Maryhelen Stevenson

Classification of the myoelectric signal using time-frequency based representations

An accurate and computationally efficient means of classifying surface myoelectric signal patterns has been the subject of considerable research effort in recent years. Effective feature extraction is crucial to reliable classification and, in the quest to improve the accuracy of transient myoelectric signal pattern classification, an ensemble of time-frequency based representations are proposed. It is shown that feature sets based upon the short-time Fourier transform, the wavelet transform, and the wavelet packet transform provide an effective representation for classification, provided that they are subject to an appropriate form of dimensionality reduction.

Time-frequency representation for classification of the transient myoelectric signal

An accurate and computationally efficient means of classifying myoelectric signal (MES) patterns has been the subject of considerable research effort in recent years. Effective feature extraction is crucial to reliable classification and, in the quest to improve the accuracy of transient MES pattern classification, many forms of signal representation have been suggested. It is shown that feature sets based upon the short-time Fourier transform, the wavelet transform, and the wavelet packet transform provide an effective representation for classification, provided that they are subject to dimensionality reduction by principal components analysis.

Improving myoelectric signal classification using wavelet packets and principal components analysis

An accurate and computationally efficient means of classifying surface myoelectric signals has been the subject of considerable research effort in recent years. This work demonstrates how this may be achieved, using a wavelet packet based feature set in conjunction with principal components analysis.

A dynamic feedforward neural network for subset classification of myoelectric signal patterns

Many biological signals are transient in nature, and the myoelectric signal (MES) is no exception. This is problematic for pattern classifiers that fail to incorporate the structure present in the temporal dimension of these signals. Standard feedforward neural network classifiers have difficulty processing temporal signals-time cannot be implicitly represented by the network architecture. A dynamic feedforward neural network architecture is described here that more effectively integrates the temporal information in transient signals, The internal representation of time also allows the dynamic network to classify subsets of the full temporal record. This reduces the time needed to obtain a classification result-an obvious benefit to real-time identification applications, such as the control of prosthetic devices.

A robust influence matrix approach to fault diagnosis

A robust scheme is proposed to detect faults, isolate them, and estimate their severity. The feature vector, which is a vector formed of the coefficients of the system transfer function, is estimated using a robust two-stage identification scheme: 1) a higher-order model is estimated using a singular value decomposition-based batch least-squares algorithm; and 2) a reduced-order model is derived by filtering-out the noise artifacts. The system is decomposed into functional units characterized by physical parameters. The influence of these physical parameters on the feature vector is captured in a vector termed the influence vector. The distance between, the inner product of the feature vector, and the influence vector are analyzed for diagnose faults. The proposed scheme is evaluated both on a simulated as well as an actual control system.

Score normalization in playback attack detection

The task of a playback attack detector (PAD) is to decide whether an incoming recording shares the same originating utterance as any of N stored recordings. All recordings are noisy channel-distorted versions of the same phrase uttered by the same person; the originating utterances of the N stored recordings are assumed to be distinct. The proposed approach makes a decision based on a set of N similarity scores which quantify the similarity between the incoming recording and each of the N stored recordings. Although satisfactory results are obtained by thresholding the maximum of the N scores using speaker and phrase (SaP)-dependent thresholds, it is shown that the use of a relative similarity score (a normalized version of the maximum similarity score) results in significant performance improvements especially in the case when the incoming recording is a severely distorted version of a stored recording utterance, as well as for the case when SaP-independent thresholds are used.

Unsupervised and uncued segmentation of the fundamental heart sounds in phonocardiograms using a time-scale representation.

A methodology is proposed to segment and label the fundamental activities, namely the first and second heart sounds, S1 and S2, of the phonocardiogram (PCG). Information supplementary to the PCG, such as a cue from a synchronously acquired electrocardiogram (ECG), subject-specific prior information, or training examples regarding the activities, is not required by the proposed methodology. A bank of Morlet wavelet correlators is used to obtain a time-scale representation of the PCG. An energy profile of the time-scale representation and a singular value decomposition (SVD) technique are used to identify segments of the PCG that contain the fundamental activities. The robustness of the methodology is demonstrated by the correct segmentation of over 90% of 1068 fundamental activities in a challenging set of PCGs which were recorded from patients with normally functioning and abnormally functioning bioprosthetic valves. The PCGs included highly varying fundamental activities that overlapped in time and frequency with other aberrant non-fundamental activities such as murmurs and noise-like artifacts

Autonomous control systems: Monitoring, diagnosis, and tuning

A systematic and unified approach which accomplishes performance monitoring, performance improvement and fault prediction in control systems is proposed. The feature vector which is a vector formed of the coefficients of the estimate of the sensitivity function and the influence matrix which is the Jacobian of the feature vector with respect to the physical parameter are shown to contain the relevant information to realize an autonomous control system. The feature vector is estimated using a robust, accurate and reliable linear predictive coding algorithm. The influence matrix is computed by perturbing the physical parameters one at a time and estimating the feature vectors for each case. The proposed scheme is evaluated both on simulated as well as on actual control systems.

A playback attack detector for speaker verification systems

A playback attack detector (PAD), which can be mobilized in guarding speaker verification systems against playback attacks, is described in this paper. To detect playback attacks, the PAD uses a feature set called peakmap, which includes the frame and FFT bin numbers of the five highest spectral peaks from each of the voiced frames in an utterance. During the detection, the peakmap of the incoming recording is first extracted and then compared to those of all the other recordings that are stored at the system end. Each comparison will yield a similarity score that represents the level of similarity between the two recordings. The incoming recording is declared to be a playback recording if its maximum similarity score is above a threshold.

Myoelectric signal classification using a finite impulse response neural network

Recent work by Hudgins (1993) has proposed a neural network-based approach to classifying the myoelectric signal (MES) elicited at the onset of movement of the upper limb. A standard feedforward artificial network was trained (using the backpropagation algorithm) to discriminate amongst four classes of upper-limb movements from the MES, acquired from the biceps and triceps muscles. The approach has demonstrated a powerful means of classifying limb function intent from the MES during natural muscular contraction, but the static nature of the network architecture fails to fully characterize the dynamic structure inherent in the MES. It has been demonstrated previously that a finite-impulse response (FIR) network has the ability to incorporate the temporal structure of a signal, representing the relationships between events in time and providing translation invariance of the relevant feature set. The application of this network architecture to limb function discrimination from the MES is described here.