separation (DMS) process is taken as an application case to illustrate the effectiveness of our proposed method via a self-developed simulation platform.

As a class of randomized learner model, stochastic configuration networks (SCNs) have been successfully applied in a few data analytics tasks. Given the industrial big data modeling tasks, however, the original SCNs potentially lead to excessive training time. To this end, this article extends SCNs to a hybrid parallel version, termed hybrid parallel SCNs (HPSCNs). In the hybrid parallel learning algorithm, two SCNs are synchronously constructed. The difference between the two of them is that one employs a point-incremental algorithm, and another one adopts a block-incremental algorithm. Moreover, a data parallel method is established based on a dynamic block strategy to accelerate the establishment of candidate node pool for each SCN. Additionally, this article proposes an adaptive hyperparameter adjustment method, which allows the hyperparameters in the supervisory mechanism to be automatically adjusted along with the learning process. Comparative experiments are first conducted through four large-scale benchmark datasets, followed by the fully discussion on the algorithm parameters. Finally, a practical industrial application case is made, where a grinding particle size soft sensor is developed based on HPSCN, showing the effectiveness of the proposed algorithm.

Optimal operation of an actual mine main fan switchover process relies heavily on a good measurement of underground airflow quantity (UAQ). However, real-time measuring the UAQ is difficult using conventional measurement techniques. In this article, a novel randomized learning model, named compact incremental random weight network (CIRWN), is proposed to estimate the UAQ. Since the hidden parameters of the original IRWN are generated in a fixed scope with a random manner, which is prone to create redundant hidden nodes, a CIRWN with new inequality constraints is proposed. The inequality constraints have several attractive properties, including dynamically guiding the generation of hidden parameters, effectively enhancing the convergence rate, and successfully establishing a universal approximator. Experiments using four benchmark datasets and an industrial dataset show that the established model possesses a more compact network structure and better modeling accuracy as well as faster convergence rate compared with other methods.

This article investigates the optimal operational control (OOC) problem for a class of industrial systems consisting of multiple unit devices with fast dynamics and an unknown operational process with slow dynamics. First, the OOC problem is formulated as a noncascade optimal control problem of two-time-scale systems with a novel performance function. Second, using singular perturbation theory, a decentralized composite control scheme is proposed by decomposing the original optimal problem into reduced-order fast and slow subsystem problems. Then, in the framework of reinforcement learning, an online controller design method for the slow subsystem is proposed by using the online measurement, and an offline controller design for the fast subsystem is proposed by using the unit device models. The obtained decentralized composite optimal controller achieves both the desired operational index tracking and disturbance rejection without requiring the dynamics of the operational process. Different from the existing cascade design methods, the proposed approach regulates the unit devices and operational process simultaneously, as well as overcomes the potential high dimensionality and ill-conditioned numerical issues. Finally, a mixed separation thickening process and a numerical example are given to illustrate the presented results.

Dense medium separation (DMS) is of great significance for coal cleaning. The DMS control system always involves dense medium density adjustment and ash content control that are operating on fast and slow time scales, respectively. The inherent time-varying and strongly nonlinear characteristics of the DMS process give rise to challenges for the design of this multitime scale control system. To address this issue, this article proposes a dual-rate adaptive optimal tracking control approach for the DMS system. For the basic loop process, a nonlinear adaptive PI controller containing a neural network (NN)-based unmodeled dynamics compensator is proposed. Then, a lifting technique is used to unify the time scales of the two loops accompanied by formulating a generalized controlled object, whose dynamics is completely unknown. On this basis, a data-driven operation optimization control method that combines adaptive dynamic programming algorithm and reference control is developed, which is implemented using NNs. Finally, the stability of the proposed method is analyzed. The simulation results indicate its effectiveness.

This paper introduces the development and implementation of a ball mill grinding circuit simulator, NEUSimMill. Compared to the existing simulators in this field which focus on process flowsheeting, NEUSimMill is designed to be used for the test and verification of grinding process control system including advanced control system such as integrated control. The simulator implements the dynamic ball mill grinding model which formulates the dynamic responses of the process variables and the product particle size distribution to disturbances and control behaviors as well. First principles models have been used in conjunction with heuristic inference tools such as fuzzy logic and artificial neural networks: giving rise to a hybrid intelligent model which is valid across a large operating range. The model building in the simulator adopts a novel modular-based approach which is made possible by the dynamic sequential solving approach. The simulator can be initiated with connection to a real controller to track the plant state and display in real-time the effect of various changes on the simulated plant. The simulation model and its implementation is verified and validated through a case of application to the design, development, and deployment of optimal setting control system.