M. Bodruzzaman

LabVIEW program design for on-line data acquisition and predictive maintenance

The online design for data acquisition and predictive maintenance for a fan-motor system using the graphical programming language, LabVIEW is presented. The data set were created for different faults under varying rpm levels from the synthetically generated fault data and the normal base line data acquired from the fan-motor system. The data were processed and the extracted features were fed to a two layer backpropagation neural network. The design is to be implemented on an online basis.

Feature extraction using wavelet transform for neural network based image classification

In order to design an image classification or recognition scheme which should have a robustness in classification approaching as close as possible to that of the human biological recognition system, two factors must be taken into account: it must be able to automatically extract global properties of the images; and it must be able to filter out the variations such as scaling and rotation in the images. Wavelet transforms of the images with high frequency components truncated off seem to be able to meet both of these conditions. This is because low frequency components are spread in the time domain and can be treated as global property while high frequency components, concentrated in time domain, can be discarded. Information at different resolution scales provided by wavelet features lead to highly discriminating, robust classifiers. Wavelets can examine data at different scales and frequencies. The theory behind the wavelets and their suitability for classification is discussed. The authors discuss extraction and how the wavelet transform is implemented. Finally, results of feature extraction are given.

A neural network-based text independent voice recognition system

A text-independent voice recognition experiment was conducted using an artificial neural network. The speech data were collected from three different speakers uttering thirteen different words. Each word was repeated ten times. The speech data were then pre-processed for signal conditioning. A total of 12 feature parameters were obtained from Cepstral coefficients via a linear predictive coding (LPC). These feature parameters then served as inputs to the neural network for speaker classification. A standard two-layer feedforward neural network was trained to identify different feature sets associated with the corresponding speakers. The network was tested for the remaining unseen words in text-independent mode. The results were very promising with a voice recognition accuracy of more than 90%. The success rate could be increased by adding more utterances from each speaker.<<ETX>>

Classification of electromyographic signals by autoregressive modeling

The classification of a set of intramuscular electromyographic (EMG) signals collected from normal, neuropathic, and myopathic patient groups is discussed. The signal is recorded in real time for 2 or 3 s, during which the patient performs a continuous ramp contraction. The time-varying dynamic nature of the neuromuscular system was observed by autoregressive (AR) modeling of the running windowed data segments. The 0.2-s length window runs along the entire data length of 1.6 seconds. The time varying nature of the model coefficients and the prediction error variance was investigated. The prediction error variance parameter is found to have significant time-varying characteristics. A first-order regression model is used to quantify the trend of this parameter. The probability density functions are estimated for the regression model parameters, and results of classifications for various pathologic classes are presented.<<ETX>>

Chaotic classification of electromyographic (EMG) signals via correlation dimension measurement

A set of intramuscular electromyographic signals were collected from various patient groups during ramp muscle contraction. The signals were collected using a real-time data acquisition system. The signals were tested for their chaotic behavior using spectral analysis and Poincare map techniques. MATLAB based software tools were developed to compute and plot the correlation function for each data set to determine the time lag for the first zero crossing. This time lag was used to create the signals state-space model. A two-dimensional state-space was created and plotted one state versus the other to observe the Poincare map. A correlation integral was computed for each state-space data set, and the correlation dimension values were then calculated by differentiating these correlation integral signals for each data set. The correlation dimension values were found to be different for different patient groups. The results show promise for online classification of neuromuscular patient groups.<<ETX>>

Sensory fusion for intelligent navigation of mobile robot

We develop techniques to fuse a robots sonar and wheel encoder information to produce a map. The sonar(s) gives the distance of the closest object(s) to the robot. The wheel encoder gives the current position of the robot from a starting reference position. Because of the physical limitations on the sonar, including poor angular resolution and poor accuracy, one cannot use sonar information directly for localization. The problem with angular resolution could be overcome by taking the reading at multiple view points. The accuracy problem could be reduced by integrating the sonar values over time. Another problem is specular reflections. A specular environment causes the pulses from the sonar to not return to the emitter, so that the object will be invisible to the sonar. A Bayesian probabilistic map is used in conjunction with a pulse coupled neural network icon of the image to locate the place. The network can handle image variations such as translation, rotation, distortion and scaling invariance, and is developed based on visual cortical processing which is suitable for identifying places. They are capable of image smoothening, image segmentation and object classification.

Pulse coupled neural network based image classification

The objective of the project is to study the feature extraction and dynamic properties of the pulse coupled neural network (PCNN) and determine its potential for classification of images. The pulse coupled neurons are significantly different from conventional artificial neurons as they intend to model the essence of the understanding of image interpretation process in biological neural system. The basis for PCNN is the linking field of Eckhorn, Reitboeck, Arndt and Dicke (1989). The model produces synchronous bursts of pulses from inputs/neurons with similar activity, effectively grouping them by phase and pulse frequency. It gives a basic new function: grouping by similarity. PCNN generates an object-specific time signal (referred to as an icon) that can be used as an object signature for object recognition. The signal detected may be made invariant to translation, scale, rotation, distortion, and intensity. The time signals generated by the PCNN were given as input to the classification network. The network recognized and classified the time signals with 90% accuracy.

Neural network based transient stability assessment of electric power systems

Abstract In this paper a systematic procedure to design an intelligent neurocontroller to assess the dynamic security of interconnected power systems is presented. This approach focuses on the integration of modern control theory together with the adaptive networks to determine critical clearing time for power systems.

Evaluating design projects in EE curriculum

The authors describe the design experiences and the evaluation of capstone design projects in the electrical engineering (BS) curriculum at Tennessee State University. Evaluation tools are also described.<<ETX>>

A neural network classifier for recycling process

A neural-network-based material classification and sorting process is presented. The network is designed, built, and trained to classify four different recycling materials, i.e., plastic, aluminum, glass, and others into four classes. The network was tested on a set of measurements, and the network performance is graphically illustrated by plotting the actual measurements and the identified class for each measurement (neural net output). Detailed explanations including detection techniques, training rules, and procedures used to accomplish the task are presented.<<ETX>>