Mohammed R. Sayeh

Neural networks for smart structures with fiber optic sensors

This paper discusses an application of artificial neural networks in smart structures with fiber optic sensors. Emphasis is on using a novel neural network approach to characterize the Impact signals from the embedded sensors. A recurrent network is proposed for control of stress in the composite material. Preliminary results on a composite with embedded multimode fiber indicates that the speckle pattern of the optical fiber can be characterized for stress measurements. This new sensing technique will simplify the measurement components as compared to the conventional methods. 2.

Neural networks for planar shape classification

A neural network approach is presented for the classification of closed planar shapes. The neural net classifier developed is robust and invariant to translation, rotation, and scaling. The primary foci are the development of an effective representation for planar shapes and the selection of a suitable neural network structure. In particular, planar shapes are represented by an ordered sequence that represents the Euclidean distance between the centroid and all contour pixels of the shape. It is also shown that for this classification problem and the representation derived, the three-layer perceptron with backpropagation training is an appropriate neural network configuration.<<ETX>>

Vibration control of flexible structures using self-sensing actuators

In this paper, the design and implementation of smart actuators for active vibration control of mechanical systems are considered. The proposed smart actuator is composed of one or several layers of piezoelectric materials that works both as a sensor and an actuator, in vibration control applications. An adaptive technique is developed for estimating the unknown equivalent capacitance of the piezoelectric material, which would be used for separating the effect of actuation from the measured (sensed) signal due to the strain in the material. This algorithm can be implemented in real time on a digital signal processor (DSP), allowing for the development of a DSP-based adaptive self-sensing actuator. This self-sensing actuator is then used in the vibration control of flexible structures. The vibration control system includes a power electronic amplifier, a data acquisition system, and a DSP for digital control implementation. A simple PID control strategy is employed for vibration reduction and motion control of cantilever beams using the proposed self-sensing actuators. Simulations and preliminary experiments show good results.

Imposing a temporal structure in neural networks

Although neural networks are very effective pattern classifiers a major limitation is that they are not suitable for classifying patterns with Inherent time-variations. This paper describes an approach to incorporate a temporal structure in a neural network system which wifi accomodate the time variations in local feature sets encountered in problems such as partial shape classification. 1.

Fiber-optic sensor based system to estimate stress in smart structures

This paper describes the development of an approach to estimate the applied stress sensed from a set of multimode fiber optic sensors which are laid on the surface of a smart structure. The estimation of the applied stress is based upon the discrimination between speckle patterns produced by different strain signals. Three approaches have been formulated to estimate/classify the applied stress from the speckle patterns: (a) neural network estimation, (b) Markov random field model classification, and (c) signature-based classification. In order to develop the neural network estimator which is trained to output an estimate of the applied strain signal vector, the dimension of the original input speckle vector is first reduced by estimating the entropy of each pixel and selecting the set of pixels which carry the most information in the training set. A statistical based clustering approach is formulated to reduce the dimension further by combining highly correlated pixels in the selected set. In the Markov random field model based approach, a Markovian model for texture is assumed to fit the speckle patterns. The model parameters, as estimated using maximum likelihood techniques, are used in conjunction with a nearest neighbor rule to classify the speckle images. The signature-based classification approach is a method which incorporates both dimensionality reduction and classification directly for the case when the reference speckle images from highly representative strain vectors are available.

Multiplexing of multimode fiber optic strain sensors using an artificial neural network

This paper introduces a relatively simple multimode fiber optic sensor built for on-line use and a package that multiplexes two fiber sensors. The multiplexing is achieved by the random nature of the multimode fiber output intensity variation (so called speckle pattern). The demultiplexing is performed by a neural network. The dynamic range of each sensor is 0.8- 120 micrometers over an effective length of 11 m. The sensors operate in 68-94 degrees F.

Image texture segmentation using a neural network

In this paper we use a neural network called the Lyapunov associative memory (LYAM) system to segment image texture into different categories or clusters. The LYAM system is constructed by a set of ordinary differential equations which are simulated on a digital computer. The clustering can be achieved by using a single tuning parameter in the simplest model. Pattern classes are represented by the stable equilibrium states of the system. Design of the system is based on synthesizing two local energy functions, namely, the learning and recall energy functions. Before the implementation of the segmentation process, a Gauss-Markov random field (GMRF) model is applied to the raw image. This application suitably reduces the image data and prepares the texture information for the neural network process. We give a simple image example illustrating the capability of the technique. The GMRF-generated features are also used for a clustering, based on the Euclidean distance.

Accomodating Temporal Variations in Neural Networks

Neural networks are very effective pattern classifiers, however, a major limitation is that they are unsuitable for classifying patterns with inherent time-variations. This paper describes an approach to incorporate a temporal structure in a neural network system which will accomodate the time variations in local feature sets encountered in problems such as partial shape classification.

Study of LiNbO3 in optical associative memory

Lithium niobate crystal has been used for storing information that can be accessed through an associative recall. This fact demonstrates the application of the crystal to the optical memory. specifically In an optical ring resonator. One important factor that determines the recall ability is the diffraction efficiency. Particularly this quantity Is also a function of the polarization of the readout beam. The diffraction efficiency of the crystal Is significant If the sufficient build-up of power in the real-time programmable ring resonator is required. To this end this paper is concentrated on the study of the dependency of diffraction efficiency on the polarization of recall beam. Experimental results are also given to indicate the significance of orientation of c-axis of the crystal in the experimental configuration. 2.

Class of learning algorithms for multilayer perceptron

A class of learning techniques for neural networks can be considered as optimization problems. The connection strengths are modified such that the difference between the network response and a desired response Is minimized. In this paper the learning techniques based on the gradient momentum Newton and quasi-Newton methods are considered. A learning algorithm is also developed based on the conjugate gradient technique. These learning techniques are applied to the Exclusive-OR problem for comparison of their performance. For this problem the algorithm based on the conjugate gradient technique converges faster than the other algorithms. 2.