1,721,046 research outputs found
A pruning technique maximizing generalization
No full textThe paper proposes an estimator of the relevance of each weight connection on generalization which relies on a probabilistic description of the training phase: based on the above estimator, a pruning technique is derived
Learning vector quantization for the probabilistic neural network
No full textA modified version of the PNN (probabilistic neural network) learning phase which allows a considerable simplification of network structure by including a vector quantization of learning data is proposed. It can be useful if large training sets are available. The procedure has been successfully tested in two synthetic data experiments. The proposed network has been shown to improve the classification performance of the LVQ (learning vector quantization) procedur
Reduction of Distortions in Periodic Signals by Wiener Filtering
No full textMost of the techniques that are used to identify the I/O characteristic of a linear system require that it be stressed at the input through a periodic-type signal. The accuracy of the results is directly related to the degree of harmonic purity of the signal used as excitation. In this paper, a procedure is proposed for generating periodic signals whose degree of spectral purity is increased to extremely high levels using a procedure based on the Wiener filter. The procedure requires that the signal is generated by the usual procedure, and the subsequent filtering has such characteristics as to remove the harmonics that occurred and that decrease its spectral purity. The method does not act on the causes of generation of the additional harmonics, but removes their effects, restoring the signal to a very high degree of spectral purity. The proposed procedure is validated through examples to verify its effectiveness
Breaking the Excitation Chirp to Improve Ultrasonic Testing
No full textSignal-to-noise ratio and axial resolution both play a primary role in ultrasonic exploration: the former is directly related to the depth of signal penetration in the innermost part of the subject under test. The latter improves with an increase of the bandwidth of the signal: this implies, however, an increase in the amount of noise superimposed on the useful signal. A tradeoff choice must be made between axial resolution (signal bandwidth) and maximum penetration depth (signal-to-noise ratio). The use of coded signals and appropriate signal processing techniques, as the pulse compression, overcomes this conflicting choice. Such use, although beneficial for the above aspects, implies that signals of relatively high duration are placed at the input to the measurement system, and this leads to an additional possible problem, related to overheating of both the object being measured and the transducer itself. The present work addresses this aspect concerning the use of coded signals and related processing techniques. By exploiting the linearity of the pulse compression procedure, a novel technique is proposed to deal with the problem of probe temperature rise in the case where very high penetration capacity and resolution are to be obtained while respecting the constraints imposed on temperature rise. The technique consists in fragmenting the Chirp signal into temporal parts, and then recomposing the signal through an appropriate processing technique. The simulation results reported show the validity of the proposed approach
Relevance of accurate filter design in Hammerstein model identification algorithms of nonlinear systems
No full textThe Hammerstein model proved to be very effective in the representation of nonlinear systems. The identification of its kernels is possible through a reliable and computationally light procedure which can be further improved through the use of low- pass filters: this latter technique has been proposed based on the use of FIR filters, which, in the application in object, need a number of coefficients in the order of one thousand. In the present paper we verify the possibility to implement lowpass filtering using IIR structures, which are characterized by a drastically lower number of parameters. In the paper we consider in particular lowpass filters derived from analogue prototypes of Butterworth and Chebyshev types, and we compare their results with those obtainable through the use of FIR filters. The results presented here, verified in an experimental situation, provide the designer of the processing system with useful tools for an optimization of the Hammerstein model identification system
Ceramic powder characterization by multilayer perceptron (MLP) data compression and classification
No full textA neural network approach for pattern classification has been explored in the present paper as part of the recent resurgence of interest in this area. Our research has focused on how a multilayer feedforward structure performs in the particular problem of particle characterization. The proposed procedure, after suitable data preprocessing, consists of two distinct phases: in the former, a feedforward neural network is used to obtain an image data compression. In the latter, a neural classifier is trained on the compressed data. All the tests have been conducted on a sample constituted by two different typologies of ceramic particles, each characterized by a different microstructure. The sample image of different particles acquired and directly digitalized by scanning electron microscopy has been processed in order to achieve the best conditions to obtain the boundary profile of each particle. The boundary is thus assumed to be representative of the morphological characteristics of the ceramic products. Using the neural approach, a classification accuracy as high as 100% on a training set of 80 sub-images was achieved. These networks correctly classified up to 96.9% of 64 testing patterns not contained in the training set.
Novel Nueral Network Feature Selection Procedure by Generalization Maximization with Application to Automatic Robot Guidance
No full text- …
