Rohit Bose

Classification of lower limb motor imagery using K Nearest Neighbor and Naïve-Bayesian classifier

For development of foot prosthetics driven by brain computer interface (BCI) for lower limb amputees, the primary requirement is the classification of right and left lower limb motor imagery movement from brain signals. It is important to detect best possible combination of feature extraction and classification algorithms efficiently and accurately recognize left and right lower limb motor imagery from Electroencephalogram (EEG) signals in minimum time possible. An optimal choice has to be reached to select a feature extraction and classification technique with highest accuracy in minimum time. Thus, in this study we direct our attention towards finding the best feature extraction technique and classifier. Preprocessing of the EEG signals are done and relevant features are extracted. The extracted features are then used to classify left and right imagery movement by k-Nearest Neighbor (kNN) and Naïve-Bayesian classifier. The best classification accuracy of 90% is obtained by kNN for power spectral density feature set requiring a time of 0.0531 sec. Thus, in future, it can be applied in real time classification to obtain best results in minimum time.

Performance analysis of left and right lower limb movement classification from EEG

The primary prerequisite for development of foot prosthetics driven by brain computer interface (BCI) is classification of left and right lower limb movement from brain signals. Moreover, it is essential to detect best possible combination of feature extraction and classification technique which will efficiently recognize left and right lower limb movement intentions from brain signals in as minimum time as possible. Thus, directed towards finding a solution to these problems, this study is aimed at determining most efficient feature extractor and classifier for recognizing left-right leg movement from brain signals acquired using electroencephalography (EEG). The acquired EEG signals are preprocessed and relevant features are extracted using Power Spectral Density (PSD), statistical parameters (mean, standard deviation, and power) and Hjorth parameter. These extracted features are classified using kNN (k Nearest Neighbor) and SVM (Support Vector Machine). Various parameters like classification accuracy, specificity, sensitivity, and run time are computed for analyzing the performance of the classifiers. kNN and SVM are also statistically analyzed along with two more classifiers (Linear Discriminant Analysis and Quadratic Discriminant Analysis) by Friedman Test. The best accuracy of 90% and sensitivity of 89% is obtained by kNN using Hjorth features taking least time of 78ms. The classifier results are also statistically validated by Friedman Test.

Electric field computation in presence of water droplets on a polymeric insulating surface

In this paper electric field enhancement factor has been computed due to presence of different volumes of single as well as multiple water droplets on a polymeric insulating surface. Outdoor polymeric insulators are exposed to different environmental conditions like rain, fog, salted fog etc. Hence the formation of discrete water droplets on the surface are very common. These water droplets cause the electric field enhancements at the triple points (interfacial point of three dielectric mediums i.e. solid, liquid and air) which may lead to external partial discharge or even flashover. Hence, in this contribution, the electric field computation has been performed due to presence of discrete water droplets using Finite Element Method under tangential AC electric field. To overcome the singularity problem of calculating electric field at the triple points, a different method of field calculation have been presented in this contribution which may provide better understanding of the electric field distribution in presence of water droplets on a insulating surface.

Comparative study of electrode configurations in different brain tumor geometry for effective Electrochemotherapy

Electrochemotherapy is an effective emerging technique for cancer treatment. Electric pulses are applied to ensure proper penetration of the drugs in the tumor cell. Electrochemotherapy has a wide area of application in cancer diagnosis. Effective Electrochemotherapy involves proper selection of electrode type and configurations. In this paper, a 2D model of a brain tumor is simulated and a comparison of parallel plate electrode and different needle electrode configuration is given. Electric field analysis is performed using Finite Element Method using COMSOL Multiphysics software. The average electric field inside the tumor and the maximum electric field outside the tumor are taken as the control parameter to determine the best electrode configuration for different tumor geometry having same area.

Transient analysis of mechanically switched capacitors with and without damping network connected to A.C grid

Mechanically switched capacitors (MSC) are very well known reactive power compensation devices which are extensively in use in power transmission systems. However, during switching of MSC in a high voltage grid, significant transients are found to appear across the capacitor bank bus as well as different elements of MSC which may lead to insulation degradation affecting the long term performance of the capacitor banks. An improved version of the MSC is MSCDN where a damping network is employed with the capacitor banks to damp out the transient oscillations. However, during switching of a MSCDN plant significant transients are observed across its different elements. Considering the afore-said fact, in this contribution, a comparative study of transient behavior of mechanically switched capacitors with and without damping network is presented during switching in a 400 kV A.C grid. Transient simulations are performed on an exemplary MSC and MSCDN model using PSCAD/EMTDC software. It has been observed that there is a significant reduction in transient overvoltage and inrush current in switching a MSCDN plant when compared to a mere MSC switching. Moreover, the percentage reduction in transient overvoltage and inrush current during switching a MSCDN plant is found to depend on the design of damping network which may be important to analyze the transient behavior during switching in a high voltage A.C grid.

Electric stress analysis of a contaminated polymeric insulating surface in presence of dry bands

In this contribution, influence of dry bands on electric field distribution on a polymeric insulating surface has been carried out using Finite Element Method. Formation of dry bands on insulating surfaces result in electric field enhancement leading to surface discharges, dry-band arcing etc. which in turn affects the long term performance of the insulators. Hence, in the present work, a polymeric insulating surface is modeled as sheath model using Comsol Multiphysics software. Electric stress analysis is performed in presence of dry bands of different length and at different locations on a contaminated insulating surface. Moreover, the effect of different contamination levels on electric field distribution in presence of dry bands have also been investigated.

Cross-correlation based feature extraction from EMG signals for classification of neuro-muscular diseases

In this contribution, classification of two main neuromuscular diseases namely Myopathy and Neuropathy and Healthy signals is performed using cross-correlation based feature extraction technique. For this purpose, cross-correlation of Healthy, Myopathy and Neuropathy disease EMG signal is done with a reference Healthy signal. Selective features like Hjorth, Adaptive Autoregressive and statistical features comprising mean, standard deviation and power are extracted from the cross-correlated signals. Support Vector Machine(SVM) and k-Nearest Neighbor(kNN) are the two classifiers used for this work. Highest classification accuracy of 100% is obtainedby SVM using Gaussian Radial Basis Function (RBF) as the kernel function with AAR and all combined features as the feature set. For kNN, k=4 yields best result of 100% accuracy using the combined feature set.