Time-frequency Peak Filtering And Its Application On Random Noise Attenuation For Seismic Exploration Data

IN practice, signals are often corrupted by noise, and this has the effect ofhindering the recovery of important information encoded in the signal. All signalprocessing fields including radar, sonar, communications, seismology, andbiomedicine suffer from this problem. The performance of signal enhancementalgorithms generally depends on the signal-to-noise ratio (SNR). Many signalprocessing algorithms work well for high SNR situations, but most perform poorlywhen SNR decreases below a given threshold. In this case, if no adequatealternative algorithm is available, preprocessing or filtering is required to improvethe SNR.Adaptive and fixed methods have been developed for signal enhancement ofnonstationary signals in noise. Although both approaches have merits, adaptivetechniques are generally superior in performance to fixed methods when the signalstatistics are nonstationary and unknown. Some of the well-known adaptivefiltering techniques are based on the least-mean-squares (LMS) algorithm, therecursive least-squares algorithm, and the Kalman filter (KF).Adaptive filters could not perform very well in certain conditions, however.An example of this is filtering of a nonstationary signal whose spectral contentchanges quickly with time. The filter designed using LMS approach, for instance,may not adapt quickly enough to track the rapidly changing signal. This is due tothe delayed convergence of the algorithm, which is a function of the signalautocorrelation matrix eigenvalues. Further, in many signal processing applicationssuch as electroencephalogram (EEG) signal, the structure of the underlying signalis often unknown and too complicated to model accurately. An attempt to provide amodel framework could probably lead to suboptimal results.Recently, efforts have been made to develop Time-frequency analysis, whichdescribes the frequency varying with time. This method uses the jointtime-frequency domain to analysis signal and maps the one-dimensional signal totwo-dimensional time-frequency plane. It shows the information of the signal inboth time domain and frequency domain. Time-frequency analysis can be used toexhibit significant energy concentration around the signal's IF and develop thetime-frequency filtering.In 2004, Time-frequency peak filtering (TFPF) have recently been used byB.Boashash to enhance signals, which may be represented as a sum of bandlimitednonstationary processes in additive white Gaussian noise(WGN). The noisy signalis encoded as the IF of a unit amplitude frequency modulated (FM) analytic signal.The instantaneous frequency (IF) of the analytic signal is then estimated usingstandard time-frequency peak detection methods based WVD to obtain an estimateof the underlying deterministic signal. TFPF which is a noncoherent method hasalready been applied to Newborn EEG data.The aim of this paper is to use TFPF to suppress random noise in a reflectionseismic data. The experiments show that a very clean recovery of seismic reflectiondata can be accomplished. Therefore it indicates the efficiency of the algorithm as anoise-eliminated method for seismic data.Seismic data is the important information resource of geological prospect andexploration. Random noise in seismic reflection data can be introduced by varioussources and is often a problem in geophysical data visualization because it obscuresfine details and complicates identification of image features.In present, there are basically two kinds of methods to reduce the noise inseismic data: signal enhancement algorithms and noise suppression algorithms.These methods including f ? x deconvolution, polynomial fitting, radialpredictive deconvolution, belong to the signal enhancement algorithms, and thevector decomposition belongs to the noise suppression algorithms. These methodswere often used in practice. However, some assumptions of seismic data for themethods above will lead to suboptimal results about the signal. Generally speaking,the assumptions are that effective primary wave is correlated in space, whereasrandom noise is not correlated in space. One-dimensional noise reduction methodswere studied very little.In this paper, Time-frequency peak filtering is used to reconstruct signals fromreflection seismic data corrupted by random noise. This method which is aone-dimensional signal enhancement algorithm doesn't need any assumptions toseismic data.Firstly, this paper discussed the basic principle and algorithm realization ofTFPF. Secondly get the signal model according to the characteristic of seismic data.This signal model showed the signal composition in reason. Thirdly, simulationwas conducted on synthetic seismic data with an event or multi-event andcomplicated seismic data. Simulations have demonstrated that TFPF could reducerandom noise effectively. In comparison with other methods, TFPF used thediscarding seismic recordings, so it could save economic resources. Finally, weanalyzed the effect of each parameter on TFPF to assist the simulation. There arethree factors: window length, sample frequency and the noise. When implementingthe experiments, we should pay attention to these factors according to the seismicdata.The different experiments conducted demonstrated that this method of signalenhancement could clearly show the position of events and the Ricker wavelets.We compared seismic data corrupted in noise with the filtering result from thewavelets, Wigner-Ville distribution and Fourier spectrum. It was concluded thatreducing random noise of seismic data by TFPF is feasible and has research value.