E. Chiarantoni

A feature extraction unsupervised neural network for an environmental data set

Environmental data sets are characterized by a huge amount of heterogeneous data from external fields. As the number of measured points grows, a strategy is needed to select and efficiently analyze the useful information from the whole data set. One efficient way of obtaining the validation-compression of data sets is the adoption of a restricted set of features that describe, with an assigned accuracy a subset of the whole data set. One characteristic feature of the environmental data is time dependency: in the medium and long term they are not stationary data sets. The aim of this work is to propose a feature extraction technique based on a new model of an unsupervised neural network suitable to analyze this kind of data. The paper reports the results obtained utilizing the above extraction and analysis procedure on a real data set on chemical pollutants. It is shown that the proposed neural network is able to identify correctly human and/or meteorological effects in the environmental data set.

A new non competitive unsupervised neural network for clustering

In this paper a new unsupervised neural network, characterized by the absence of competition between its elements, is introduced. The kernel of this network is a new neural unit able to perform clustering even acting alone. It is shown how this network overcomes some of the major drawbacks of classical unsupervised competitive architectures.<<ETX>>

Centroid estimation by means of uncompetitive unsupervised neural element

In this paper the drawbacks of classical unsupervised learning laws are discussed and the paradigms of an alternative clustering algorithm are carried out. Then a new model of neuron element able to search the centroid of clusters without competition with other neurons, as in an unsupervised competitive learning law, is singled out.<<ETX>>

Multivariate data projection techniques based on a network of enhanced neural elements

In this paper two techniques to project high dimensional data into a bidimensional space are introduced. These techniques are based on an unsupervised neural network of enhanced processing elements. The proposed approaches are compared with some widely known projection techniques based on unsupervised neural networks. These comparisons show that the new projection techniques perform comparably or slightly better than the traditional techniques and are promising in term of computational burden.

An automatic method to detect missing components in manufactured products

In this paper we describe a method to recognize missing components on manufactured products. The proposed approach exploits the wavelet transform to extract features from the acquired data, while the diagnosis is performed by means of a neural network. The results show that this method achieves an high recognition rate. At the same time the method allows to use a very cheap diagnostic system.

Applying fixed point homotopy to nonlinear DAEs deriving from switching circuits

The time-domain analysis of switching circuits is a time consuming process as the change of switch state produces sharp discontinuities in switch variables. In this paper, a method for fast time-domain analysis of switching circuits is described. The proposed method is based on piecewise temporal transient analysis windows joined by DC analysis at the switching instants. The DC analysis is carried out by means of fixed point homotopy to join operating points between consecutive time windows. The proposed method guarantees accurate results reducing the number of iterations needed to simulate the circuit.

A new method for efficient time-domain simulation of power electronic circuits

In this paper a method for fast time-domain simulation of power electronic circuit with switches is described. The method is based on transient windows analysis joined by DC analysis on reduced circuits at the switching instants. It shows that the proposed approach provides the same results of standard analysis methods with a consistent reduction of the iterations required.

Transient Analysis of Three-Phase Non-linear Circuits by means of Homotopy Method

In this paper a class of homotopy functions for transient analysis of non-linear circuits is described. This class of homotopy is based on the hypothesis that the modified nodal analysis is used to solve the circuit. The method is employed for the simulation of a three-phase converter. The results show the reduction of the iteration number to trace the waveforms of the circuital variables preserving the accuracy of numerical results

Automatic Evaluation of Flaws in Pipes by means of Ultrasonic Waveforms and Neural Networks

The ultrasonic inspection technique takes a relevant place in not destructive defect detection. It can be very useful to determine the state of not accessible structure. In this paper a method based on ultrasonic waves inspection to evaluate the dimensions of flaws in not accessible pipes is shown. The method performs the extraction of time and frequency features from simulated ultrasonic waves and the proper reduction of the number of these features. Then a neural network classification evaluates the dimension of the flaws in the pipe under test. The results show low error rates for all classes considered.

Robust Unsupervised Competitive Neural Network by Local Competitive Signals

Unsupervised competitive neural networks have been recognized as a powerful tool for pattern analysis, feature extraction and clustering analysis. The global competitive structures tend to critically depend on the number of elements in the networks and on the noise property of the space. In order to overcome these problems in this work is presented an unsupervised competitive neural network characterized by units with an adaptive threshold and local inhibitory interactions among its cells. Each neural unit is based on a modified competitive learning law in which the threshold changes in learning stage. It is shown that the proposed neuron is able, during the learning stage, to perform an automatic selection of patterns that belong to a cluster, moving towards its centroid. The properties of this network, are examined in a set of simulations adopting a data set composed of Gaussian mixtures.