L_2-L_∞ learning of dynamic neural networks

This paper proposes an y2-y∞ learning law as a new learning method for dynamic neural networks with external disturbance. Based on linear matrix inequality （LMI） formulation, the y2-y∞ learning law is presented to not only guarantee asymptotical stability of dynamic neural networks but also reduce the effect of external disturbance to an y2-y∞ induced norm constraint. It is shown that the design of the y2-y∞ learning law for such neural networks can be achieved by solving LMIs, which can be easily facilitated by using some standard numerical packages. A numerical example is presented to demonstrate the validity of the proposed learning law.

Study on Ecological Change Remote Sensing Monitoring Method Based on Elman Dynamic Recurrent Neural Network

In this paper, Hailin City of Heilongjiang Province, China is taken as the research area. As an important city in Heilongjiang Province, China, the sustainable development of its ecological environment is related to the opening up, economic prosperity and social stability of Northeast China. In this paper, the remote sensing ecological index (RSEI) of Hailin City in recent 20 years was calculated by using Landsat 5/8/9 series satellite images, and the temporal and spatial changes of the ecological environment in Hailin City were further analyzed and the influencing factors were discussed. From 2003 to 2023, the mean value of RSEI in Hailin City decreased and increased, and the ecological environment decreased slightly as a whole. RSEI declined most significantly from 2003 to 2008, and it increased from 2008 to 2013, decreased from 2013 to 2018, and increased from 2018 to 2023 again, with higher RSEI value in the south and lower RSEI value in the northwest. It is suggested to appropriately increase vegetation coverage in the northwest to improve ecological quality. As a result, the predicted value of Elman dynamic recurrent neural network model is consistent with the change trend of the mean value, and the prediction error converges quickly, which can accurately predict the ecological environment quality in the future study area.

DyPipe: A Holistic Approach to Accelerating Dynamic Neural Networks with Dynamic Pipelining

Dynamic neural network(NN)techniques are increasingly important because they facilitate deep learning techniques with more complex network architectures.However,existing studies,which predominantly optimize the static computational graphs by static scheduling methods,usually focus on optimizing static neural networks in deep neural network(DNN)accelerators.We analyze the execution process of dynamic neural networks and observe that dynamic features introduce challenges for efficient scheduling and pipelining in existing DNN accelerators.We propose DyPipe,a holistic approach to optimizing dynamic neural network inferences in enhanced DNN accelerators.DyPipe achieves significant performance improvements for dynamic neural networks while it introduces negligible overhead for static neural networks.Our evaluation demonstrates that DyPipe achieves 1.7x speedup on dynamic neural networks and maintains more than 96%performance for static neural networks.

Design and performance analysis of tracking controller for uncertain nonlinear composite system using neural networks

Based on high order dynamic neural network, this paper presents the tracking problem for uncertain nonlinear composite system, which contains external disturbance, whose nonlinearities are assumed to be unknown. A smooth controller is designed to guarantee a uniform ultimate boundedness property for the tracking error and all other signals in the dosed loop. Certain measures are utilized to test its performance. No a priori knowledge of an upper bound on the “optimal” weight and modeling error is required; the weights of neural networks are updated on-line. Numerical simulations performed on a simple example illustrate and clarify the approach.

Recursive recurrent neural network:A novel model for manipulator control with different levels of physical constraints

Manipulators actuate joints to let end effectors to perform precise path tracking tasks.Recurrent neural network which is described by dynamic models with parallel processing capability,is a powerful tool for kinematic control of manipulators.Due to physical limitations and actuation saturation of manipulator joints,the involvement of joint constraints for kinematic control of manipulators is essential and critical.However,current existing manipulator control methods based on recurrent neural networks mainly handle with limited levels of joint angular constraints,and to the best of our knowledge,methods for kinematic control of manipulators with higher order joint constraints based on recurrent neural networks are not yet reported.In this study,for the first time,a novel recursive recurrent network model is proposed to solve the kinematic control issue for manipulators with different levels of physical constraints,and the proposed recursive recurrent neural network can be formulated as a new manifold system to ensure control solution within all of the joint constraints in different orders.The theoretical analysis shows the stability and the purposed recursive recurrent neural network and its convergence to solution.Simulation results further demonstrate the effectiveness of the proposed method in end‐effector path tracking control under different levels of joint constraints based on the Kuka manipulator system.Comparisons with other methods such as the pseudoinverse‐based method and conventional recurrent neural network method substantiate the superiority of the proposed method.

Hybrid dynamic model of polymer electrolyte membrane fuel cell stack using variable neural network

The polymer electrolyte membrane（PEM） fuel cell has been regarded as a potential alternative power source,and a model is necessary for its design,control and power management.A hybrid dynamic model of PEM fuel cell,which combines the advantages of mechanism model and black-box model,is proposed in this paper.To improve the performance,the static neural network and variable neural network are used to build the black-box model.The static neural network can significantly improve the static performance of the hybrid model,and the variable neural network makes the hybrid dynamic model predict the real PEM fuel cell behavior with required accuracy.Finally,the hybrid dynamic model is validated with a 500 W PEM fuel cell.The static and transient experiment results show that the hybrid dynamic model can predict the behavior of the fuel cell stack accurately and therefore can be effectively utilized in practical application.

A dynamical neural network approach for distributionally robust chance-constrained Markov decision process

In this paper,we study the distributionally robust joint chance-constrained Markov decision process.Utilizing the logarithmic transformation technique,we derive its deterministic reformulation with bi-convex terms under the moment-based uncertainty set.To cope with the non-convexity and improve the robustness of the solution,we propose a dynamical neural network approach to solve the reformulated optimization problem.Numerical results on a machine replacement problem demonstrate the efficiency of the proposed dynamical neural network approach when compared with the sequential convex approximation approach.

Adaptive Kalman filter and dynamic recurrent neural networks-based control design of macro-micro manipulator

In this paper, a composite control scheme for macro-micro dual-drive positioning stage with high accel- eration and high precision is proposed. The objective of control is to improve the precision by reducing the influence of system vibration and external noise. The positioning stage is composed of voice coil motor （VCM） as macro driver and piezoelectric actuator （PEA） as micro driver. The precision of the macro drive positioning stage is improved by the com- bined PID control with adaptive Kalman filter （AKF）. AKF is used to compensate VCM vibration （as the virtual noise） and the external noise. The control scheme of the micro drive positioning stage is presented as the integrated one with PID and intelligent adaptive inverse control approach to compensate the positioning error caused by macro drive positioning stage. A dynamic recurrent neural networks （DRNN） based inverse control approach is proposed to offset the hysteresis nonlinearity of PEA. Simulations show the positioning precision of macro-micro dual-drive stage is clearly improved via the proposed control scheme.

Hybrid Dynamic Neural Network and PID Control of Pneumatic Artificial Muscle Using the PSO Algorithm

Pneumatic artificial muscles(PAM)have been recently considered as a prominent challenge regarding pneumatic actuators specifically for rehabilitation and medical applications.Since accomplishing accurate control of the PAM is comparatively complicated due to time-varying behavior,elasticity and ambiguous characteristics,a high performance and efficient control approach should be adopted.Besides of the mentioned challenges,limited course length is another predicament with the PAM control.In this regard,this paper proposes a new hybrid dynamic neural network(DNN)and proportional integral derivative(PID)controller for the position of the PAM.In order to enhance the proficiency of the controller,the problem under study is designed in the form of an optimization trend.Considering the potential of particle swarm optimization,it has been applied to optimally tune the PID-DNN parameters.To verify the performance of the proposed controller,it has been implemented on a real-time system and compared to a conventional sliding mode controller.Simulation and experimental results show the effectiveness of the proposed controller in tracking the reference signals in the entire course of the PAM.

Hydraulic Optimization of a Double-channel Pump's Impeller Based on Multi-objective Genetic Algorithm

Computational fluid dynamics（CFD） can give a lot of potentially very useful information for hydraulic optimization design of pumps, however, it cannot directly state what kind of modification should be made to improve such hydrodynamic performance. In this paper, a more convenient and effective approach is proposed by combined using of CFD, multi-objective genetic algorithm（MOGA） and artificial neural networks（ANN） for a double-channel pump＇s impeller, with maximum head and efficiency set as optimization objectives, four key geometrical parameters including inlet diameter, outlet diameter, exit width and midline wrap angle chosen as optimization parameters. Firstly, a multi-fidelity fitness assignment system in which fitness of impellers serving as training and comparison samples for ANN is evaluated by CFD, meanwhile fitness of impellers generated by MOGA is evaluated by ANN, is established and dramatically reduces the computational expense. Then, a modified MOGA optimization process, in which selection is performed independently in two sub-populations according to two optimization objectives, crossover and mutation is performed afterword in the merged population, is developed to ensure the global optimal solution to be found. Finally, Pareto optimal frontier is found after 500 steps of iterations, and two optimal design schemes are chosen according to the design requirements. The preliminary and optimal design schemes are compared, and the comparing results show that hydraulic performances of both pumps 1 and 2 are improved, with the head and efficiency of pump 1 increased by 5.7% and 5.2%, respectively in the design working conditions, meanwhile shaft power decreased in all working conditions, the head and efficiency of pump 2 increased by 11.7% and 5.9%, respectively while shaft power increased by 5.5%. Inner flow field analyses also show that the backflow phenomenon significantly diminishes at the entrance of the optimal impellers 1 and 2, both the area of vortex and intensity of vortex decreases in the whole flow channel. This paper provides a promising tool to solve the hydraulic optimization problem of pumps＇ impellers.

Intelligent Multivariable Modeling of Blast Furnace Molten Iron Quality Based on Dynamic AGA-ANN and PCA

Blast furnace （BF） ironmaking process has complex and nonlinear dynamic characteristics. The molten iron temperature （MIT） as well as Si, P and S contents of molten iron is difficult to be directly measured online, and large-time delay exists in offline analysis through laboratory sampling. A nonlinear multivariate intelligent modeling method was proposed for molten iron quality （MIQ） based on principal component analysis （PCA） and dynamic ge- netic neural network. The modeling method used the practical data processed by PCA dimension reduction as inputs of the dynamic artificial neural network （ANN）. A dynamic feedback link was introduced to produce a dynamic neu- ral network on the basis of traditional back propagation ANN. The proposed model improved the dynamic adaptabili- ty of networks and solved the strong fluctuation and resistance problem in a nonlinear dynamic system. Moreover, a new hybrid training method was presented where adaptive genetic algorithms （AGA） and ANN were integrated, which could improve network convergence speed and avoid network into local minima. The proposed method made it easier for operators to understand the inside status of blast furnace and offered real-time and reliable feedback infor- mation for realizing close-loop control for MIQ. Industrial experiments were made through the proposed model based on data collected from a practical steel company. The accuracy could meet the requirements of actual operation.

Application of Neural Network in Prediction of Radionuclide Diffusion in Receiving Water

It needs long time to predict radioactive contaminant diffusion in receiving water under accident condition by using computational fluid dynamics （CFD） model. In order to shorten the computation time, a hybrid model based on CFD and time series neural network （TSNN） is proposed in this paper. The concentration change of radioactive contamination in an inland reservoir after a postulated accident is studied as a case. The result shows that this hybrid model can predict the contaminant diffusion trend and shorten at least 50% of iteration time. Priori knowledge integrated into the neural network model is able to reduce the mean square error of network output to 9.66×10 8 , which makes neural network output more close to the simulated contaminant concentration.

Model-free adaptive optimal design for trajectory tracking control of rocket-powered vehicle

An adaptive optimal trajectory tracking controller is presented for the Solid-RocketPowered Vehicle(SRPV)with uncertain nonlinear non-affine dynamics in the framework of adaptive dynamic programming.First,considering that the ascent model of the SRPV is non-affine,a model-free Single Network Adaptive Critic(SNAC)method is developed based on the dynamic neural network and the traditional SNAC method.This developed model-free SNAC method overcomes the limitation of the traditional SNAC method that can only be applied to affine systems.Then,a closed-form adaptive optimal controller is designed for the non-affine dynamics of SRPVs.This controller can adjust its parameters under different flight conditions and converge to the approximate optimal controller through online self-learning.Finally,the convergence to the approximate optimal controller is proved.The theoretical analysis of the uniformly ultimate boundedness of the tracking error is also presented.Simulation results demonstrate the effectiveness of the proposed controller.

Extracting governing system for the plastic deformation of metallic glasses using machine learning

This paper shows hidden information from the plastic deformation of metallic glasses using machine learning.Ni_(62)Nb_(38)(at.%)metallic glass(MG)film and Zr_(64.13)Cu_(15.75)Al_(10)Ni_(10.12)(at.%)BMG,as two model materials,are considered for nano-scratching and compression experiment,respectively.The interconnectedness among variables is probed using correlation analysis.The evolvement mechanism and governing system of plastic deformation are explored by combining dynamical neural networks and sparse identification.The governing system has the same basis function for different experiments,and the coefficient error is≤0.14%under repeated experiments,revealing the intrinsic quality in metallic glasses.Furthermore,the governing system is conducted based on the preceding result to predict the deformation behavior.This shows that the prediction agrees well with the real value for the deformation process.