Research On Waveform Clustering Method Based On Stacked Denoising Autoencoder

Oil and gas are strategic resources that we are short of.China's deep carbonate rock formations are rich in natural oil and gas resources,but reservoir identification is difficult.Seismic waveform data implies amount of reservoir information,and clustering analysis is an effective method for reservoir identification.The ability to identify reservoirs based on clustering original seismic waveform data directly is weak.To further mine the reservoir information in seismic waveform data and better perform deep reservoir detection,this paper studies the use of Stacked Denoising AutoEncoding(SDAE)method for seismic waveform clustering.The method combines low-level features to form more abstract high-level representation attribute categories or features,in order to discover the distributed feature representation of seismic waveform data,provide more abstract features for clustering,and improve the reliability and effectiveness of reservoir gas-bearing detection.The method also has the function of denoising.The main research contents and results of this paper are as follows:(1)On the basis of researching the basic theories of various deep learning networks,this paper focused on studying Stacked Denoising AutoEncoding(SDAE)algorithm.SDAE is superimposed by Denoising AutoEncoding(DAE),which is one of the unsupervised shallow learning models,while DAE is a type of AutoEncoding that accepts corrupted data as input and trains to predict the original uncorrupted data as output.SDAE not only has the ability of de-noising,but also can learn the potential features of the data from different hidden layers.Based on this,this paper completes the programming of the whole deep learning feature extraction model with deep learning framework——tensorflow and matlab.(2)The training of SDAE models were completed by adjusting the parameters of hyperparameters,hidden layers and hidden layer nodes.In order to avoid problems such as the disappearance of gradients in deep networks,deep learning layer-by-layer training is used.And the parameters such as reconstruction error,different hidden layer feature category discrimination and correlation are used to adjust hyperparameters,number of hidden layers and hidden layer nodes to train the best model.After a lot of experiments,the final parameters are determined as follows: learning rate is 0.001,the decay rate was 0.1,and the batch size was 32,the optimizer is Adam optimizer,and the activation function is sigmoid.Meanwhile,different hidden layer numbers and hidden layer nodes of both coding phase and decoding phase are also eventually determined,which complete a reasonable and effective network structure.(3)Selected seismic waveform time window dataset for SDAE model training.Compare and analyse results of SDAE model training with different time window data bodies.The results show that the narrow time window of 20 ms is easily affected by the interpretation noise and seismic data noise.It comes to us that the result of narrow time window has relatively less stability and larger variation range,while the wider time window of 40 ms and 60 ms are relatively stable.Therefore,it can concluded that wider time window of 40 ms and 60 ms can get more reliable results.(4)The methods of correlation analysis and feature visualization were studied for feature selection.Correlation analysis is used to select features that are highly correlated with the input data,which then are visualized to find features that are more representative of the specified categories,thereby performing cluster analysis.The features of the same category have similar feature activation,and target features are highly related to target samples.The higher the correlation,the more obvious the classification of hydrocarbon-bearing categories.(5)Compared different clustering results of different selected hidden layer features of SDAE and original data.The experimental results show that the effect of selected high-level features extracted by SDAE is better than that of shallow ones and original data,and the classification of hydrocarbon-bearing categories is also clearer.However,this does not mean that the characteristics of the shallow layer cannot be classified as hydrocarbon-bearing,and a comprehensive analysis is required.(6)Used the multi-GPU parallel computing and deep learning inference optimizer TensorRT to accelerate the training phase and the inference phase of the model respectively.The speed of the training phase is determined by the performance of the graphics card and the number of graphics cards.The more the number of graphics cards,the faster the training speed.In the inference phase,the model is converted to the FLOAT16 format using TensorRT.Under the same conditions,the inference speed can be increased by more than 3 times.