Search results

Independent component analysis (ICA) is a widely-used blind source separation technique. ICA has been applied to many applications. ICA is usually utilized as a black box, without understanding its internal details. Therefore, in this paper, the basics of ICA are provided to show how it works to serve as a comprehensive source for researchers who are interested in this field. This paper starts by introducing the definition and underlying principles of ICA. Additionally, different numerical examples in a step-by-step approach are demonstrated to explain the preprocessing steps of ICA and the mixing and unmixing processes in ICA. Moreover, different ICA algorithms, challenges, and applications are presented.

This paper seeks to present a novel computational intelligence technique to generate concise neural network models for distributed dynamic systems.

The approach used in this paper is based on artificial neural network architectures that incorporate linear and nonlinear principal component analysis, combined with generalized dimensional analysis.

Neural network principal component analysis coupled with generalized dimensional analysis reduces input variable space by about 90 percent in the modeling of oil reservoirs. Once trained, the computation time is negligible and orders of magnitude faster than any traditional discretisation schemes such as fine‐mesh finite difference.

Finding the minimum number of input independent variables needed to characterize a system helps in extracting general rules about its behavior, and allows for quick setting of design guidelines, and particularly when evaluating changes in the physical properties of systems.

The methodology can be used to simulate dynamical systems characterized by differential equations, in an interactive CAD and optimization providing faster on‐line solutions and speeding up design guidelines.

In Part I, an intelligent acoustic emission (AE) locator is described while the Part II discusses a blind source separation, time delay estimation and location of two continuous AE sources. AE analysis is used for characterization and location of developing defects in materials. AE sources often generate a mixture of various statistically independent signals. A difficult problem of AE analysis is a separation and characterization of signal components when the signals from various sources and the mode of mixing are unknown. Recently, blind source separation (BSS) by independent component analysis (ICA) has been used to solve these problems. The purpose of this paper is to demonstrate the applicability of ICA to locate two independent simultaneously active AE sources on an aluminum band specimen. The method is promising for non‐destructive testing of aircraft frame structures by AE analysis.

Acoustic emission analysis (AE) is used for characterization and location of developing defects in materials. AE sources often generate a mixture of various statistically independent signals. One difficult problem of AE analysis is the separation and characterization of signal components when the signals from various sources and the way in which the signals were mixed are unknown. Recently, blind source separation (BSS) by independent component analysis (ICA) has been used to solve these problems. The main purpose of this paper is to demonstrate the applicability of ICA to time‐delay estimation of two independent continuous AE sources on an aluminum beam. It is shown that it is possible to estimate time delays by ICA, and thus to locate two independent simultaneously emitted sources.

To develop a new method for estimation of damage configuration in composite laminate structure using acoustic wave propagation signal and a reduction‐prediction neural network to deal with high dimensional spectral data.

A reduction‐prediction network, which is a combination of an independent component analysis (ICA) and a multi‐layer perceptron (MLP) neural network, is proposed to quantify the damage state related to transverse matrix cracking in composite laminates using acoustic wave propagation model. Given the Fourier spectral response of the damaged structure under frequency band‐selective excitation, the problem is posed as a parameter estimation problem. The parameters are the stiffness degradation factors, location and approximate size of the stiffness‐degraded zone. A micro‐mechanics model based on damage evolution criteria is incorporated in a spectral finite element model (SFEM) for beam type structure to study the effect of transverse matrix crack density on the acoustic wave response. Spectral data generated by using this model is used in training and testing the network. The ICA network called as the reduction network, reduces the dimensionality of the broad‐band spectral data for training and testing and sends its output as input to the MLP network. The MLP network, in turn, predicts the damage parameters.

Numerical demonstration shows that the developed network can efficiently handle high dimensional spectral data and estimate the damage state, damage location and size accurately.

Only numerical validation based on a damage model is reported in absence of experimental data. Uncertainties during actual online health monitoring may produce errors in the network output. Fault‐tolerance issues are not attempted. The method needs to be tested using measured spectral data using multiple sensors and wide variety of damages.

The developed network and estimation methodology can be employed in practical structural monitoring system, such as for monitoring critical composite structure components in aircrafts, spacecrafts and marine vehicles.

A new method is reported in the paper, which employs the previous works of the authors on SFEM and neural network. The paper addresses the important problem of high data dimensionality, which is of significant importance from practical engineering application viewpoint.

New approaches for non‐classical neural‐based computing are introduced. The developed approaches utilize new concepts in three‐dimensionality, invertibility and reversibility to perform the required neural computing. The various implementations of the new neural circuits using the introduced paradigms and architectures are presented, several applications are shown, and the extension for the utilization in neural‐systolic computing is also introduced.

The new neural paradigms utilize new findings in computational intelligence and advanced logic synthesis to perform the functionality of the basic neural network (NN). This includes the techniques of three‐dimensionality, invertibility and reversibility. The extension of implementation to neural‐systolic computing using the introduced reversible neural‐systolic architecture is also presented.

Novel NN paradigms are introduced in this paper. New 3D paradigm of NL circuits called three‐dimensional inverted neural logic (3DINL) circuits is introduced. The new 3D architecture inverts the inputs and weights in the standard neural architecture: inputs become bases on internal interconnects, and weights become leaves of the network. New reversible neural network (RevNN) architecture is also introduced, and a RevNN paradigm using supervised learning is presented. The applications of RevNN to multiple‐output feedforward discrete plant control and to reversible neural‐systolic computing are also shown. Reversible neural paradigm that includes reversible neural architecture utilizing the extended mapping technique with an application to the reversible solution of the maze problem using the reversible counterpropagation NN is introduced, and new neural paradigm of reversibility in both architecture and training using reversibility in independent component analysis is also presented.

Since the new 3D NNs can be useful as a possible optimal design choice for compacting a learning (trainable) circuit in 3D space, and because reversibility is essential in the minimal‐power computing as the reduction of power consumption is a main requirement for the circuit synthesis of several emerging technologies, the introduced methods for non‐classical neural computation are new and interesting for the design of several future technologies that require optimal design specifications such as three‐dimensionality, regularity, super‐high speed, minimum power consumption and minimum size such as in low‐power control, adiabatic signal processing, quantum computing, and nanotechnology.

Depression is a mental health problem characterized by a persistent sense of sadness and loss of interest. EEG signals are regarded as the most appropriate instruments for diagnosing depression because they reflect the operating status of the human brain. The purpose of this study is the early detection of depression among people using EEG signals.

(i) Artifacts are removed by filtering and linear and non-linear features are extracted; (ii) feature scaling is done using a standard scalar while principal component analysis (PCA) is used for feature reduction; (iii) the linear, non-linear and combination of both (only for those whose accuracy is highest) are taken for further analysis where some ML and DL classifiers are applied for the classification of depression; and (iv) in this study, total 15 distinct ML and DL methods, including KNN, SVM, bagging SVM, RF, GB, Extreme Gradient Boosting, MNB, Adaboost, Bagging RF, BootAgg, Gaussian NB, RNN, 1DCNN, RBFNN and LSTM, that have been effectively utilized as classifiers to handle a variety of real-world issues.

1. Among all, alpha, alpha asymmetry, gamma and gamma asymmetry give the best results in linear features, while RWE, DFA, CD and AE give the best results in non-linear feature. 2. In the linear features, gamma and alpha asymmetry have given 99.98% accuracy for Bagging RF, while gamma asymmetry has given 99.98% accuracy for BootAgg. 3. For non-linear features, it has been shown 99.84% of accuracy for RWE and DFA in RF, 99.97% accuracy for DFA in XGBoost and 99.94% accuracy for RWE in BootAgg. 4. By using DL, in linear features, gamma asymmetry has given more than 96% accuracy in RNN and 91% accuracy in LSTM and for non-linear features, 89% accuracy has been achieved for CD and AE in LSTM. 5. By combining linear and non-linear features, the highest accuracy was achieved in Bagging RF (98.50%) gamma asymmetry + RWE. In DL, Alpha + RWE, Gamma asymmetry + CD and gamma asymmetry + RWE have achieved 98% accuracy in LSTM.

A novel dataset was collected from the Central Institute of Psychiatry (CIP), Ranchi which was recorded using a 128-channels whereas major previous studies used fewer channels; the details of the study participants are summarized and a model is developed for statistical analysis using N-way ANOVA; artifacts are removed by high and low pass filtering of epoch data followed by re-referencing and independent component analysis for noise removal; linear features, namely, band power and interhemispheric asymmetry and non-linear features, namely, relative wavelet energy, wavelet entropy, Approximate entropy, sample entropy, detrended fluctuation analysis and correlation dimension are extracted; this model utilizes Epoch (213,072) for 5 s EEG data, which allows the model to train for longer, thereby increasing the efficiency of classifiers. Features scaling is done using a standard scalar rather than normalization because it helps increase the accuracy of the models (especially for deep learning algorithms) while PCA is used for feature reduction; the linear, non-linear and combination of both features are taken for extensive analysis in conjunction with ML and DL classifiers for the classification of depression. The combination of linear and non-linear features (only for those whose accuracy is highest) is used for the best detection results.

We present an approach for blind separation of acoustic sources produced from multiple speakers mixed in realistic room environments. We first transform recorded signals into the time‐frequency domain to make mixing become instantaneous. We then separate the sources in each frequency bin based on an independent component analysis (ICA) algorithm. For the present paper, we choose the complex version of fixedpoint iteration (CFPI), i.e. the complex version of FastICA, as the algorithm. From the separated signals in the time‐frequency domain, we reconstruct output‐separated signals in the time domain. To solve the so‐called permutation problem due to the indeterminacy of permutation in the standard ICA, we propose a method that applies a special property of the CFPI cost function. Generally, the cost function has several optimal points that correspond to the different permutations of the outputs. These optimal points are isolated by some non‐optimal regions of the cost function. In different but neighboring bins, optimal points with the same permutation are at almost the same position in the space of separation parameters. Based on this property, if an initial separation matrix for a learning process in a frequency bin is chosen equal to the final separation matrix of the learning process in the neighboring frequency bin, the learning process automatically leads us to separated signals with the same permutation as that of the neighbor frequency bin. In each bin, but except the starting one, by chosen the initial separation matrix in such a way, the permutation problem in the time domain reconstruction can be avoided. We present the results of some simulations and experiments on both artificially synthesized speech data and real‐world speech data, which show the effectiveness of our approach.

We study the informational content of factor structures in discrete triangular systems. Factor structures have been employed in a variety of settings in cross-sectional and panel data models, and in this chapter we formally quantify their identifying power in a bivariate system often employed in the treatment effects literature. Our main findings are that imposing a factor structure yields point-identification of parameters of interest, such as the coefficient associated with the endogenous regressor in the outcome equation, under weaker assumptions than usually required in these models. In particular, we show that a “non-standard” exclusion restriction that requires an explanatory variable in the outcome equation to be excluded from the treatment equation is no longer necessary for identification, even in cases where all of the regressors from the outcome equation are discrete. We also establish identification of the coefficient of the endogenous regressor in models with more general factor structures, in situations where one has access to at least two continuous measurements of the common factor.