Hidden Markov Model Smoothing Estimate Theory And Its Application In Suppressing Seismic Data Stochastic Noise

As far as seismic data are processed, enhancing signal-to-noise ratio(SNR) and the resolution indicating its quality simultaneously is the research subject holding prospector's interest all the while. Raising SNR is the precondition and base of enhancing resolution, and whether the lengthways one or the transverse one depends on the seismic signal frequencies and the data's SNR. Because of the diversity of the formation cause of stochastic noise acting as an important factor to deteriorate SNR and the complexity of the geologic structure of the exploration region, the underground reflection energy received by the detector is rather weak. Especially for the data area with the SNR lower than 0.5dB that the signal is comparatively feeble in the background of the strong stochastic noise, it is difficult for some common methods to obtain the ideal results of noise elimination. Nay, sometimes, noise suppression measures lead to the signal being damnified to the different extent. So the research trying for the robust technique of noise suppression is a challenging task all the same. In addition, most of the current means are based on the model of the multiple traces and the supposition that signal approximates the linearity within the some transverse range and the waveforms resemble each other, and the idea suprresssing stochastic noise based on the single trace model of one dimension time is waiting for the further development.The dissertation commences with processing single trace and treats the following aspects as the main goal probing into the problem of stochastic noise suppression. One is to design a method implementing the effective separation between signal and noise and ensuring an ideal SNR level. The other is supposed to avoid the distinct aberration of the reconstructed signal incurred by some common methods and presume upon decreasing it as low as possible, if possible.Founding on the above objects, the dissertation embraces the tasks summarized as follows,a) Study the problem of stochastic noise suppression of seismic data, and construct its mathematical model within the framework of the nonlinear/linear state space model.b) On the basis of the structure similarity and approximate periodicity of seismic signal, create the model for the short-time signal segment(frame) describing the non-stationarity of the seismic reflections. Premised by considering signal detection uniformly and state estimation and reaching the rational amalgamation, the combination model of noise submodel and signal frame submodel characterizes the data variation law of synthetical seismic records. c) Apply the continuous hidden Markov models(HMM) smoothing estimate to stochastic noise suppression. By defining the state chain parameter corresponding to the main frequency of Ricker wavelet, discuss its influencing the performance of continuous HMM smoothing estimate.d) Use continuous HMM smoothing estimate to deal with noisy synthetical seismic records with the events of the multiform types and do the emulation experiments of the theoretical models as well as give the performance contrast with other methods.e) Apply continuous HMM smoothing estimate to the practical seismic section, and analyze its application effect.The dissertation consists of six chapters, and the basic content of each one is as follows,Chapter one reiterates the types of stochastic noise and explains many common classifications suppressing it after introducing the research background and significance of stochastic noise suppression of seismic data processing simply. Summarize the status quo of its current research according to the classification way of time-space domain and transformed domain. Finally elicit the substantial objects, the structure and the concrete contents of the dissertation.Chapter two specifies the HMM principles and the algorithms of HMM filtering. Giving the example of the discrete HMM, explain such three basic problems as probability evaluation,'optimal'state sequence extraction and model parameter estimation. Subsequently, elucidate the continuous HMM to which the dissertation attaches importance and the five problems which must be solved in the course of the application. Regarding the view that HMM is essentially a Bayesian finite state process as the cut-in, introduce the state space framework and Bayesian filtering principles along with HMM filtering algorithms.Chapter three studies the stochastic noise suppression method of HMM smooting estimate(HMM-S) taking synthetical seismic records which play an important role in the geology property analysis of the reflection groups and geology layer determination as the object. Its primary expectation is to buck for an idea to remove the noise of the noisy records. Firstly laying the foundation for the application of Bayesian statitics inference, establish the state pace model of noisy synthetical seismic records and produce the noise-signal frame continuous HMM describing their data variation law. Secondly detail the parameter training flow of the wavelet frame continuous HMM. By comparing with the results of time-frequency peak filtering(TFPF) and Wiener filtering, evaluate the performance of HMM-S reducing noise. Discuss the varieties of the factors influencing its performance, the emulation experiments indicate the relations between the wavelet frame HMM state chain parameter settings and signal reconstruction accuracy.Chapter four integrates constructing the many models of single hyperbolic reflection events or multiple hyperbolic reflection events and lucubrates HMM-S by the theoretical emulation experiments. First of all, make the models of single hyperbolic reflection events using zero-phase wavelet(Ricker wavelet) or mixed phase wavelet according to the different parameters such as wavelet frequency, arrival time and layer velocity, etc. By a series of experments, investigate HMM-S performance influenced by the intensity variation of stochastic noise, the absorption attenuation of seismic wave frequency, phase alteration arosed by souce wavelet traveling through the underground medium. Not only does HMM-S eliminate the noise of synthetical section with SNR lower to -0.236dB and improve its quality obviously, but also embodies the definite robustness while processing the sections whose quality is unbalanced in some local areas. In the next place, make synthetical seismic records using multiple horizontal and gradient events as well as hyperbolic events respectively, and analyze the HMM-S performance with multiple events coexisting in the same section. Experiments indicate that the components of the multiple events of the complex sections including the faultage, folium and incomeplete events, etc, are enhanced without exception, and their continuity is ameliorated with the favourable reconstruction effects for the section details.Chapter five dicusses the noise suppression application of the practical seismic data for HMM-S. For the concrete data, introduce the implementation procedures orienting common shot point seismic record, the parameter settings, the results'comparison analysis and the problems demanding the solutions so as to attain high robustness.In chapter six, a summary of the dissertation is given. The further research direction is illuminated.To sum up, the author of the dissertation evolves the study and probe separately from the two aspects of theory and application addressing the problems that exist in the area of stochastic noise suppression area and wait for the further reseach based on systematicly learning HMM principles, and reaps some fruits also. The applicability for HMM-S to attenuate the noise doesn't halt in the phase of the theory research at all, and it reaches the goal to suppress the noise when processing the practical seismic data. If letting it exert its particular advantage adequately in the extensive applications, it should be modified and perfected more delicately according to the actual.