header images

Output
and
Outcome

Signal processing

Signal processing is often used for feature extraction and classification in medical disease diagnosis, industrial process control, fault detection and many other fields. The primary goal of signal processing in the aforementioned applications is to provide underlying information on specific problems for decision making. These techniques can be classified either as time, frequency or time-frequency domain based algorithms. At the classification level, there also exist several different methodologies.

Understanding of the problem at hand is crucial in deciding which framework to employ for feature analysis. Some features, such as amplitude levels in the time domain, are easily extracted and classified, but are susceptible to noise. Others, such as energy concentration in the time-frequency domain, even though require more involved operations, can lead to more robust feature extraction and more accurate classification. Furthermore, not every feature yields plausible conclusions. For example, in the analysis of heart sounds, which are nonstationary, the amplitude rarely provides conclusive information. The intensity of the recorded heart sounds is affected by many factors, which are not necessarily pathological. On the other hand, the amplitude in the time domain will provide sufficient information when considering control of the liquid level in a tank.

Depending upon whether the phenomenon under analysis is stationary or nonstationary, and on the nature of the desired feature, different algorithms have to be used. The question is what signal processing algorithms should be used for feature analysis in a given situation? The answer simply depends on a priori knowledge about the phenomenon under consideration. Parametric signal processing algorithms can be used for feature extraction and classification if an accurate model of the signal exists in a selected representation space. However, such modeling techniques have limitations as well. Modeling of nonstationary signals is more difficult and consistent parametric models often do not exist, except in very few special cases, e.g., mono or multi component chirp signals. Most of the signals encountered in practice do not satisfy the stationarity conditions, which explains our interest in nonstationary signal processing tfr
ece
Innovative Medical Engineering Developments Laboratory
Department of Electrical and Computer Engineering
Swanson School of Engineering
University of Pittsburgh