Accurate decision-making for complex industrial processes is essential for improving resource utilization and optimizing operating conditions. However, current decision work-flows operate within a hierarchical closed loop between the operations layer and the process-control layer. Cross-layer transmission of objectives and constraints and feedback coupling can degrade performance; the effect is pronounced under nonstationary conditions or when objectives or constraints change. To address these challenges, we propose the Bi-level Evolving Neural Networks (BENNs) framework with four components: (i) a bi-level neural architecture that explicitly models cross-layer information flow and constraint propagation; (ii) a graph neural network-based structural similarity assessment that reduces redundant evaluations to improve search quality; (iii) an online evolving network strategy that adapts rapidly to process dynamics without extensive retraining; and (iv) a convex-hull knee-point-based compromise selection that encodes engineering preferences. Experiments on real mineral processing data and on the Tennessee Eastman Process (TEP) show improvements over baselines in prediction accuracy, performance retention, and adaptation speed, demonstrating the potential of BENNs for complex industrial settings.

Multirate industrial processes pose significant challenges for accurate forecasting due to varying sampling frequencies and missing data. This article proposes a novel hybrid deep learning framework that effectively addresses these issues. Our approach uses a combination of time series decomposition, inverted transformer (iTransformer)-based feature extraction, and a modified minimal gated unit (MGU) network. To handle missing quality variables, we introduce a robust adaptive parameter update algorithm based on dead-zone Kalman filtering. Through extensive experiments conducted on real-world industrial datasets, our method achieves a mean absolute error (MAE) reduction of 61.42%, a root-mean-square error (RMSE) reduction of 64.11%, and a high qualification rate improvement of 14.73% compared to the average performance of state-of-the-art technologies, thereby outperforming existing state-of-the-art techniques in terms of both forecasting accuracy and robustness.

This paper introduces a novel transfer learning framework for time series forecasting that uses Concept Echo State Network (CESN) and a multi-objective optimization strategy. Our approach addresses the challenges of feature extraction and knowledge transfer in heterogeneous data environments. By optimizing CESN for each data source, we extract targeted features that capture the unique characteristics of individual datasets. Additionally, our multi-network architecture enables effective knowledge sharing among different ESNs, leading to improved forecasting performance. To further enhance efficiency, CESN reduces the need for extensive hyperparameter tuning by focusing on optimizing only the concept matrix and output weights. Our proposed framework offers a promising solution for forecasting problems where data is diverse, limited, or missing.

As one of the by-products of the municipal solid waste incineration (MSWI) process, dioxin (DXN) is not only difficult to detect but also potential harm to humans and the environment. The article proposes a method for detecting DXN emissions. It addresses the challenge of poor generalization performance in detection models due to the dynamic nature of the MSWI process. Firstly, the method constructs a historical soft sensor model and a drift detection model based on historical samples. Secondly, it assesses online samples for drift detection. When drift is detected, it calculates a distance threshold to prune the ensemble model. Subsequently, it reconstructs new ensemble sub-models and integrates them with the historical model to form a preliminary online ensemble model. Finally, it conducts local pruning and reconstruction on each sub-model, refining the final online ensemble model based on weighted posterior information. The efficacy of this approach is validated using synthetic, benchmark, and real DXN datasets from an MSWI plant in Beijing.

The flue gas oxygen content is critical for ensuring the efficient and stable operation of the municipal solid waste incineration (MSWI) process. This article proposes a novel control method based on an interval type-2 fuzzy broad learning system and proportional-integral-derivative (IT2FBLS-PID). First, a parameter self-learning complex controller based on IT2FBLS is designed, alongside a simple adaptive PID controller. To reduce computational complexity and resource consumption while maintaining control accuracy, a double event-triggering (DET) mechanism is introduced. Additionally, to enhance the controller’s adaptability and robustness, structural self-organization adjustments are made to the IT2FBLS. Finally, a stability analysis of the parameter learning and structural self-organization algorithms of the complex controller is conducted. The superiority of the proposed control method is demonstrated through actual data from a MSWI power plant in Beijing.

The current control of alumina digestion process (ALDP) relies on infrequent manual assays for caustic ratio, leading to open-loop operation between assays and suboptimal performance. This paper proposes a novel cloud-edge collaboration (CEC) architecture utilizing deep learning for real-time caustic ratio prediction, enabling closed-loop operation at all times. Our method combines slow-sampled assays with fast-sampled process measurements. An autoregressive moving average with extra inputs (ARMAX) model is used for feature extraction, while an improved bidirectional gated recurrent unit (I-BiGRU) that incorporates a multi-head attention (MHA) mechanism bridges data gaps through prediction. A soft-collaboration mechanism ensures smooth model updates at the edge, enabling operators to make informed control decisions. Theoretical analysis guarantees the convergence of prediction errors. Real-world data from a large-scale alumina plant demonstrates the superior performance of our prediction method compared to the baseline methods. Additionally, industrial experiments validate the effectiveness of the CEC architecture in supporting human operators for ALDP closed-loop operation. Note to Practitioners—This paper was motivated by the challenges of achieving closed-loop operation in the alumina digestion process. The concept of cloud-edge collaboration (CEC) architecture with deep learning offers a powerful solution for various industrial scenarios that rely on infrequent manual measurements, where real-time monitoring of key performance indexes is essential. CEC can continuously predict these values, bridging the gaps between slow, accurate assays and frequent, less precise sensor data. This approach allows for closed-loop operation based on real-time predictions, potentially improving production efficiency and yield. The ability to handle data gaps makes our method applicable to processes with inherent measurement challenges. By implementing CEC in various industries, practitioners can transition from open-loop control to data-driven closed-loop systems, optimizing processes and achieving substantial benefits.

Precise online prediction of caustic concentration is crucial for optimal control and operating efficiency in alumina production. However, existing supervised learning models rely heavily on a large amounts of labeled assay samples, which are time-consuming and resource intensive to obtain. To overcome this challenge, we propose a novel semisupervised learning approach that effectively uses both labeled and unlabeled data. Our semisupervised approach, based on a transformer variational autoencoder (TVAE), incorporates a multihead decoder for joint reconstruction and prediction tasks. Through robust regularizing and jointly learning these tasks, the TVAE constructs a smooth, low-dimensional, and continuous latent representation that captures hidden patterns and relationships within the data, enabling accurate predictions with significantly fewer labeled samples compared to existing methods. Experiments on a real-world alumina production process demonstrate the superior performance of our approach in predicting caustic concentration, outperforming state-of-the-art semisupervised time-series prediction models.

The furnace temperature (FT) control is the key for ensuring the stable operation and effective pollution reduction in municipal solid waste incineration (MSWI) processes. However, conventional control strategies encounter challenges in effectively managing FT due to uncertainties associated with material composition, feeding modes, and equipment maintenance. In response to these challenges, this article introduces a control approach utilizing a Bayesian optimization-based interval type-2 fuzzy neural network (BO-IT2FNN), which achieves offline optimization and online control through the FT controller constructed by IT2FNN. In offline optimization process, the BO algorithm is used to optimize the learning rate of multiple types parameter of IT2FNN controller. In the online control process, fine-tuned by gradient descent method with multiple LR for adaptability. In addition, the stability of control system is confirmed using theorem of Lyapunov, providing the theoretical foundation. Experiments with real MSWI data, tested on a hardware-in-loop platform, prove the effectiveness of the proposed method.

Deep learning offers promising solutions for quality prediction in industrial processes, improving decision-making and performance monitoring. In this article, we propose a novel deep learning framework that incorporates variable time delay (VTD) estimation and uncertainty quantification into quality prediction. The framework employs a collaborative method that integrates deep Q-network with random forest to estimate VTD values. It then utilizes a hybrid BMCR model, consisting of parallel bidirectional minimal gated unit and 1-D convolutional layers, along with a residual connection, specifically designed to capture both long-term and short-term features in industrial data. The framework produces prediction intervals directly to quantify the uncertainty in the prediction results. This combined method offers high-precision point predictions alongside uncertainty quantification, providing valuable insights for industrial decision-making. The effectiveness of the proposed method is validated through two numerical examples, a benchmark, and a real-world industrial case from the alumina digestion process.

Consistent product quality and efficient operations in high-pressure grinding roll (HPGR) rely heavily on real-time anomaly detection. Complexities arise from fluctuations in raw materials, feeding processes, and unforeseen disruptions, along with the inherent spatio-temporal dynamics of sensor data. This article addresses these challenges by proposing a collaborative anomaly monitoring architecture that leverages the cloud, edge devices, and a powerful algorithm: The spatio-temporal attention-based (STA) minimal gated unit (MGU) adversarial autoencoder (AAE). The proposed algorithm, trained in the cloud, analyzes sensor data encompassing information, material, and energy flows within the HPGR. Its core strength lies in capturing the intricate interplay between spatial and temporal data patterns through a novel spatio-temporal attention mechanism. In addition, adversarial training enhances the model's ability to distinguish normal operations from anomalies. Edge devices perform real-time monitoring and transmit preprocessed data to the cloud for STA–MGU–AAE analysis. The extracted features not only enable accurate anomaly detection in process variables, but also facilitate root cause analysis, leading to significant improvements in process stability and reliability. The effectiveness of the proposed architecture is validated through practical beneficiation experiments, demonstrating its potential to revolutionize HPGR anomaly monitoring in production processes.

In mineral processing, the dynamic nature of industrial data poses challenges for decision-makers in accurately assessing current production statuses. To enhance the decision-making process, it is crucial to predict comprehensive production indices (CPIs), which are influenced by both human operators and industrial processes, and demonstrate a strong dual-scale property. To improve the accuracy of CPIs' prediction, we introduce the high-frequency (HF) unit and low-frequency (LF) unit within our proposed dual-scale deep learning (DL) network. This architecture enables the exploration of nonlinear dynamic mapping in dual-scale industrial data. By integrating the Cloud-Edge collaboration mechanism with DL, our training strategy mitigates the dominance of HF data and guides networks to prioritize different frequency information. Through self-tuning training via Cloud-Edge collaboration, the optimal model structure and parameters on the cloud server are adjusted, with the edge model self-updating accordingly. Validated through online industrial experiments, our method significantly enhances CPIs' prediction accuracy compared to the baseline approaches.

This paper presents a novel approach for predicting bi-level human decision-making in mineral processing. This method aims to enhance the adaptability of human decision-making support systems and reduce dependence on expert knowledge. To achieve this, it leverages a combination of evolutionary computation and neural networks. The proposed method utilizes a hierarchical neural network framework optimized through evolutionary computation. This framework takes advantage of the sequential nature of decision-making within the mineral processing system. The effectiveness of the method is validated through experiments using 1,324 real-world data entries from a mineral processing plant in China. The results demonstrate the method’s capability to improve prediction accuracy.

This paper presents a novel and cost-effective approach for structural health monitoring using smartphones. By using built-in accelerometers, smartphones can collect data on building motion, facilitating the detection of potential damage. Traditional methods often rely on classification techniques, requiring extensive training data encompassing both damaged and undamaged scenarios. However, this proves impractical for smartphones due to their limited computational resources for complex classification tasks. We propose a paradigm shift, transforming the classification problem into a regression problem. This enables robust structural health assessment using a neural network specifically designed for this purpose: the echo state network (ESN). ESNs offer inherent robustness to noise and perturbations, making them ideal for real-world applications with sensor data. Compared to traditional methods, the proposed smartphone-based system offers significant advantages in terms of cost-effectiveness, user-friendliness, and computational efficiency. The effectiveness of the proposed method is evaluated through several experiments, demonstrating its capability in identifying structural damage.

The slurry concentration is a crucial factor in mineral processing production, affected by both upstream and downstream systems, including closed-loop control systems. The variability in slurry concentration presents a challenge due to its complex, nonlinear nature and the difficulty in accurately modeling this dynamic system for real-time monitoring. This article introduces an innovative approach for the real-time prediction of slurry concentration. Our method comprises two key components: a mechanistic model and a nonlinear dynamic system. The mechanism model is developed using the closed-loop input error method, effectively mitigating the influence of control systems. We have rigorously demonstrated the convergence of parameters and the stability of the identification process. Further, a hybrid deep neural network, called convolutional neural network-gated recurrent unit (GRU), is proposed to tackle challenges, such as intervariable dependencies and the inherent nonlinear dynamics of the process. The network integrates an autoregressive integrated moving average (ARIMA) model for laboratory delays. A parallel GRU network captures nonlinear dynamic characteristics after the ARIMA model. This architecture ensures real-time online concentration forecasting. The effectiveness of our approach has been substantiated with actual production data from a large-scale mineral processing facility.

Dioxin (DXN) is a persistent organic pollutant produced from municipal solid waste incineration (MSWI) processes. It is a crucial environmental indicator to minimize emission concentration by using optimization control, but it is difficult to monitor in real time. Aiming at online soft-sensing of DXN emission, a novel fuzzy tree broad learning system (FTBLS) is proposed, which includes offline training and online measurement. In the offline training part, weightedk-means is presented to construct a typical sample pool for reduced learning costs of offline and online phases. Moreover, the novel FTBLS, which contains a feature mapping layer, enhance layer, and increment layer, by replacing the fuzzy decision tree with neurons applied to construct the offline model. In the online measurement part, recursive principal component analysis is used to monitor the time-varying characteristic of the MSWI process. To measure DXN emission, offline FTBLS is reused for normal samples; for drift samples, fast incremental learning is used for online updates. A DXN data from the actual MSWI process is employed to prove the usefulness of FTBLS, where the RMSE of training and testing data are 0.0099 and 0.0216, respectively. This result shows that FTBLS can effectively realize DXN online prediction.

Broad learning system based on neural network (BLS-NN) has poor efficiency for small data modeling with various dimensions. Tree-based BLS (TBLS) is designed for small data modeling by introducing nondifferentiable modules and an ensemble strategy to the traditional broad learning system (BLS). TBLS replaces the neurons of BLS with the tree modules to map the input data. Moreover, we present three new TBLS variant methods and their incremental learning implementations, which are motivated by deep, broad, and ensemble learning. Their major distinction is reflected in the incremental learning strategies based on: 1) mean square error (mse); 2) pseudo-inverse; and 3) pseudo-inverse theory and stack representation. Therefore, this study further explores the domain of BLS based on the nondifferentiable modules. The simulations are compared with some state-of-the-art (SOTA) BLS-NN and tree methods under high-, medium-, and low-dimensional benchmark datasets. Results show that the proposed method outperforms the BLS-NN, and the modeling accuracy is remarkably improved with the small training data of the proposed TBLS.

For most mechanical systems such as buildings, aerodynamics, aerospace applications, etc., acceleration measurement is the only variable for practical purposes. Therefore, the displacement and velocity signals required for control schemes are generally estimated through state observers or by means of integral filters. For instance, the Luenberger observer is designed based on a suitable observer gain, whereas the filters request an appropriate cut-off frequency to avoid phase shifts. Consequently, this article provides an algebraic state observer to estimate displacement, integral of displacement, and velocity simultaneously. The main advantage of this approach is that the observer does need tuning gains, only depending on the system properties. In addition, this work presents an alternative for the semi-active vibration control problem via a PID plus Active Disturbance Rejection Controller (ADRC), which is robust to seismic movement. A nonlinear second-order equation models the building structure equipped with a Magnetorheological Damper (MRD).

PD (proportional-derivative) control is a widely used model-free method for controlling robots. However, it does not guarantee optimal performance. Model-based optimal control methods, such as the linear quadratic regulator (LQR), can achieve the desired control performance, but they are only suitable for linear systems that are well understood. In this paper, we propose a novel approach to design an optimal PD control for unknown robot systems using Conditional Adversarial Generative Networks (C-GAN) and long-short term memory (LSTM) to approximate LQR PD control. This new control mechanism ensures both stability and optimal performance. We apply this method to control lower limb prostheses and our results demonstrate that the optimal PD control using GAN and LSTM outperforms classical controllers.

Fuzzy decision trees (FDTs) is one of the considerably excellent methods. Most of the existing FDTs’ methods are oriented to classification tasks. Applying FDTs to regression tasks may solve complex industrial modeling problems. In this article, we propose the Takagi–Sugeno (T–S) fuzzy regression tree (TSFRT), which uses the hypothesis of “feature screening followed by T–S fuzzy reasoning.” In the TSFRT, the growth process (crisp set theory) can be deemed as feature screening, and each leaf node (fuzzy set theory) is viewed as a T–S inference reasoning system. Thus, the TSFRT becomes a top-down structure. We develop multiple strategies to identify the parameters of the T–S system in the leaf node using sample-by-sample and batch samples. To improve the method's generalization performance, we also generalize an ensemble method with pseudoinverse and ridge regression. The proposed methods are evaluated by several high- and low-dimensional complex industrial processes. The experimental results show that the proposed method remarkably outperforms other popular regression methods.

This study is summarising the design of an output hybrid feedback controller for biped robots where the motion range for each joint is being considered. The design considers the hybrid nature of biped device when it is developing an entire gait cycle including the interaction with the environment touching the support floor. A simplified hybrid model is proposed to represent the dynamics of the robotic device in a realistic form. The proposed biped device is driven using linear actuators with a motion transfer system. A hybrid formulation for the control with adaptive state dependent gains regulates the articulations motion considering the limits of actuator motion. The dynamics of the adaptive gains is obtained with the application of a control barrier-like Lyapunov function for hybrid systems. The explicit structure of these gains are derived in a formal way. A set of numerical simulations is used to demonstrate the applicability of the developed controller analysing the tracking of bio-inspired reference trajectories obtained from reported biomechanical information. The numerical simulations used a virtual model of a biped robotic device where the interaction with the environment was considered for analysing the effect of hybrid evolution. A comparison between trajectories produced by the hybrid controller and a traditional state feedback offers a class of validation for the application of the restricted barrier inspired output feedback control strategy,

We are currently far from being able to make an accurate and timely prediction of earthquakes, that can tell us even roughly Where will it occur? When will it occur? What will its magnitude be? And at what depth will it occur? However, this information would be incredibly valuable to avoid loss of lives, damage to constructions, and a great economic loss. In this paper we propose a using recurrent neural networks (RNN) which have proven to be very efficient in time series analysis and based on the Gated recurrent unit (GRU) cell which is simple but powerful. This architecture achieves to solve these questions seconds before the earthquake and is applied to the seismicity of Italy from 1995 to 2018, considering the events with magnitude larger or equal to 1.5 and with a depth smaller than 60 km. Although a lot of work still needs to be done so that it can be applied in the real world as a prevention method, our results indicate that we are on the right track.

The prediction of the magnitude of an earthquake is still a challenge in the studies of seismic processes. The machine learning approaches have been developed recently to predict only the magnitude of next incoming earthquake using historical data. In this study, we combine neural networks with the meta-learning to predict not only the magnitude of the next earthquake, but also that of several earthquakes in the future. We successfully applied this novel method to predict the magnitude of Italian earthquakes above three different threshold magnitude. The experimental results show that the meta-learning based neural networks perform much better than the classical machine learning.

In this article, we discuss continuous-time$\mathcal {H}_{2}$control for the unknown nonlinear system. We use differential neural networks to model the system, then apply the$\mathcal {H}_{2}$tracking control based on the neural model. Since the neural$\mathcal {H}_{2}$control is very sensitive to the neural modeling error, we use reinforcement learning to improve the control performance. The stabilities of the neural modeling and the$\mathcal {H}_{2}$tracking control are proven. The convergence of the approach is also given. The proposed method is validated with two benchmark control problems.

This study has the aim of introducing a new type of trajectory tracking robust controllers for a class of rehabilitation robotic system considering the articulations restrictions. The robotic device consists of a suspended biped configuration. The suggested robust control considers the application of state depending gains which provide finite-time convergence for the tracking deviation. The state restrictions are fulfilled by the implementation of controller gains estimated by a class of the controlled tangent barrier Lyapunov function. Stability analysis for the tracking error yields the explicit design of the state dependent gains. The rate of convergence for the controller design is enhanced using a matrix inequality convex optimization method. Based on the forward complete characteristic of the suggested rehabilitation device, it is allowed using a finite-time convergent super-twisting-based differentiator to concrete an output feedback realization of the proposed controller. A computerized model of the tendered rehabilitation robot provides a reliable testing platform to the suggested roust controller. Numerical evaluations appear to serve as an indirect confirmation for the tracking error convergence, satisfying the articulation restrictions, and the effect of the gain optimization design. For comparison purposes, the regular state feedback control design is considered as benchmark. The faster convergence of the mean square estimation of the tracking error justifies the design of the proposed control design as well as the state feedback structure justifies the origin is a fixed-time stable equilibrium point for the space of tracking error at the same time that state space restrictions remain satisfied. The experimental evaluations of the proposed controller justifies the barrier controller which, in spite of the modeling uncertainties and the implementation issues, tracked the reference trajectories.

Bayesian inference shows that future event distribution not only depends on the past events (prior), but also depends on the relation between the past and the future events (likelihood). However, the classical Bayesian methods do not consider the important contributions of recent data.In this paper, we propose a new Bayesian inference based training method, which can be used as on-line training for Bayesian methods. We give the training methods for the exponential and the normal models. We successfully apply this method for the seismic parameter prediction using the data of central Italy from years 2014 to 2017. Comparisons on disscuss section in this work show our method is more effective than the other Bayesian methods.

In this article, we discuss$\mathcal {H}_{2}$control for unknown nonlinear systems in discrete time. A discrete-time recurrent neural network is used to model the nonlinear system, and then, the$\mathcal {H}_{2}$tracking control is applied based on the neural model. Since this neural$\mathcal {H}_{2}$control is very sensitive to the neural modeling error, we use reinforcement learning and another neural approximator to improve tracking accuracy and robustness of the controller. The stabilities of the neural identifier and the$\mathcal {H}_{2}$tracking control are proven. The convergence of the approach is also given. The proposed method is validated with the control of the pan and tilt robot and the surge tank.

Neural networks (NNs) are effective models for data-driven system identification. However, these data-based models do not give the probability properties of the system. Is it possible to use NNs to learn both the dynamics and input–output probability distributions to improve the modeling? In this article, we propose a special neural model, which combines the restricted Boltzmann machine (RBM) and the feedforward NNs. The stochastic RBM learns the input–output probability distributions. The feedforward NN learns the dynamics of nonlinear systems. For the input–output probability distributions, we give the calculation methods of their conditional and joint probabilities. The approximation abilities of this neural model are analyzed. We use two nonlinear systems to compare our guaranteed probability distribution method with the other black-box identification methods. The results show that this novel model is much better, when there are big noises and the dynamics of the system are complex.

This paper presents the development of an alternative Proportional-Derivative (PD) controller designed based only on system bandwidth to achieve the reduction of vibrations in buildings structures subjected to seismic excitation. A feature of the proposed controller is that it can be designed around an experimental test and the stability of the closed-loop system is ensured through the Kharitonov theorem and Bilharz criteria. The effectiveness of the proposed PD controller is illustrated using simulation results of a reduced-scale two-story building structure and the proposed algorithm is tested with data from the 1985 Mexico City earthquake. Due to the simplicity of the design, the methodology makes it a promising controller for practical applications.

The data driven black-box and gray-box models, like the neural networks and fuzzy systems, have some disadvantages, such as the high and uncertain dimensions and complex learning process. In this paper to affront these disadvantages, we use the Takagi-Sugeno fuzzy model and LSTM cells to propose a new fuzzy-network model. This novel model takes the advantages of the interpretability of the fuzzy system and the good approximation ability of the LSTM. We also propose a fast and stable learning algorithm for this model. Comparisons with others similar black-box and grey-box models are made, in order to observe the advantages of the proposal.

In this paper we discuss discrete-time H2control for unknown nonlinear system. We use recurrent neural networks to model the system identification, then apply H2tracking control. The neural networks based critic control does not require the system dynamics. Our optimal control policy uses a recursive solution of the discrete algebraic Riccati equation and reinforcement learning. The stabilities of system identification and H2tracking control are proven. he convergence of the approach is also given by the use of Lyapunov stability theory. The proposed method is validated with the control of a surge tank.

Reinforcement learning (RL) is an effective method to design robust control. Uncertainty in the worst case requires large state-action learning space. The continuous time RL can solve this computational problem. In this paper, we modify the classical continuous time RL. Compared with the actor-critic (AC) algorithm, our method is more simple and more robust under the worst-case uncertainty.

A Bayesian approach has been applied to estimate the distribution of magnitudes, interevent distances and times of earthquakes occurred in 2017 in central Italy by using a small amount of random samples drawn from the distribution of the same seismic parameters for the earthquakes occurred in 2014-2016. We applied the method to the whole and aftershock-depleted seismicity by using the exponential and the normal model to fit the distributions of the seismic parameters. Our findings indicate that the exponential model fits the distributions of the seismic parameters much better than the normal model. Furthermore, in the whole seismicity case, the method requires at least 2100 to 2300 random samples to estimate the distributions of the seismic parameters of earthquakes occurred in 2017 with an estimation error less than 0.01; while in the aftershock-depleted case, a minimum number of random samples varying between 360 and 1470 occurred in 2014-2017 is required to estimate the distributions of the seismic parameters of earthquakes occurred in 2017 with an estimation error less than 0.01.

Robot position/force control provides an interaction scheme between the robot and the environment. When the environment is unknown, learning algorithms are needed. But, the learning space and learning time are big. To balance the learning accuracy and the learning time, we propose a hybrid reinforcement learning method, which can be in both discrete and continuous domains. The discrete-time learning has poor learning accuracy and less learning time. The continuous-time learning is slow but has better learning precision. This hybrid reinforcement learning learns the optimal contact force, meanwhile it minimizes the position error in the unknown environment. Convergence of the proposed learning algorithm is proven. Real-time experiments are carried out using the pan and tilt robot and the force/torque sensor.

Fuzzy modeling has many advantages over nonfuzzy methods, such as robustness with respect to uncertainties and less sensitivity to the varying dynamics of nonlinear systems. Data-driven fuzzy modeling needs to extract fuzzy rules from input and output data, and to train the fuzzy parameters of the fuzzy model. This paper takes advantages from deep learning, probability theory, fuzzy modeling, and extreme learning machines (ELMs). Restricted Boltzmann machine (RBM) and probability theory are used to overcome some common problems in data-driven modeling methods. The RBM is modified such that it can be trained with continuous values. A probability-based clustering method is proposed to partition the hidden features from the RBM. The obtained fuzzy rules have probability measurement. ELM and an optimization method are applied to train the fuzzy model. The proposed method is validated with two benchmark problems.

In this work a large space dimension reinforcement learning (RL) approximation is developed for a Discrete Impedance Position/Force control of robot manipulators that interacts with an unknown environment model. The Q-value function is designed in the sense of optimal control theory. The approximator is based on normalized radial basis functions (NRBFs), and are built using the K-means clustering algorithm which generates a family of approximators for the Q-value function. The RL algorithms learn on-line the optimal impedance model which is equivalent to the desired force without any prior knowledge of the environment dynamics; this feeds a force controller and its output feeds the position controller. Real time experiments are shown using a 2 degree of freedom (DOF) pan and tilt robot and a 6-DOF force/torque (F/T) sensor.

Autonomous navigation in unknown environment is a big challenge. In this paper, we combine the SLAM (simultaneous localization and mapping) with the path planning method. We first modify the classical SLAM with sliding mode technique, such that it is robust in the unknown environment. Then we analyze the algorithm using the “known space” and “free space” conditions, and propose the polar histogram path planning based on these conditions. We use Monte Carlo method to evaluate the performance of our algorithms. Simulation results show that our autonomous navigation algorithms are better than the others in unknown environment.

Online vehicle routing is an important task of the modern transportation service provider. Contributed by the ever-increasing real-time demand on the transportation system, especially small-parcel last-mile delivery requests, vehicle route generation is becoming more computationally complex than before. The existing routing algorithms are mostly based on mathematical programming, which requires huge computation time in city-size transportation networks. To develop routes with minimal time, in this paper, we propose a novel deep reinforcement learning-based neural combinatorial optimization strategy. Specifically, we transform the online routing problem to a vehicle tour generation problem, and propose a structural graph embedded pointer network to develop these tours iteratively. Furthermore, since constructing supervised training data for the neural network is impractical due to the high computation complexity, we propose a deep reinforcement learning mechanism with an unsupervised auxiliary network to train the model parameters. A multisampling scheme is also devised to further improve the system performance. Since the parameter training process is offline, the proposed strategy can achieve a superior online route generation speed. To assess the proposed strategy, we conduct comprehensive case studies with a real-world transportation network. The simulation results show that the proposed strategy can significantly outperform conventional strategies with limited computation time in both static and dynamic logistic systems. In addition, the influence of control parameters on the system performance is investigated.

Critical Infrastructure Systems (CISs) play an essential role in our life, when disasters, attacks, failures happen, such complex systems are expected to be reliable and safety, even react to undesirable accidents. Modeling CISs and developing methods to analyze their safety and dependability is of utmost importance. CIS modeling formalisms must be able to describing both discrete and continuous quantities, a hybrid system modelling approach is natural. In this work, CISs are modeled from two aspects: logic and continuous; Adaptive fuzzy Petri nets (AFPN) and neural networks are combined in our framework, where AFPN is adopted to model the logic parts, and dynamic neural networks are applied to continuous parts. Two hybrid system examples are illustrated to show the effectiveness of the proposed approach.

Normal SLAMs use the extended Kalman filter to estimate robot localization and the mapping simultaneously. They do not work well under big disturbances and bounded noises. In this paper, the sliding mode method is applied for the SLAM. The proposed sliding model SLAM only requires the noises and the disturbances are bounded. The estimation errors are analyzed, and the stability of the novel SLAM is proposed. A mobile robot is applied in the experiment to show the effectiveness of the sliding mode SLAM in the presence of bounded noises.

This paper takes advantages from probability theory and fuzzy modeling. We use probability theory to overcome some common problems in data based modeling methods. A probability based clustering method is proposed to partition the hidden features, and extract fuzzy rules with probability measurement. An optimization method are applied to train the consequent part of the fuzzy rules and the probability parameters. The proposed method is validated with two benchmark problems.

The identification problem incorporated in feedback control of uncertain nonlinear systems exhibiting complex behavior has been solved in different ways. Some of these solutions have used artificial intelligence methods like fuzzy logic and neural networks. However, their individual implementation suffers from certain drawbacks, such as the black-box nature of neural network and the problem of finding suitable membership functions for fuzzy systems. These weaknesses can be avoided by implementing a hybrid structure combining these two approaches, the so-called neuro-fuzzy system. In this paper, a neuro-fuzzy system that implements differential neural networks (DNNs) as consequences of Takagi-Sugeno (T-S) fuzzy inference rules is proposed. The DNNs substitute the local linear systems that are used in the common T-S method. In this paper, DNNs are used to provide an effective instrument for dealing with the identification of the uncertain nonlinear system, while the T-S rules are used to provide the framework of previous knowledge of the system. The main idea is to carry out an online identification process of an uncertain nonlinear system with the aim to design a closed-loop trajectory tracking controller. The methodology developed in this study that supports the identification and trajectory control designs is based on the Lyapunov formalism. The DNN implementation results in a time-varying T-S system. As a consequence, the solution of two time-varying Riccati equations was used to adjust the learning laws in the DNN as well as to adjust the gains of the controller. Two results were provided to justify the existence of positive-definite solutions for the class of Riccati equations used in the learning laws of DNNs. A complete description of the learning laws used for the set of DNN identifiers is also obtained. An autonomous underwater vehicle system is used to demonstrate the performance of the controller on tracking a desired 3-D path by this combination of the DNN and the T-S system.

Many Takagi-Sugeno (T-S) fuzzy systems use the linear matrix inequality (LMI) to design the controllers. Although LMI can be solved off-line, there are still some difficulties in T-S fuzzy control design with LMI, such as complexities in the analysis and the computation, and conservativeness for complex systems. In this paper, we first transform the T-S fuzzy system into a time-varying nonlinear system. Then we apply the Riccati differential equation to design the fuzzy control. The stability of the T-S fuzzy control is proven. The trajectory tracking error converges to a bounded zone. This novel design method for T-S fuzzy control is more simple and effective than LMI-based methods. The on-line solutions of the Riccati equation can be obtained directly from a designed differential equation.

Many cascade processes, such as wastewater treatment plant, include complex nonlinear sub-systems and many variables. The normal input-output relation only represent the first block and the last block of the cascade process. In order to model the whole process. We use hierarchical dynamic neural networks to identify the cascade process. The internal variables of the cascade process are estimated. Two stable learning algorithms and theoretical analysis are given. Real operational data of a wastewater treatment plant are applied to illustrate this new neural modeling approach.

With fuzzy set theory, the uncertainty nonlinear systems can be modeled with fuzzy equations or fuzzy differential equations (FDEs). The solutions of them are applied to analyze many engineering problems. However, it is very difficult to obtain solutions of FDEs. In this paper, the solutions of FDEs are approximated by two type of Bernstein neural networks. We first transform the FDE into four ordinary differential equation (ODEs) with Hukuhara differentiability. Then we construct neural models with the structure of ODEs. With modified backpropagation method for fuzzy variables, the neural networks are trained. The theory analysis and simulation results show that these new models, Bernstein neural networks, are effective to estimate the solutions of FDEs.

Deep learning techniques have been successfully used for pattern classification. These advantage methods are still not applied in nonlinear systems identification. In this paper, the neural model has deep architecture which is obtained by a random search method. The initial weights of this deep neural model is obtained from the denoising autoencoders model. We propose special unsupervised learning methods for this deep learning model with input data. The normal supervised learning is used to train the weights with the output data. The deep learning identification algorithms are validated with three benchmark examples.

Because the range of mobile robot sensors is limited and navigation maps are not always accurate, autonomous navigation in dynamic and unknown environments is a big challenge. In this article, we propose two novel autonomous navigation algorithms, which are based on the analysis of three conditions for unobserved and uncertain environments during navigation.The algorithm for a dynamic environment uses the “known space” and “free space” conditions. It corrects false obstacles in the map when the conventional path is stuck. The navigation algorithm for unknown environments uses the “unknown space” and “free space” conditions. We use the Monte Carlo method to evaluate the performance of our algorithms and the other methods. Experimental results show that our autonomous navigation algorithms are better than the others.

Data-driven modeling based on the shell vibration and acoustic signals of ball mills is normally applied to overcome the subjective errors of human inference. Many previously proposed selective ensemble (SEN) modeling approaches are based on “the manipulation of input features” from the multiinformation fusion perspective, which cannot selectively and jointly fuse the information hidden in multiscale spectral features and under several operating conditions (training samples). Therefore, this study suggests a new soft measuring procedure based on ensemble empirical mode decomposition (EEMD) and SEN. An improved kernel partial least-squares algorithm for SEN that is based on “subsample training samples” is utilized to construct a soft measuring model with the selected features and training samples. This study compares such data-driven soft measuring methods. The comparative results of bootstrap-based prediction performance estimation show that different methods have specific advantages in terms of simplicity, prediction accuracy, and interpretability. The industrial application of the EEMD-SEN method is discussed in this paper, and a new virtual sample generation method is proposed to address the modeling problem based on small sample spectral data.

In this paper, we use a deep learning method, restricted Boltzmann machine, for nonlinear system identification. The neural model has deep architecture and is generated by a random search method. The initial weights of this deep neural model are obtained from the restricted Boltzmann machines. To identify nonlinear systems, we propose special unsupervised learning methods with input data. The normal supervised learning is used to train the weights with the output data. The modified algorithm is validated by modeling two benchmark systems.

Many uncertain nonlinear systems can be modeled by the linear-in-parameter model, and the parameters are uncertain in the sense of fuzzy numbers. Dual fuzzy equations are alternative models for these nonlinear systems identification, while the solutions of the fuzzy equations are the controllers. In this paper, we propose a novel fuzzy controller via dual fuzzy equations. Two types of neural networks are applied to approximate the solutions of the fuzzy equations. These solutions are then transformed into the fuzzy controllers. The novel methods are validated with two benchmark examples.

Passivity method is very effective to analyze large-scale nonlinear systems with strong nonlinearities. However, when most parts of the nonlinear system are unknown, the published neural passivity methods are not suitable for feedback stability. In this brief, we propose a novel sliding mode learning algorithm and sliding mode feedback passivity control. We prove that for a wide class of unknown nonlinear systems, this neural sliding mode control can passify and stabilize them. This passivity method is validated with a simulation and real experiment tests.

Although normal fuzzy sliding mode controllers can reduce the chattering problem in building structure control, there are some problems such as they need the equivalent control and the upper bounds of the uncertainties. In this paper, we use fuzzy logic to approximate the sliding surface for the sliding mode control. The stability of the proposed controller is established. A six-story building prototype equipped with an active mass damper is used to demonstrate the effectiveness of the proposed controller towards the wind-induced vibration.

Gasoline blending is an important unit operation in gasoline industry. A good model for the blending system is beneficial for supervision operation, prediction of the gasoline qualities and performing model-based optimal control. Gasoline blending process involves two types of proprieties: static blending and dynamic in the blending tanks. The blending process cannot be modeled exactly, because it does not follow ideal mixing rules in practice. In this paper we propose a hybrid neural network, which uses static and dynamic neural networks to approximate the blending properties. Numerical simulations are provided to illustrate the neuro modeling approach.

Human guide robots need to generate a trajectory from human training. The popular work space methods have to calculate the inverse kinematics. While the joint space methods need the dynamic time warping. These destroy the accuracy of the trajectory model. In this paper, we use Lloyd's algorithm to hidden Markov model (HMM). The advantages of the method over the other HMM are the time difference does not affects the HMM training, and the training data can be generated in joint space. We also modify the traditional HMM such that the model in the joint space works similar as the task space. Simulation and experimental results show that the modified HMM with Lloyd's algorithm in joint space is effective to generate the desired trajectory.

Path planning of multi-agent is much harder than single-agent. Reinforcement learning (RL) is a popular method for it. However, it cannot solve the path planning problem directly in unknown environment. In this paper, neural network (NN) is applied to estimate the unvisited space. The traditional multi-agent reinforcement learning is modified by the neural approximation. The path planning of this paper includes two stages: we first use RL to generate training samples for NN; then the trained NN gives an approximate action to agents. The advantage of this method is we do not need to repeat RL for the unvisited state. Experiment results show the proposed algorithm can generate suboptimal paths in the unknown environment for multiple agents.

Because the range of mobile robot sensor is limited and the navigation map is not accurate, autonomous navigation in dynamic and unknown environment is a big challenge. In this paper, we propose two novel autonomous navigation algorithms, which are based on the analysis of three conditions for unobserved and uncertainty environment during the navigation. The algorithm for dynamic environment uses the “known space” and “free space” conditions. It corrects false obstacles in the map when conventional path is stuck. The navigation algorithm for unknown environment uses the “unknown space” and “free space” conditions. We use Monte Carlo method to evaluate the performance of our algorithms and the other methods. Experimental results show that our autonomous navigation algorithms are better than the others.

Many real data sets are imbalanced and contain a large number of a certain type of patterns, but a very small number of another type of patterns. Normal classification methods, such as support vector machine (SVM), do not work well for these imbalanced data sets (IDS). It is difficult for SVMs to get the optimal separation hyperplane when they are trained with imbalanced data. In this paper, we propose a genetic algorithm (GA)-based classification method. A draft hyperplane and support vectors are first generated by SVMs. Then, GA is applied to compensate the imbalanced data. Finally, SVM is used again to find the best hyperplane from the generated data points. Compared with the other popular classification algorithms, our method has better classification accuracy for several IDS.

In this paper we propose a novel feedback passive controller which uses a two-neuro dynamic neural network. It is robust for a wide class of nonlinear systems with a priory incomplete model description. By means of a Lyapunov-like analysis, both identification and passivation effects are guaranteed. Based on this neuro model we design an feedback passive controller. The example illustrate the effectiveness of the suggested approach.

Many real data sets are imbalanced, which contain a large number of certain type objects and a very small number of opposite type objects. Normal classification methods, such as support vector machine (SVM), do not work well for these skewed data sets. In this paper we propose a genetic algorithm (GA) based classification method. We first use SVM to generate a draft hyper plane and support vectors. Then GA is applied to find new data points in the sensible region or classification margin. Finally, SVM is used again to find the best hyper plane from the generated data points. Compared with the other popular classification algorithms, the proposed method has better classification accuracy for several skewed data sets.

In this paper we develop a wireless communication human machine interface (HMI) system, which uses a smart-phone. This HMI has some advantages over the other HMIs, such as it is small, cheap, and space sensing. It uses the accelerometer and the gyroscope of the smartphone to generate six commands to control a robot via Wi-Fi network. The experiment results show that our HMI is convenient and effective for remote control.

Due to complex dynamic characteristics of the ball mill system, it is difficult to measure load parameters inside the ball mill. It has been noticed that the traditional single-model and ensemble-model based soft sensor approaches demonstrate weak generalization power. Also, mill motor current, feature subsets of the shell vibration and acoustical frequency spectra contain different useful information. To achieve better solutions and overcome these problems mentioned above, a selective ensemble multisource information approach is proposed in this paper. Only the useful feature subsets of vibration and acoustical frequency spectra are portioned and selected. Some modeling techniques, such as fast Fourier transform (FFT), mutual information (MI), kernel partial least square (KPLS), brand and band (BB), and adaptive weighting fusion (AWF), are combined effectively to model the mill load parameters. The simulation is conducted using real data from a laboratory-scale ball mill. The results show that our proposed approach can effectively fusion the shell vibration, acoustical and mill motor current signals with improved model generalization.

In order to minimize steady-state error with respect to uncertainties in robot control, proportional-integral-derivative (PID) control needs a big integral gain, or a neural compensator is added to the classical proportional-derivative (PD) control with a large derivative gain. Both of them deteriorate transient performances of the robot control. In this paper, we extend the popular neural PD control into neural PID control. This novel control is a natural combination of industrial linear PID control and neural compensation. The main contributions of this paper are semiglobal asymptotic stability of the neural PID control and local asymptotic stability of the neural PID control with a velocity observer which are proved with standard weight training algorithms. These conditions give explicit selection methods for the gains of the linear PID control. An experimental study on an upper limb exoskeleton with this neural PID control is addressed.

Support Vector Machine (SVM) is an important classification method used in a many areas. The training of SVM is almost O(n^{2}) in time and space. Some methods to reduce the training complexity have been proposed in last years. Data selection methods for SVM select most important examples from training data sets to improve its training time. This paper introduces a novel data reduction method that works detecting clusters and then selects some examples from them. Different from other state of the art algorithms, the novel method uses a decision tree to form partitions that are treated as clusters, and then executes a guided random selection of examples. The clusters discovered by a decision tree can be linearly separable, taking advantage of the Eidelheit separation theorem, it is possible to reduce the size of training sets by carefully selecting examples from training sets. The novel method was compared with LibSVM using public available data sets, experiments demonstrate an important reduction of the size of training sets whereas showing only a slight decreasing in the accuracy of classifier.

Support Vector Machines (SVMs) are high-accuracy classifiers. However, normal SVM algorithms are unsuitable for classification of large data sets because of their training complexity. In this paper, we propose a novel SVM classification approach for

In the construction of a Wavelet Neural Network, the number of neurons is determined by the traslation coefficient and by the dilations coefficient. Exists two ways to set the value of the traslation coefficients and dilation, one is considering the coefficients like a hidden layer of the network and the other way is establishing fixed values to those coefficients, where there remains the problem of establishing the number of fixed values to be taken, in this paper we present an algorithm to determine the number of fixed values, that they minimize a rate that depends on the approximation error and the number of neurons that are used.

Both neural network and sliding mode technique can compensate the steady-state error of proportional-derivative (PD) control. The tracking error of PD control with sliding mode is asymptotically stable, but the chattering is big. PD control with neural networks is smooth, but it is not asymptotically stable. PD control combining both neural networks and sliding mode cannot reduce chattering, because the sliding mode control (SMC) is always applied. In this paper, neural control and SMC are connected serially: first a dead-zone neural PD control assures that the tracking error is bounded, then super-twisting second-order sliding-mode is used to guarantee finite time convergence of the sliding mode PD control.

Component concentration of sodium aluminate solution is an important quality index for alumina production. In this article, we propose a new on-line soft sensing strategy for measuring component concentration of sodium aluminate solution. With this method, on-line control can be realised in aluminate production plants. Several advance techniques are used, such as principal component analysis (PCA), neural modelling and the least square algorithm. Industry experiments are conducted in the alumina production process and the results show the effectiveness of this method.

It is difficult to establish a black-box model for sparse data, because not enough data can be applied for training. This paper presents a novel identification approach using multiple fuzzy neural networks. It focuses on structure and parameters uncertainty which have been widely explored in the literature. Firstly, the sparse data are used within a fixed time interval to generate model structure. Then kernel regression methods are used to generate training data, a stable updating algorithm is proposed to train the membership functions. To cope structure change, a hysteresis strategy is proposed to enable multiple fuzzy neural identifier switching with guaranteed performance. Both theoretic analysis and simulation example show the efficacy of the proposed method.

Compared to normal learning algorithms, for example backpropagation, Kalman filter-based algorithm has some better properties, such as faster convergence, although this algorithm is more complex and sensitive to the nature of noises. In this paper, extended Kalman filter is applied to train recurrent wavelets neural networks for nonlinear system identification. In order to improve robustness of Kalman filter algorithm dead-zone robust modification is applied to Kalman filter. Lyapunov method is used to prove that this new training approach is stable.

Cascade process, such as wastewater treatment plant, includes many nonlinear sub-systems and many variables. When the number of sub-systems is big, the input-output relation in the first block and the last block cannot represent the whole process. In this paper we use two techniques to overcome the above problem. Firstly we propose a new neural model: hierarchical neural networks to identify the cascade process; then we use serial structural mechanism model based on the physical equations to connect with neural model. A stable learning algorithm and theoretical analysis are given. Finally, this method is used to model a wastewater treatment plant. Real operational data of wastewater treatment plant is applied to illustrate the modeling approach.

This paper, presents a novel identification approach using fuzzy neural networks. It focuses on structure and parameters uncertainties which have been widely explored in the literatures. The main contribution of this paper is that an integrated analytic framework is proposed for automated structure selection and parameter identification. A kernel smoothing technique is used to generate a model structure automatically in a fixed time interval. To cope with structural change, a hysteresis strategy is proposed to guarantee finite times switching and desired performance.

In general, neural networks cannot match non-linear systems exactly. Neuro identifier has to include robust modification in order to guarantee Lyapunov stability. In this paper input-to-state stability approach is applied to access robust training algorithms of Fourier neural network (FoNN). It is successfully applied on modeling lettuce growth in green-house.

Combination of neural networks and sliding mode control (SMC) can reduce chattering, because the upper bound of uncertainties becomes smaller when neural networks are used to model unknwn nolinear systems. The tracking error of normal neural sliding mode control is asymptotically stable, while neural control and SMC are applied at same time. In this paper, neural control and SMC are connected serially: first a deadzone neural control assures that the tracking error is bounded, then super-twisting secondorder slidingmode is used to guarantee finite time convergence of the contoller.

Support vector machines (SVMs) are known to be excellent algorithms for classification problems. The principal disadvantage of SVMs is due to its excessive training time in large data set, such as DNA sequences. This paper presents a novel SVMs classification method which reduces significantly the input data set using Bayesian technique. Using this system, we are able to predict with a high accuracy huge data sets in a reasonable time. The system has been tested successfully on large splice-junction gene sequences (DNA). Experimental results show that the accuracy obtained by the proposed algorithm is comparable (98.2) with other SVMs implementations such as SMO (98.4%), LibSVM (98.4%), and Simple SVM (97.6%). Furthermore the proposed approach is scalable to large data sets with high classification accuracy.

This brief is concerned with the global robust exponential stability of a class of interval Cohen-Grossberg neural networks with both multiple time-varying delays and continuously distributed delays. Some new sufficient robust stability conditions are established in the form of state transmission matrix, which are different from the existing ones. Furthermore, a sufficient condition is also established to guarantee the global stability for this class of Cohen-Grossberg neural networks without uncertainties. Three examples are used to show the effectiveness of the obtained results.

Bounding ellipsoid (BE) algorithms offer an attractive alternative to traditional training algorithms for neural networks, for example, backpropagation and least squares methods. The benefits include high computational efficiency and fast convergence speed. In this paper, we propose an ellipsoid propagation algorithm to train the weights of recurrent neural networks for nonlinear systems identification. Both hidden layers and output layers can be updated. The stability of the BE algorithm is proven.

This paper discusses the identification of nonlinear dynamic system using fuzzy neural networks. It focuses on both the structure uncertainty and the parameter uncertainty which have been widely explored in the literature of nonlinear system identification. The main contribution is that an integrated analytic framework is proposed for automated fuzzy neural network structure selection, parameter identification and hysteresis network switching with guaranteed neural identification performance. Firstly, an automated support vector machine is proposed within a fixed time interval for a given network construction criterion. Then the network parameter updating algorithm is proposed with guaranteed bounded identification error. To cope structure uncertainty, a hysteresis strategy is proposed to enable fuzzy neural identifier switching with guaranteed network performance along the switching process. Both theoretic analysis and simulation example show the efficacy of the proposed method.

In this paper, we propose a new type of neural adaptive control via dynamic neural networks. For a class of unknown nonlinear systems, a neural identifier-based feedback linearization controller is first used. Dead-zone and projection techniques are applied to assure the stability of neural identification. Then four types of compensator are addressed. The stability of closed-loop system is also proven.

In recent years support vector machines (SVM) has received considerable attention due to its high generalization ability and performance for a wide range of applications. However, the most important problem of this method is slow training for classification problems with a large data sets because the quadratic form is completely dense and the memory requirements grow with the square of the number of data points. This paper presents a novel SVM classification approach for large data sets by reducing training data and train the support vector machine using only these data. In this algorithm, a first stage uses SVM classification on a small data set in order to gets a sketch of classes distribution and labels the support vectors as a data set with label +1 and the other points as a data set with label -1. We call this change of classes. Then the algorithm obtains the classification hyperplane and classify the original input data set, the data points obtained with label +1 constitute the data points in the boundary of each original class and represent the most important data points, these data points are used as training data for a posterior SVM classification. The effectiveness of the approach proposed is supported by experimental results.

Since the fuzzy cerebellar model articulation controller (FCMAC) uses linguistic variables, it is highly intuitive and easily comprehended. Despite the FCMAC's good local generalization capability for approximating nonlinear functions and fast learning, a normal FCMAC requires huge memory, and its dimension increases exponentially with the number of inputs. In order to overcome the memory explosion problem, this paper proposes two types of hierarchical FCMAC (HFCMAC). Another contribution of the paper is that we give stable learning algorithms for these two HFCMACs. Backpropagation-like approach is applied to train each block with a time-varying learning rate, which is obtained by the input-to-state stability technique.

Normal fuzzy CMAC neural network performs well for nonlinear systems identification because of its fast learning speed and local generalization capability for approximating nonlinear functions. However, it requires huge memory and the dimension increases exponentially with the number of inputs, and it is difficult for its static structure to model a dynamic system. In this paper, we use two types of recurrent techniques for fuzzy CMAC to overcome the above problems. The new CMAC neural networks are named recurrent fuzzy CMAC (RFCMAC) which add feedback connections in the inner layers (local feedback) or the output layer (global feedback). The corresponding learning algorithms are presented that have time-varying learning rates, the stabilities of the neural identifications are proven.

Despite of good theoretic foundations and high classification accuracy of support vector machine (SVM), normal SVM is not suitable for classification of large data sets, because the training complexity of SVM is very high. This paper presents a novel two stages SVM classification approach for large data sets by randomly selecting training data. The first stage SVM classification gets a sketch of support vector distribution. Then the neighbors of these support vectors in original data set are used as training data for the second stage SVM classification. Experimental results demonstrate that our approach have good classification accuracy while the training is significantly faster than other SVM classifiers.

In this paper we propose a novel online clustering approach which can be applied for nonlinear system modeling. Fuzzy neural networks are used as models whose structure and parameters are updated online. The new idea for the structure identification is that the input (precondition) and the output (consequent) spaces partitioning are carried out in the same time index. This idea gives better explanation for input-output mapping of nonlinear system. An application on modeling of crude oil blending is proposed.

Dynamic neural networks with different time-scales include the aspects of fast and slow phenomenons. Some applications require that the equilibrium points of these networks be stable. The objective of the paper is to develop sufficient conditions for stability of the dynamic neural networks with different time scales. Lyapunov function and singularly perturbed technique are combined to access several new stable properties of different time-scales neural networks. Exponential stability and asymptotic stability are obtained by sector and bound conditions. Compared to other papers, these conditions are simpler. Numerical examples are given to demonstrate the effectiveness of the theoretical results.

This paper considers the problem of PD control of overhead crane in the presence of uncertainty associated with crane dynamics. By using radial basis function neural networks, these uncertainties can be compensated effectively. This new neural control can resolve the two problems for overhead crane control: 1) decrease steady-state error of normal PD control. 2) guarantee stability via neural compensation. Lyapunov method and input-to-state stability technique, we prove that these robust controllers with neural compensators are stable. Real-time experiments are presented to show the applicability of the approach presented in this paper.

Dynamic neural networks with different time-scales include the aspects of fast and slow phenomenons. Some applications require that the equilibrium points of these networks to be stable. The main contribution of the paper is that Lyapunov function and singularly perturbed technique are combined to access several new stable properties of different time-scales neural networks. Exponential stability and asymptotic stability are obtained by sector and bound conditions. Compared to other papers, these conditions are simpler. Numerical examples are given to demonstrate the effectiveness of the theoretical results.

In this paper we propose a novel on-line clustering approach which can be applied for nonlinear system identification. Both structure and parameters of fuzzy neural networks are updated on-line. The new clustering method for the structure identification can divide input/output data into different groups (rule number) by on-line data. For the parameter learning, our algorithm has two advantages over the others. First, the normal methods for parameter identification are based on a fixed structure and whole data, for example ANFIS by C. F. Jang and C. Teng Lin (1998), but after clustering we know each group corresponds to one rule, so we train each rule by its group data, it is more effective. Second, we give a time-varying learning rate for the common used backpropagation algorithm, we prove that the new algorithm is stable and faster than backpropagation algorithm

Compared to normal learning algorithms, for example backpropagation, Kalman filter-based algorithm has some better properties, such as faster convergence. In this paper, Kalman filter is modified with a risk-sensitive cost criterion, we call it as risk-sensitive Kalman filter. This new algorithm is applied to train recurrent neural networks for nonlinear system identification. Input-to-state stability is used to prove that the risk-sensitive Kalman filter training is stable. The contributions of this paper are: 1) the risk-sensitive Kalman filter is used for the state-space recurrent neural networks training, 2) the stability of the risk-sensitive Kalman filter is proved.

Crude oil blending is an important unit in petroleum refining industry. Many blend automation systems use real-time optimizer (RTO), which apply current process information to update the model and predict the optimal operating policy. The key unites of the conventional RTO are on-line analyzers. Sometimes oil fields cannot apply these analyzers. In this paper, we propose an off-line optimization technique to overcome the main drawback of RTO. We use the history data to approximate the output of the on-line analyzers, then the desired optimal inlet flow rates are calculated by the optimization technique. After this off-line optimization, the inlet flow rates are used for on-line control, for example PID control, which forces the flow rate to follow the desired inlet flow rates. Neural networks are applied to model the blending process from the history data. The new optimization is carried out via the neural model. The contributions of this paper are: (1) Stable learning for the discrete-time multilayer neural network is proposed. (2) Sensitivity analysis of the neural optimization is given. (3) Real data of a oil field is used to show effectiveness of the proposed method.

Gasoline blending is an important operation in chemical industry. A good model for the blending process is beneficial for supervision operation, prediction of gasoline qualities and realizing model-based optimal control. Gasoline blending process includes static and dynamic properties which are corresponded to thermodynamic and the storage tank respectively. Since the blending does not follow the ideal mixing rule in practice, we propose static and dynamic neural networks to approximate the blending process. Input-to-state stability approach is applied to access new robust learning algorithms of the neural networks. Numerical simulations are provided to illustrate the neuro modelling approaches.

Nonlinear system online identification via dynamic neural networks is studied in this paper. The main contribution of the paper is that the passivity approach is applied to access several new stable properties of neuro identification. The conditions for passivity, stability, asymptotic stability, and input-to-state stability are established in certain senses. We conclude that the gradient descent algorithm for weight adjustment is stable in an L(infinity) sense and robust to any bounded uncertainties.

In this paper the popular PD controller of robot manipulator is modified. RBF neural networks are used to compensate the gravity and friction. No exact knowledge of the robot dynamics is required. The suggested learning law of neuro compensator is similar to the well-known backpropagation algorithm but with additional robust terms. Lyapunov-like analysis is used to derive the stability of learning algorithm.

Since knowledge in an expert system is vague and modified frequently, expert systems are fuzzy and dynamic. It is very important to design a dynamic knowledge inference framework which is adjustable according to knowledge variation as human cognition and thinking. A generalized fuzzy Petri net model, called adaptive fuzzy Petri net (AFPN), is proposed with this object in mind. AFPN not only has the descriptive advantages of the fuzzy Petri net, it also has learning ability like a neural network. Just as other fuzzy Petri net (FPN) models, AFPN can be used for knowledge representation and reasoning, but AFPN has one important advantage: it is suitable for dynamic knowledge, i.e., the weights of AFPN are adjustable. Based on the AFPN transition firing rule, a modified backpropagation learning algorithm is developed to assure the convergence of the weights.

We develop a new neurocontroller for robot manipulators. A simple dynamic neural network is used to estimate the unknown robot manipulators, then the direct linearization controller is derived via this neuro-identifier. Since the approximation capability is limited, another robust sliding mode compensator is addressed. In this paper we give a bound for the identification error of the parallel neuro-identifier, and establish a bound for the tracking error of the hybrid controller.

This paper proposes a new online identification method for a class of partial-state measurement nonlinear systems. Only input and output are available, the inner state and the structure are unknown. The design of this paper is based on the combination of the state observer with the neuro identifier. As no information of the nonlinear system can be used, first a model-free high-gain observer is designed to estimate the inner state. Then a dynamic multilayer neural network is used to identify the nonlinear system based on the full observed states. By means of a Lyapunov-like analysis we determine the stable learning algorithms for the observer-based neuro identifier.

To identify online a quite general class of nonlinear systems, this paper proposes a new stable learning law for dynamic multilayer neural networks. A Lyapunov-like analysis is used to derive this stable learning procedure for the hidden layer as well as for the output layer.

In this paper the adaptive nonlinear identification and trajectory tracking are discussed via dynamic neural networks. By means of a Lyapunov-like analysis we determine stability conditions for the identification error. Then we analyze the trajectory tracking error by a local optimal controller. An algebraic Riccati equation and a differential one are used for the identification and the tracking error analysis. As our main original contributions, we establish two theorems: the first one gives a bound for the identification error and the second one establishes a bound for the tracking error. We illustrate the effectiveness of these results by two examples: the second-order relay system with multiple isolated equilibrium points and the chaotic system given by Duffing equation.