Research On The Realization Method Of Autoencoder Based On Singular Spectrum Analysis

Singular spectrum analysis(SSA)uses the characteristics of seismic data has a low rank structure when constructing the Hankel pre-transformation matrix,and the interference of random noise increases the rank of the pre-transformation matrix or tensor.It is a classical denoising algorithm to suppress the random noise of seismic data by reducing the rank of Hankel matrix.However,this algorithm has some problems,such as the rank of Hankel matrix of seismic signal is difficult to determine and the operation speed is slow.In recent years,deep learning has become the most rapidly developing research direction in the field of artificial intelligence.In the field of seismic exploration,it is widely used in the prediction,reconstruction,denoising of seismic data.The model keeps learning the features of massive data and converts them into abstract feature representations.Using the high robustness model to process the same type of data to achieve fast and efficient seismic data processing.However,the deep learning model is a black box,and there are disadvantages such as poor interpretability of the training process and unmanageable control of the training direction.Therefore,this paper introduces the singular spectrum analysis algorithm into the autoencoders to find the eigenvector and process the seismic data in frequency division.Using the theoretical basis of SSA and the sparsity constraint of the network,it adapts to the fixed rank and guides the selection of the number of hidden layer neurons,so as to realize an unsupervised learning algorithm models which have fast operation speed and retain the physical interpretability of traditional methods.Synthetic records and actual seismic data processing show that this method has good performance in seismic random noise suppression.