Research And Application Of Oil And Gas Reservoir Information Extraction Based On HHT And Neural Network

Seismic signals are complex non-linear and non-stationary signals whose timefrequency properties contain rich formation information and have been proven to reflect the characteristics of carbonate reservoirs.The time-frequency analysis methods convert the one-dimensional time signal into the two-dimensional time-frequency domain and can analyze the time and frequency characteristics of the signal in the local range.They are effective methods for analyzing non-linear and non-stationary signals.Huang created a complete time-frequency analysis method system,Hilbert-Huang Transform(HHT),which is based on the characteristics of the signal itself and the definition of constraints.It can perform high-precision time-frequency decomposition of signals and has been applied to reservoir prediction,but modal aliasing,endpoint effects,spectral analysis with singular values limit its application.In this thesis,we systematically improve the defects existing in the HHT method to obtain the low-frequency shadow profiles of seismic data in western Sichuan and predict the gas-bearing characteristics of the reservoir in the target area.Combined with logging data,we apply a supervised neural network to achieve adaptive high-precision identification of reservoir gas.In summary,the main contents of this thesis are as follows:(1)Based on comparing the principles and shortcomings of conventional timefrequency analysis tools such as Short-Time Fourier Transform,Wigner-Ville Distribution,Wavelet Transform and S-transform,this thesis introduces HHT method to process non-linear and non-stationary signals.We systematically analyze the basic principles of the HHT method,the definition of the Intrinsic Mode Function(IMF),the process of Empirical Mode Decomposition(EMD)and the method of Hilbert Transform(HT)to extract the instantaneous properties of the IMF.HHT shows high precision in both time and frequency domains.(2)For the deficiencies in the HHT method,this thesis analyzes the causes one by one and improves them to form a systematic improved HHT method.Among them,we use the Complete Ensemble Empirical Mode Decomposition(CEEMD)to add noise to the signal,so that each frequency component is continuous on the time scale,effectively suppressing the modal aliasing of EMD;we perform waveform matching and extrema mirror extension on the original signal to make the endpoint error of the IMF isolated outside the effective range;we use the Normalized Hilbert Transform(NHT)to avoid the instantaneous parameter singular values in HT.The simulation analysis shows that the improved HHT has higher precision and less error and is more suitable for the processing of seismic signals.(3)For the actual seismic data in western Sichuan,the Hilbert spectrum of seismic signals is extracted by the improved HHT method to construct the frequency component profile.The reservoir location has the characteristics of “higher energy in the low frequency part and weaker energy in the high frequency part”.Based on this feature,we predict the fluid distribution of marine carbonate reservoirs in western Sichuan.The T-K energy operator is introduced when constructing the crossover profile,which improves the energy concentration in the crossover profile.In this thesis,we established a low-frequency shadow model to quantitatively describe the energy difference of different frequency profiles.(4)Neural network is a kind of complex abstract mathematical model built on the biological neural system,which can be used to classify and identify non-linear mapping relationships.In this thesis,we extract Hilbert spectrum of seismic signals from the perspective of time-frequency analysis,construct multiple frequency component profiles as parameters to train the BP neural network,and finally realize the identification of reservoir fluids in western Sichuan.The method has high precision and strong adaptability without influence by subjective factors.