A new rule-based control system for rotary ventricular assist devices (rVADs) is proposed. The control system is comprised of two modules: a suction detector and a rule-based controller. The suction detector can classify pump flow patterns, based on a discriminant analysis (DA) model that combines several indices derived from the pump flow signal, to make a decision about the pump status. The indices considered in this approach are frequency, time, and time-frequency-domain indices. These indices are combined in a DA decision system to generate a suction alarm. The suction detector performance was assessed using experimental data and in simulations. Experimental results comprise predictive discriminant analysis (classification accuracy: 100% specificity, 93% sensitivity on training set and 97% specificity, 86% sensitivity on test set) of the detector and descriptive discriminant analysis (explained variance) of the DA model. To perform the simulation studies, the suction detector was coupled to a cardiovascular-pump model that included a suction model. Simulations were carried out to access the detector performance, under different physiological conditions, i.e., by varying preload and the contractility state of the left ventricle. To verify its robustness to noise, simulations were carried out to verify how the accuracy of the detector is affected when increasing levels of noise are added to the pump flow signal. The rule-based controller uses fuzzy logic to combine the discriminant scores from the DA model to automatically adjust the pump speed. The effects on controller performance of symmetric or asymmetric membership output sets and the dimension of the rule base were evaluated in simulations. The same parameter changes, i.e., preload and contractility, were used to assess the control system performance under different physiologic scenarios in simulations. The proposed control system is capable of automatically adjusting pump speed, providing pump flow according to the patient's level of activity, while sustaining adequate perfusion pressures and avoiding suction. In addition, the control system performance was not adversely affected by noise until SNR was less than 20dB, which is a higher noise level than is commonly encountered in flow sensors used clinically for this type of application.

The brains motor control can be studied by characterizing the activity of spinal motor nuclei to brain control, expressed as motor unit activity recordable by surface electrodes. When a specific area is under consideration, the first step in investigation of the motor control system pertinent to it is the system identification of that specific body part or area. The aim of this research is to characterize the working of the brains motor control system by carrying out system identification of the wrist joint area and quantifying tremor observed in Parkinsons disease patients. We employ the ARMAX system identification technique to gauge the intrinsic and reflexive components of wrist stiffness, in order to facilitate analysis of problems associated with Parkinsons disease. The intrinsic stiffness dynamics comprise majority of the total stiffness in the wrist joint and the reflexive stiffness dynamics contribute to the tremor characteristic commonly found in Parkinsons disease patients. The quantification of PD tremor entails using blind source separation of convolutive mixtures to obtain sources of tremor in patients suffering from movement disorders. The experimental data when treated with blind source separation reveals sources exhibiting the tremor frequency components of 3-8 Hz. System identification of stiffness dynamics and assessment of tremor can reveal the presence of additional abnormal neurological signs and early identification or diagnosis of these symptoms would be very advantageous for clinicians and will be instrumental to pave the way for better treatment of the disease.

We introduce new spatial filtering methods in the spherical harmonics domain for constraining magnetoencephalographic (MEG) multichannel measurements to user-specified spherical regions of interests (ROI) inside the head. The main idea of the spatial filtering is to emphasize those signals arising from an ROI, while suppressing the signals coming from outside the ROI. We exploit a well-known method called the signal space separation (SSS), which can decompose MEG data into a signal component generated by neurobiological sources and a noise component generated by external sources outside the head. The novel methods presented in this work, expanded SSS (exSSS) and generalized expanded SSS (genexSSS) utilize a beamspace optimization criterion in order to linearly transform the inner signal SSS coefficients to represent the sources belonging to the ROI. The filters mainly depend on the radius and the center of the ROI. The simplicity of the derived formulations of our methods stems from the natural appropriateness to spherical domain and orthogonality properties of the SSS basis functions that are intimately related to the vector spherical harmonics. Thus, unlike the traditional MEG spatial filtering techniques, exSSS and genexSSS do not need any numerical computation procedures on discretized headspace. The validation and performance of the algorithms are demonstrated by experiments utilizing both simulated and real MEG data.

Model order reduction (MOR) is a very powerful technique that is used to deal with the increasing complexity of dynamic systems. It is a mature and well understood field of study that has been applied to large linear dynamic systems with great success. However, the continued scaling of integrated micro-systems, the use of new technologies, and aggressive mixed-signal design has forced designers to consider nonlinear effects for more accurate model representations. This has created the need for a methodology to generate compact models from nonlinear systems of high dimensionality, since only such a solution will give an accurate description for current and future complex systems. The goal of this research is to develop a methodology for the model order reduction of large multidimensional nonlinear systems. To address a broad range of nonlinear systems, which makes the task of generalizing a reduction technique difficult, we use the concept of transforming the nonlinear representation into a composite structure of well defined basic functions from multiple projection bases. We build upon the concept of a training phase from the trajectory piecewise-linear (TPWL) methodology as a practical strategy to reduce the state exploration required for a large nonlinear system. We improve upon this methodology in two important ways: First, with a new strategy for the use of multiple projection bases in the reduction process and their coalescence into a unified base that better captures the behavior of the overall system; and second, with a novel strategy for the optimization of the state locations chosen during training. This optimization technique is based on using the Hessian of the system as an error bound metric. Finally, in order to treat the overall linear/nonlinear reduction task, we introduce a hierarchical approach using a block projection base. These three strategies together offer us a new perspective to the problem of model order reduction of nonlinear systems and the tracking or preservation of physical parameters in the final compact model.

Studies have shown that visual features extracted from the lips of a speaker (visemes) can be used to automatically classify the visual representation of phonemes. Different visual features were extracted from the audio-visual recordings of a set of phonemes and used to define Linear Discriminant Analysis (LDA) functions to classify the phonemes. . Audio-visual recordings from 18 speakers of Native American English for 12 Vowel-Consonant-Vowel (VCV) sounds were obtained using the consonants /b,v,w,ð,d,z/ and the vowels /ɑ,i/. The visual features used in this study were related to the lip height, lip width, motion in upper lips and the rate at which lips move while producing the VCV sequences. Features extracted from half of the speakers were used to design the classifier and features extracted from the other half were used in testing the classifiers. When each VCV sound was treated as an independent class, resulting in 12 classes, the percentage of correct recognition was 55.3% in the training set and 43.1% in the testing set. This percentage increased as classes were merged based on the level of confusion appearing between them in the results. When the same consonants with different vowels were treated as one class, resulting in 6 classes, the percentage of correct classification was 65.2% in the training set and 61.6% in the testing set. This is consistent with psycho-visual experiments in which subjects were unable to distinguish between visemes associated with VCV words with the same consonant but different vowels. When the VCV sounds were grouped into 3 classes, the percentage of correct classification in the training set was 84.4% and 81.1% in the testing set. In the second part of the study, linear discriminant functions were developed for every speaker resulting in 18 different sets of LDA functions. For every speaker, five VCV utterances were used to design the LDA functions, and 3 different VCV utterances were used to test these functions. For the training data, the range of correct classification for the 18 speakers was 90-100% with an average of 96.2%. For the testing data, the range of correct classification was 50-86% with an average of 68%. A step-wise linear discriminant analysis evaluated the contribution of different features towards the dissemination problem. The analysis indicated that classifiers using only the top 7 features in the analysis had a performance drop of 2-5%. The top 7 features were related to the shape of the mouth and the rate of motion of lips when the consonant in the VCV sequence was being produced. Results of this work showed that visual features extracted from the lips can separate the visual representation of phonemes into different classes.

Field Programmable Gate Arrays (FPGAs) mitigate many of the problems encountered with the development of ASICs by offering flexibility, faster time-to-market, and amortized NRE costs, among other benefits. While FPGAs are increasingly being used for complex computational applications such as signal and image processing, networking, and cryptology, they are far from ideal for these tasks due to relatively high power consumption and silicon usage overheads compared to direct ASIC implementation. A reconfigurable device that exhibits ASIC-like power characteristics and FPGA-like costs and tool support is desirable to fill this void. In this research, a parameterized, reconfigurable fabric model named as domain specific fabric (DSF) is developed that exhibits ASIC-like power characteristics for Digital Signal Processing (DSP) style applications. Using this model, the impact of varying different design parameters on power and performance has been studied. Different optimization techniques like local search and simulated annealing are used to determine the appropriate interconnect for a specific set of applications. A design space exploration tool has been developed to automate and generate a tailored architectural instance of the fabric. The fabric has been synthesized on 160 nm cell-based ASIC fabrication process from OKI and 130 nm from IBM. A detailed power-performance analysis has been completed using signal and image processing benchmarks from the MediaBench benchmark suite and elsewhere with comparisons to other hardware and software implementations. The optimized fabric implemented using the 130 nm process yields energy within 3X of a direct ASIC implementation, 330X better than a Virtex-II Pro FPGA and 2016X better than an Intel XScale processor.

A primary objective of active sonar systems is to detect, locate, and classify objects, such as mines, ships, and biologics, based on their sonar backscatter.  A shallow-water ocean channel is a challenging environment in which to classify sonar echoes because interactions of the sonar signal with the ocean surface and bottom induce frequency-dependent changes (especially dispersion and damping) in the signal as it propagates, the effects of which typically grow with range.  Accordingly, the observed signal depends not only on the initial target backscatter, but also the propagation channel and how far the signal has propagated.  These propagation effects can increase the variability of observed target echoes and degrade classification performance.    Furthermore, uncertainty of the exact propagation channel and random variations within a channel cause classification features extracted from the received sonar echo to behave as random variables. With the goal of improving sonar signal classification in shallow-water environments, this work develops a phase space framework for studying sound propagation in channels with dispersion and damping. This approach leads to new moment features for classification that are invariant to dispersion and damping, the utility of which is demonstrated via simulation.  In addition, the accuracy of a previously developed phase space approximation method for range-independent pulse propagation is analyzed and shown to be greater than the accuracy of the standard stationary phase approximation for both large and small times/distances.  The phase space approximation is also extended to range dependent propagation.  Finally, the phase space approximation is used to investigate the random nature of moment features for classification by calculating the moments of the moment features under uncertain and random channel assumptions. These moments of the moment features are used to estimate probability distribution functions for the moment features, and we explore several ways in which this information may be used to improve sonar classification performance.

Studies have shown that transient speech, which is associated with consonants, transitions between consonants and vowels, and transitions within some vowels, is an important cue for identifying and discriminating speech sounds. However, compared to the relatively steady-state vowel segments of speech, transient speech has much lower energy and thus is easily masked by background noise. Emphasis of transient speech can improve the intelligibility of speech in background noise, but methods to demonstrate this improvement have either identified transient speech manually or proposed algorithms that cannot be implemented to run in real-time. We have developed an algorithm to automatically extract transient speech in real-time. The algorithm involves the use of a function, which we term the transitivity function, to characterize the rate of change of wavelet coefficients of a wavelet packet transform representation of a speech signal. The transitivity function is large and positive when a signal is changing rapidly and small when a signal is in steady state. Two different definitions of the transitivity function, one based on the short-time energy and the other on Mel-frequency cepstral coefficients, were evaluated experimentally, and the MFCC-based transitivity function produced better results. The extracted transient speech signal is used to create modified speech by combining it with original speech. To facilitate comparison of our transient and modified speech to speech processed using methods proposed by other researcher to emphasize transients, we developed three indices. The indices are used to characterize the extent to which a speech modification/processing method emphasizes (1) a particular region of speech, (2) consonants relative to, and (3) onsets and offsets of formants compared to steady formant. These indices are very useful because they quantify differences in speech signals that are difficult to show using spectrograms, spectra and time-domain waveforms. The transient extraction algorithm includes parameters which when varied influence the intelligibility of the extracted transient speech. The best values for these parameters were selected using psycho-acoustic testing. Measurements of speech intelligibility in background noise using psycho-acoustic testing showed that modified speech was more intelligible than original speech, especially at high noise levels (-20 and -15 dB). The incorporation of a method that automatically identifies and boosts unvoiced speech into the algorithm was evaluated and showed that this method does not result in additional speech intelligibility improvements.

Video transmission plays a critical role in robotic telesurgery because of the high bandwidth and high quality requirement. The goal of this dissertation is to find a preprocessing method based on human visual perception for telesurgical video, so that when preprocessed image sequences are passed to the video encoder, the bandwidth can be reallocated from non-essential surrounding regions to the region of interest, ensuring excellent image quality of critical regions (e.g. surgical region). It can also be considered as a quality control scheme that will gracefully degrade the video quality in the presence of network congestion. The proposed preprocessing method can be separated into two major parts. First, we propose a time-varying attention map whose value is highest at the gazing point and falls off progressively towards the periphery. Second, we propose adaptive spatial filtering and the parameters of which are adjusted according to the attention map. By adding visual adaptation to the spatial filtering, telesurgical video data can be compressed efficiently because of the high degree of visual redundancy removal by our algorithm. Our experimental results have shown that with the proposed preprocessing method, over half of the bandwidth can be reduced while there is no significant visual effect for the observer. We have also developed an optimal parameter selecting algorithm, so that when the network bandwidth is limited, the overall visual distortion after preprocessing is minimized.

Automatic speech recognition has a wide variety of uses in this technological age, yet speech distortions present many difficulties for accurate recognition. The research presented provides solutions that counter the detrimental effects that some distortions have on the accuracy of automatic speech recognition. Two types of speech distortions are focused on independently. They are distortions due to speech coding and distortions due to additive noise. Compensations for both types of distortion resulted in decreased recognition error. Distortions due to the speech coding process are countered through recognition of the speech directly from the bitstream, thus eliminating the need for reconstruction of the speech signal and eliminating the distortion caused by it. There is a relative difference of 6.7% between the recognition error rate of uncoded speech and that of speech reconstructed from MELP encoded parameters. The relative difference between the recognition error rate for uncoded speech and that of encoded speech recognized directly from the MELP bitstream is 3.5%. This 3.2 percentage point difference is equivalent to the accurate recognition of an additional 334 words from the 12,863 words spoken. Distortions due to noise are offset through appropriate modification of an existing noise reduction technique called minimum mean-square error log spectral amplitude enhancement. A relative difference of 28% exists between the recognition error rate of clean speech and that of speech with additive noise. Applying a speech enhancement front-end reduced this difference to 22.2%. This 5.8 percentage point difference is equivalent to the accurate recognition of an additional 540 words from the 12,863 words spoken.