Search

In this paper, we consider the problem of sampled-data active anti-disturbance control design for multi-rate control system subject to delayed sampled-data output. The plant considered in this paper is equipped with a slow-rate sensor and a fast-rate actuator. In the presence of non-vanishing disturbance, a new output predictor-based observer is first proposed by only using the delayed sampled-data output, so as to effectively estimate system state and disturbance in the inter-sample time interval. A new sampled-data controller that updates at a fast-rate is then designed with the help of active anti-disturbance technique. Thanks to prediction and active anti-disturbance techniques, the proposed sampled-data control method can effectively reject the influence of non-vanishing disturbance despite slow-rate sampling and delay of sensor. Finally, the effectiveness of the proposed control method is demonstrated via simulation results of a practical example of the visual servoing control for an inertially stabilized platform.

There are two challenging problems with emotional recognition using physiological signals. The first problem is to fuse different signals to improve the recognition accuracy, and the second problem is to reveal the importance of different signals for recognition. Data fusion methods have been used to solve the first problem in many studies. However, few methods consider the second problem. In this paper, a new method is proposed to solve both problems simultaneously, named multi-task multi-kernel logistic regression (MTMKLR). The proposed method regards different types of physiological signals as multiple kernels and fuses them by kernel weights which reflect the importance of signals. The classifications of low/high valence and low/high arousal are considered as multiple tasks. A publicly available dataset named DEAP is used to test the proposed method. Compared with conventional kernel logistic regression, the results show that the classification accuracy of MTMKLR is increased by more than 10%. Additionally, the proposed MTMKLR method achieves better classification performance than other methods in the case of using fewer signals and channels.

In this paper we discuss the impulsive containment control problem of continuous-time linear multi-agent systems. Firstly, an impulsive control protocol with self-feedback and aperiodic sampling is proposed, in which the coupling matrix do not have to be Laplacian. Utilizing the proposed protocol, the impulsive containment control problem for continuous-time linear multi-agent systems is equivalently transformed into the stabilization problem of some independent linear discrete-time switched systems. Then a mode-dependent switching-type Lyapunov function is constructed, some sufficient conditions are showed by using some linear matrix inequalities to achieve the containment control purpose. In addition, based on a linear matrix inequality, the controller gain matrix can be designed to solve the containment control problem. In obtained results, the self-feedback gain, the size and the switching frequency of the sampling intervals can play all important roles. Finally, two simulation examples are employed to validate the effectiveness of the theoretical results.

Adversarial examples cause the deep neural network prediction error, which is a great threat to the deep neural network. How to generate more natural adversarial examples and improve the robustness of deep neural networks has received attention. In this paper, we propose an improved blackbox attack (IBBA) algorithm based on query and perturbation distribution. This algorithm only needs the top-l label of the attacked model to generate the adversarial examples. Based on the existing black-box attacks, we optimize the performance of the algorithm from two aspects: query distribution and perturbation distribution. In the aspect of query distribution, we adopt different strategies for nontargeted attack and targeted attack; in the aspect of perturbation distribution, we choose different low-frequency noise according to the difference between the targeted attack and nontargeted attack. The experimental results on ImageNet show that the proposed algorithm is better than the existing algorithms in low query number, and the targeted attack is better in each specified query number.

Non-Rigid Structure from Motion (NRSfM) is the task of reconstructing the 3D point set of a non-rigid object from an ensemble of images with 2D correspondences, which has been a long-lasting challenging research topic. Compared to the state-of-the-art methods for NRSfM, the Procrustean Markov Process (PMP) model has obtained a relatively good performance. However, the estimation error and the convergence time of the PMP model will increase simultaneously when noise is present. To address this problem, in this paper, a coherent constraint is constructed to suppress the noise in the initialization step of the PMP algorithm. Moreover, an Accelerated Expectation Maximization (AEM) algorithm is devised to optimize the PMP estimation model. Experimental results on several widely used sequences demonstrate that our proposed algorithm achieves state-of-the-art performance, as well as its effectiveness and feasibility.

Recent years, convolutional neural network (CNN) has achieved excellent performance in computer vision, but recent researches show that they’re very vulnerable to adversarial examples, which results security risks to practical applications of CNN. Research on the generation algorithm of adversarial examples is conducive to evaluate the robustness of CNNs and promotes the study of defense algorithm. Although existing attack algorithms have achieved good performance in white-box manner, they behave unsatisfactorily in black-box manner. In this paper, we propose the mini-batch gradient based iterative method which could dramatically improve the attack ability in black-box manner. The previous iterative methods such as Iterative Fast Gradient Sign Method (I-FGSM) always make the generated adversarial examples fall into over-fitting, however we use minibatch gradient information to help them jump out of the local minimum. Moreover, the proposed method can also be considered as self-ensemble or implicit ensemble attack thus have a stronger attack power. The parallel randomization layer plays core role for providing the mini-batch gradient and the implement of it is very simple since there are no trainable variables. In order to further improve the attack success rate, we combined the momentum and the proposed method. We conducted several experiments on seven networks and the highest success rate of black-box attack has reached 94.1%, which is close to the whitebox manner. Since the proposed method greatly improve the transferability of adversarial examples, we think it could help evaluate the robustness of various deep networks.

In recent years, deep learning models have been widely adopted for transfer learning tasks. The deep models, however, were shown to be easily attacked by adversarial examples, which is generated from original samples with the carefully designed small perturbations. Thus, the transfer learning models based on deep networks will also face this problem. Because of the particularity of transfer learning tasks, using the conventional adversarial training to improve the robustness of deep transfer models is very difficult. In this paper, we propose a novel Double Transfer Learning with Adversarial Training (DTLAT) method to enhance the robustness of deep transfer learning models. Our intuition is using the instances of the source domain and target domain to help with adversarial training. At the same time, we design a reverse transfer learning method to weaken the effect of attack methods and improve the performance of deep transfer models of target domain. We regard the adversarial examples from some kind of attack method, like FGSM, as an adversarial domain while try to improve the generalization ability on other attacks method. Experiments demonstrate that DTLAT exceed many other methods about improving the robustness of the deep transfer model on several benchmark datasets.

Transfer learning aims to enhance performance in a target domain by exploiting useful information from related source domains. Transfer learning is important for many applications where the target domain instances are difficult to obtain while many data are available in source domains. Online transfer learning is established as an effective technology in many applications where the target data is received in an online manner. However, most existing methods ignore that target domain model does not have any prior knowledge when combined with source domain models. It is inappropriate that using conventional weight update rules to combine target domain model with source domain models. In this paper, we put forward a novel weight update rule for online transfer learning. Specifically, we prove the shortcomings of the previous methods and propose an effective improvement method which does not require a large amount of computation. Extensive experiments verify that our method can outperform several state-of-the-art methods on real-world data sets.

In this paper, a novel pinning controller, which is applied to the selected partial neurons with partial errors needed, is designed for global ultimate Mittag-Leffler lag quasi-synchronization (GUMLQS) of delayed fractional-order memristive neural networks (FMNNs) with switching jumps mismatch. Since the right-hand sides of the equations of FMNNs are discontinuous, FMNNs have no solution in the ordinary sense, and Filippov solutions are adopted for FMNNs. With the assistance of fractional Halanay inequality, a Lyapunov function in quadratic form and the theory of fractional derivative, an LMI-based GUMLQS criterion is derived. For a prescribed GUMLQS error bound, the control gain matrices can be calculated out via a group of LMIs. Meanwhile, the sufficient condition for the feasibility of the LMIs is obtained, which is an important basis for the selection of controlled neurons. Finally, a numerical example shows that the designed pinning controller can achieve GUMLQS between master-slave FMNNs with the prescribed ultimate error bound.

It has been demonstrated that biological whisker sensing system has excellent capacity to achieve the pattern recognition, and whisker sensor based robot tactile perception has caused widespread concern. However, most of current existed whisker sensors are too large in size, and are difficult to be integrated into a small robot. In this paper, we proposed a MEMS based biomimetic whisker sensor inspired by rats, which integrates four sensing units on a square silicon wafer with a side length of 6.8mm and makes it possible to be installed on a small robot with similar size to a real rat. Our sensor measure the bending moment at the root of the whisker based on the piezoresistive effect. We have explored the range of functions that whisker arrays can serve on and it is shown that our whisker sensor can determine obstacle distance, and is also capable of gathering surface information including shape and texture.

The present paper deal with the containment control problem for a class of continuous-time linear multi-agent systems. Using the neighborhood information of aperiodic sampled data, we propose a protocol based on aperiodic sampled data and with self-feedback. Utilizing the proposed protocol, the containment control problem is equivalently transformed into the stability problem for some complex-value linear discrete-time equations. By using the differential inequality technique, a very simple necessary and sufficient condition is obtained to ensure linear multi-agent systems achieves state containment, which gives a formula for calculating the maximum sampling interval.

This paper investigates a prescribed performance controller design method for DC converter system in DC Microgrid based on finite-time disturbance observer (FTDO) by means of nonsingular terminal sliding mode control (NTSMC) design procedure. As a main contribution, prescribed performance controller design manner is employed to realize the layout of specific requirements of system performance. Then, FTDO is developed to estimate disturbances quickly and improve the robustness of the system. In addition, considering dynamic response speed of system, NTSMC is utilized to render a finite-time tracking result with tunable parameters for the DC converter system. Finally, numerical simulation results are displayed to demonstrate the dynamic performance of the proposed controller.

Adaptive firing rate of neuron plays an indispensable role in stabilizing the neural system, which means that the firing rate of neuron could be adjusted adaptively within an inherent range. In this paper, two aspects of implementing the adaptive firing rate are proposed at the circuit level. First, a memristor model is used in the neuron circuit to represent membrane sensitivity. Second, the threshold voltage of neuron circuit can be adjusted adaptively to change the firing rate. Combined these two methods, the adaptive firing rate of neuron circuit is realized effectively, which is in accordance with its biological counterpart. Furthermore, the proposed neuron circuit is applied in the spiking neural network to verify its functionality, where pattern recognition could be realized. All the simulations are carried out on PSPICE.

Although the Capsule Network (CapsNet) has a better proven performance for the recognition of overlapping digits than Convolutional Neural Networks (CNNs), a large number of matrix-vector multiplications between lower-level and higher-level capsules impede efficient implementation of the CapsNet on conventional hardware platforms. Since three-dimensional (3-D) memristor crossbars provide a compact and parallel hardware implementation of neural networks, this paper provides an architecture design to accelerate convolutional and matrix operations of the CapsNet. By using 3-D memristor crossbars, the PrimaryCaps, DigitCaps, and convolutional layers of a CapsNet perform the matrix-vector multiplications in a highly parallel way. Simulations are conducted to recognize digits from the USPS database and to analyse the work efficiency of the proposed circuits. The proposed design provides a new approach to implement the CapsNet on memristor-based circuits.

For the design and operation of interim product manufacturing systems, large-scale one-of-a-kind production (OKP) features disruptions and lack of synchronization, which challenges real-time production process systematic control. The difficulty lies in two aspects. First, large-scale OKP is generally characterized by product design uncertainty and changes in resource availability, which often causes varied working time with workforce allocation variations. Second, specific types of large-scale OKP provides discrete production with complicated structures. There is the complicated work breakdown structure (WBS) of top-down and simple-to-complex refinement. Moreover, the production task relation structure in WBS, which consists of tandem structure, parallel structure and double-level-nested parallel structure, is complex. Indeed, many large-scale OKP enterprises do not have an effective method to both control and plan adjustment in a systematic way for dealing with the progress of OKP operations under workforce allocation disruptions and different working time scenarios. The traditional production project management and control system, theory, and methods do not handle this situation well since these technologies are developed with a view to managing production periods instead of controlling time-related cost systematically during production. Regarding the interim product production process in large-scale OKP as a manufacturing system, in this article we propose a cost dynamic control and optimization method based on the multilevel hierarchical PERT-Petri net (MLHPP) to achieve a closed-loop production dynamic control structure in large-scale OKP. The proposed modeling method and algorithms, through an industrial implementation in shipbuilding interim production (e.g., a ship block building), demonstrate a computing model for structuring and controlling large-scale OKP systems.

The imitation of classical conditioning reflex at circuit level is a significant procedure to achieve biology-like circuit systems. In order to realize the imitation, the important process are to imitate synaptic behaviors. As an emerging device, the memristor has some excellent properties like nonvolatility, adjustability, nanoscale. Therefore, it is an appropriate candidate to simulate the synaptic behavior in artificial neural network circuits. This paper presents a memristor-CMOS hybrid circuit to imitate classical conditioning reflex of aplysia californica. Besides, the proposed circuit can simulate additional forgetting stages that are activated by the single unconditional stimulus (US) or the single conditioned stimulus (CS) after the learning stage. The learning and forgetting stages can be described as: when US and CS are given to an aplysia californica together, it forms classical conditioning reflex through associative learning. Then, giving US or CS alone to the aplysia californica after the learning, it will forget the classical conditioning reflex gradually. The proposed circuit is simulated on PSPICE to demonstrate the effectiveness.

Differential-algebraic neural networks is established by using the current-voltage characteristic of the mem-ristor. The basic problem of this new neural networks are investigated, such as the existence and uniqueness problem and the stability problem. In addition, a numerical example and simulation are given to show the effectiveness of the our results. The approach and results of the paper have a leading significance in the design and study of memristive neural networks and differential-algebraic neural networks.

In this paper, a dynamic aperiodic sampling approach is proposed for nonlinear model reference control systems. An exogenous dynamic model is introduced to drive the aperiodic sampling sequences. The uniformly ultimate boundedness of the reference error is proved and the proposed design is shown to be effective in reducing the sampling times. The theoretical results are validated by simulations of a nonlinear example.

Flip-flops are basic units for all kinds of sequential logic circuits and complex digital electronics systems, which can be used to store binary data owing to their two stable logic states 0 and 1. SR flip-flop and D flip-flop are widely employed in digital circuits as representative types of basic flip-flops. In this paper, logic circuits of nonvolatile synchronous SR flip-flop and D flip-flop are proposed, which utilize nanoscale memristors as the nonvolatile memory elements to store the information with different conductance states. The resistive switching property of memristors is well-suited for implementing trigger function in flip-flop circuits. Compared with traditional flip-flop ones, the proposed memristor-based SR flip-flop and D flip-flop have nonvolatile characteristic, which is suitable for occasions with unstable power supply. Meanwhile, the designed circuits provide experimental references for the development of digital circuit structures based on the feasibility of tight integration of memristors with CMOS circuitry.

This brief proposes an improved XOR logic gate and a novel full adder. The presented XOR has simple structure which consists of 2 CMOS transistors and 4 memristors, and it reduces 4 CMOS transistors compared with the previous. In addition, the novel full adder consists of 7 CMOS transistors and 10 memristors. This novel full adder reduces 8 CMOS transistors and 2 memristors, which means the size of the circuit is smaller and power consumption is lower. Through analysis and simulation, the feasibilities of the improved XOR logic gate and novel full adder are demonstrated.

Accurate electric load forecasting is significant for the operation of the power systems and electricity markets. This paper proposes a particle swarm optimization with support vector machine (PSOSVM) to forecast annual power load. Based on radial basis function, support vector machine (SVM) is utilized to determine the structure and initial values of the parameters. Then, particle swarm optimization (PSO) is employed to optimize the parameters of the SVM model. In order to utilize the proposed method, practical data are divided into two parts, one is for training, the other is for testing. The combined method, PSOSVM, can effectively predict annual power load.

Memristor is a nonlinear resistor with the character of memory and is proved to be suitable for simulating synapse of neuron. This paper introduces two memristors in series with the same polarity (back-to-back) as simulator for neuron's synapse and presents the model of recurrent neural networks with such back-to-back memristors. By analysis techniques and fixed point theory, some sufficient conditions are obtained for recurrent neural network having single attractor flow and multiple attractors flow. At last, simulation with numeric examples is presented to illustrate our results.

Blood vessel segmentation is an important problem for quantitative structure analysis of retinal images, and many diseases are related to the structure changes. Manual segmentation is time consuming and computer aided segmentation is required to deal with large amount images. This paper presents a new supervised method for segmentation of blood vessels in retinal photographs. Multiple kernel learning (MKL) is introduced to deal with the problem, utilizing features from Hessian matrix based vesselness measure, response of multiscale Gabor filter, and multiple scale line strength features. The method is evaluated on the publicly available DRIVE and STARE databases. The performance of the MKL method is evaluated and experimental results show the high accuracy of the proposed method.

In this paper, a new intelligence paradigm scheme to forecast landslide based on functional networks is presented. Both methodology and learning algorithm for this kind of intelligence system paradigm using the minimax method are derived. The performance and validity of the new functional networks intelligence paradigm are demonstrated by using real-world example. The results show that the landslide prediction using functional networks is reasonable, effective and achieves a high-quality performance.

An efficient and accurate method for landslide displacement prediction is very important to reduce the casualties and property losses caused by this type of natural hazard. In recent years, many kinds of artificial neural networks (ANNs) have been widely applied to landslide displacement prediction. But we can't know which type of ANN is the best until we have calculated the prediction error. An improper choice of ANN may result in bad prediction results. In this paper, we use a neural networks combination prediction method based on the discounted MSFE (mean squared forecast error) to reduce the risk of selecting the types of ANNs. Four popular ANNs, radial basis function neural network (RBFNN), support vector regression (SVR), least squares support vector machine (LSSVM) and extreme learning machine (ELM), are selected as candidate neural networks. The performance of our model is verified through two case studies in Baishuihe landslide and Bazimen landslide. Experimental results reveal that the combining neural networks can improve the generalization abilities of ANNs.

The property of changing resistance according to applied currents of memristors makes them candidates for emulating synapses in artificial neural networks. In this paper, we introduce a memristive synapse design into neural network circuits. Combined with modified integrate-and-fire (I&F) complementary metal-oxide-semiconducter (CMOS) neurons, the memristive neural network shows similarities to its biological counterpart, in respect of biologically realistic, current-controlled spikes and adaptive synaptic plasticity. Then, the spike-rate-dependent plasticity (SRDP) of the synapse, an extended protocol of the Hebbian learning rule, is originally implemented by the circuit. And some advanced neural activities including learning, associative memory and forgetting are realized based on the SRDP rule. These activities are comprehensively validated on a neural network circuit inspired by famous Pavlov's dog-experiment with simulations and quantitative analyses.

In this paper, we analyze the character of memristor. A four-dimensional autonomous hyper chaotic circuit based on memristor is given. By using differential inclusion, linear matrix inequalities (LMIs) and Lyapunov stability theory, a sufficient condition for the global asymptotically stability of nonlinear memristor-based hyper chaotic circuit is obtained. Finally, a numerical example is given to illustrate the derived theoretical results.

In time series prediction tasks, dynamic models are less popular than static models, while they are more suitable for modeling the underlying dynamics of time series. In this paper, a novel architecture and supervised learning principle for recurrent neural networks, namely echo state networks, are adopted to build dynamic time series predictors. Ensemble techniques are employed to overcome the randomness and instability of echo state predictors, and a dynamic ensemble predictor is therefore established. The proposed predictor is tested in numerical experiments and different strategies for training the predictor are also comparatively studied. A case study is then conducted to test the predictor's performance in realistic prediction tasks.

The prediction of landslide displacement is essential for carrying out to improve the disaster warning system and reduce casualties and property losses. This study applies a novel neural network technique, extreme learning machine (ELM) with kernel function, to landslide displacement prediction problem. However, the generalization performance of ELM with kernel function depends closely on the kernel types and the kernel parameters. In this paper, we use a convex combination of Gaussian kernel function and polynomial kernel function in ELM, which may use these two types of kernel functions' advantages. In order to avoid blindness and inaccuracy in parameter selection, a novel hybrid optimization algorithm based on the combination of Particle Swarm Optimization (PSO) and Gravitational Search Algorithm (GSA) is used to optimize the regularization parameter C, the Gaussian kernel parameter γ, the polynomial kernel parameter q and the mixing weight coefficient η. The performance of our model is verified through two case studies in Baishuihe landslide and Yuhuangge landslide.

This paper discusses passivity of delayed neural networks with discontinuous activations which are unbounded. Using generalized Lyapunov approach, sufficient conditions for this problem are obtained based on differential inclusion theory, where the results are proposed in form of linear matrix inequality. In addition, a numerical example is given to show the effectiveness and rightness of the results.

The problem of stability of multiple equilibria is studied in this paper for two kinds of recurrent neural networks with time-varying delays and activation functions symmetrical with respect to the origin on the phase plane. Some sufficient conditions are obtained to ensure that two kinds of recurrent neural networks can have (2m + 1)n equilibrium points and (m + 1)n of them are locally exponentially stable. The derived conditions are valuable extensions to the existing results on stability of multiple equilibria for recurrent neural networks with time-varying delays in the literature.

In this paper, a general class of hybrid Lotka-Volterra recurrent neural networks with memristor characteristics is formulated and studied. Some sufficient conditions on nondivergence, global attractivity and complete stability of the network are obtained, respectively. These results can be applied to the memristive dynamic memories. The analysis in the paper employs results from the theory of differential equations with discontinuous right-hand side as introduced by Filippov. These theoretical analysis can characterize the fundamental electrical properties of memristor devices and provide convenience for applications. A numerical example is given to illustrate the theoretical findings via computer simulations.

This paper is concerned with the problem of passivity and passification for a class of singularly perturbed nonlinear systems (SPNS) via neural network. By constructing a proper functional as well as the linear matrix inequalities (LMIs) technique, some novel sufficient conditions are derived to make SPNS passive. The allowable perturbation bound ξ* can be determined via certain algebra inequalities, the proposed controller based on neural network will make SPNS passive for all ξ* ∈ (0, ξ*). Finally, a numerical example is given to illustrate the theoretical results.

The problem of exponential stability of multiple equilibria in recurrent neural networks with time-varying delays and concave-convex characteristics is addressed in this paper. The focus is placed upon derivation of some sufficient conditions under which an neural network of order n can have (2k + 2m − 1)n equilibrium points with (k + m)n of them having local exponential stability. The new results represent important extensions of the existing results on multistability of delayed recurrent neural networks.