Why Spectral Analysis? The spectrum provides an alternate representation of the same information that is contained in the time-domain representation of the waveform (a different but more useful perspective!).
When we examine two waveforms one with higher frequency content than the other they generally appear to be visually different in the time domain in terms of the rapidity of variation with respect to time, if there is clear separation of their frequency content.
But what if we needed to go further and quantify this difference? What if the two signals overlap in frequency content, in which case they may not appear very different visually in the time domain?
Possible Applications Speech/speaker recognition Engine diagnosis based on vibration analysis, detection of abnormal combustion in an internal combustion engine. Prediction of machine tool breakdown or wear (pseudo or quasi stationary process) Communication systems determining possible source of noise interference or typing the noise source (which may or may not be stationary) Processing of EEG, ECG, and other signals for medical diagnosis Predicting earthquakes, weather
Comment Even though the fundamental mathematical relationships of random processes are often obtained based on assumption of some form of stationarity, in real life we often have to deal with non-stationary signals. Typically they are slowly changing stationary signals (which still makes them non-stationary!), which enables the use of theoretical results based on assumption of stationarity.
Curtin University, Australia 8-2 YH Leung (2005, 2012) Classification of Techniques The body of spectral estimation techniques falls into two categories:
Basic Principle These methods of estimation are based on using the DFT (actually the FFT) directly on the samples of a sample function of the (random) process. They make no assumptions about how the data was generated.
Advantages Since they are based on the FFT, they can be quite fast.
Disadvantages The spectral resolution is low when only short data records are available.
2 Parametric (Modern) Techniques Autoregressive (AR) model Moving Average (MA) model Autoregressive Moving Average (ARMA) model Advanced techniques
Basic Principle A model (characterized by parameters) is postulated for the production process of the random signal. These parameters are estimated in some optimal way using a sample function of the process.
Advantages These methods generally enable increased spectral resolution even with shorter data lengths.
Disadvantages An appropriate model has to be chosen based on some intuition about the process and additional tests. The model order has to be selected and the parameters have to be computed from the available data.
"Adaptive Noise Cancellation Using The Adaptive Neuro-Fuzzy Inference System Algorithm" Yassir A. Granillo April 20, 2006 EE 5390: Fuzzy Logic and Engineering Professor: Dr. Thompson Sarkodie-Gyan