Lyapunov Theory-Based Fusion Neural Networks for the Identification of Dynamic Nonlinear Systems

This paper introduces a novel fusion neural architecture and the use of a novel Lyapunov theory-based algorithm, for the online approximation of the dynamics of nonlinear systems. The proposed neural system, in combination with the proposed update rule of the neural weights, achieves fast convergence of the identification process, ensuring at the same time stability of the error system in the sense of Lyapunov theory. The fusion neural system combines the features that are extracted from two-independent neural streams, a feedforward and a diagonal recurrent one, satisfying different design criteria of the identification task. Simulation results for five cases reveal the approximation strength of both proposed fusion neural architecture and proposed learning algorithm. Also, additional experiments demonstrate the effectiveness in cases of parameter variations and additive noise.

THE SUPER-TURING COMPUTATIONAL POWER OF PLASTIC RECURRENT NEURAL NETWORKS

We study the computational capabilities of a biologically inspired neural model where the synaptic weights, the connectivity pattern, and the number of neurons can evolve over time rather than stay static. Our study focuses on the mere concept of plasticity of the model so that the nature of the updates is assumed to be not constrained. In this context, we show that the so-called plastic recurrent neural networks (RNNs) are capable of the precise super-Turing computational power — as the static analog neural networks — irrespective of whether their synaptic weights are modeled by rational or real numbers, and moreover, irrespective of whether their patterns of plasticity are restricted to bi-valued updates or expressed by any other more general form of updating. Consequently, the incorporation of only bi-valued plastic capabilities in a basic model of RNNs suffices to break the Turing barrier and achieve the super-Turing level of computation. The consideration of more general mechanisms of architectural plasticity or of real synaptic weights does not further increase the capabilities of the networks. These results support the claim that the general mechanism of plasticity is crucially involved in the computational and dynamical capabilities of biological neural networks. They further show that the super-Turing level of computation reflects in a suitable way the capabilities of brain-like models of computation.

INDIRECT ADAPTIVE CONTROL OF NONLINEAR SYSTEMS BASED ON BILINEAR NEURO-FUZZY APPROXIMATION

In this paper, we investigate the indirect adaptive regulation problem of unknown affine in the control nonlinear systems. The proposed approach consists of choosing an appropriate system approximation model and a proper control law, which will regulate the system under the certainty equivalence principle. The main difference from other relevant works of the literature lies in the proposal of a potent approximation model that is bilinear with respect to the tunable parameters. To deploy the bilinear model, the components of the nonlinear plant are initially approximated by Fuzzy subsystems. Then, using appropriately defined fuzzy rule indicator functions, the initial dynamical fuzzy system is translated to a dynamical neuro-fuzzy model, where the indicator functions are replaced by High Order Neural Networks (HONNS), trained by sampled system data. The fuzzy output partitions of the initial fuzzy components are also estimated based on sampled data. This way, the parameters to be estimated are the weights of the HONNs and the centers of the output partitions, both arranged in matrices of appropriate dimensions and leading to a matrix to matrix bilinear parametric model. Based on the bilinear parametric model and the design of appropriate control law we use a Lyapunov stability analysis to obtain parameter adaptation laws and to regulate the states of the system. The weight updating laws guarantee that both the identification error and the system states reach zero exponentially fast, while keeping all signals in the closed loop bounded. Moreover, introducing a method of "concurrent" parameter hopping, the updating laws are modified so that the existence of the control signal is always assured. The main characteristic of the proposed approach is that the a priori experts information required by the identification scheme is extremely low, limited to the knowledge of the signs of the centers of the fuzzy output partitions. Therefore, the proposed scheme is not vulnerable to initial design assumptions. Simulations on selected examples of well-known benchmarks illustrate the potency of the method.

APPLICATION OF HIGHER ORDER CUMULANT FEATURES FOR CARDIAC HEALTH DIAGNOSIS USING ECG SIGNALS

Electrocardiogram (ECG) is the electrical activity of the heart indicated by P, Q-R-S and T wave. The minute changes in the amplitude and duration of ECG depicts a particular type of cardiac abnormality. It is very difficult to decipher the hidden information present in this nonlinear and nonstationary signal. An automatic diagnostic system that characterizes cardiac activities in ECG signals would provide more insight into these phenomena thereby revealing important clinical information. Various methods have been proposed to detect cardiac abnormalities in ECG recordings. Application of higher order spectra (HOS) features is a seemingly promising approach because it can capture the nonlinear and dynamic nature of the ECG signals. In this paper, we have automatically classified five types of beats using HOS features (higher order cumulants) using two different approaches. The five types of ECG beats are normal (N), right bundle branch block (RBBB), left bundle branch block (LBBB), atrial premature contraction (APC) and ventricular premature contraction (VPC). In the first approach, cumulant features of segmented ECG signal were used for classification; whereas in the second approach cumulants of discrete wavelet transform (DWT) coefficients were used as features for classifiers. In both approaches, the cumulant features were subjected to data reduction using principal component analysis (PCA) and classified using three layer feed-forward neural network (NN) and least square — support vector machine (LS-SVM) classifiers. In this study, we obtained the highest average accuracy of 94.52%, sensitivity of 98.61% and specificity of 98.41% using first approach with NN classifier. The developed system is ready clinically to run on large datasets.

SINGLE-TRIAL MOTOR IMAGERY CLASSIFICATION USING ASYMMETRY RATIO, PHASE RELATION, WAVELET-BASED FRACTAL, AND THEIR SELECTED COMBINATION

An electroencephalogram (EEG) analysis system is proposed for single-trial classification of motor imagery (MI) data in this study. Applying event-related brain potential (ERP) data acquired from the sensorimotor cortices, the system mainly consists of enhanced active segment selection, feature extraction, feature selection and classification. In addition to the original use of continuous wavelet transform (CWT) and Student's two-sample t-statistics, the 2D anisotropic Gaussian filter is proposed to further refine the selection of active segments. We then extract several features, including spectral power and asymmetry ratio, coherence and phase-locking value, and multiresolution fractal feature vector, for subsequent classification. Next, genetic algorithm (GA) is used to select features from the combination of above-mentioned features. Finally, support vector machine (SVM) is used for classification. Compared with "without enhanced active segment selection," several potential features and linear discriminant analysis (LDA) on MI data from two data sets for 10 subjects, the results indicate that the proposed method achieves 86.7% average classification accuracy, which is promising in BCI applications.

HUMAN ACTION RECOGNITION USING META-COGNITIVE NEURO-FUZZY INFERENCE SYSTEM

We propose a sequential Meta-Cognitive learning algorithm for Neuro-Fuzzy Inference System (McFIS) to efficiently recognize human actions from video sequence. Optical flow information between two consecutive image planes can represent actions hierarchically from local pixel level to global object level, and hence are used to describe the human action in McFIS classifier. McFIS classifier and its sequential learning algorithm is developed based on the principles of self-regulation observed in human meta-cognition. McFIS decides on what-to-learn, when-to-learn and how-to-learn based on the knowledge stored in the classifier and the information contained in the new training samples. The sequential learning algorithm of McFIS is controlled and monitored by the meta-cognitive components which uses class-specific, knowledge based criteria along with self-regulatory thresholds to decide on one of the following strategies: (i) Sample deletion (ii) Sample learning and (iii) Sample reserve. Performance of proposed McFIS based human action recognition system is evaluated using benchmark Weizmann and KTH video sequences. The simulation results are compared with well known SVM classifier and also with state-of-the-art action recognition results reported in the literature. The results clearly indicates McFIS action recognition system achieves better performances with minimal computational effort.