Research On Microseismic Signal Recognition Based On Online Multi-Kernel Boosting Learning

Microseismic signal recognition is an important work in unconventional oil and gas development.Realizing rapid and accurate identification of effective microseismic signals has important guiding significance in the field of oil and gas reservoir exploration and development.Traditional recognition methods mainly rely on pattern recognition algorithms under offline batch processing,without considering that the microseismic data monitored by the geophone is a continuous real-time data stream with time-varying characteristics,prone to conceptual drift,and due to the impact of noise,there is a problem of category imbalance between effective signals and noise signals.The introduction of online kernel learning,integrated learning,and other technologies has realized the automatic recognition of microseismic signals,effectively improving the recognition efficiency.This paper proposes an online multiple kernel boosting model based on data streams with imbalanced class distribution,and solves the problem of microseismic signal recognition by this model.The main work is reflected in the following aspects:(1)An online kernel learning algorithm SA-CNOGD for imbalanced data streams is proposed.Introducing a priori information about the class distribution of sample data,this paper proposes an adaptive weighted online kernel learning method that automatically adjusts the penalty parameters of arriving training samples to learn imbalanced data streams online.At the same time,in order to improve the learning efficiency of large-scale imbalanced data streams,first of all,combining Nystr?m and the principal component Cholesky decomposition iterative algorithm,a new approximate method,the CNystr?m method,is proposed.The approximate error bound of the algorithm is analyzed theoretically,and experimental results verify the rationality and computational efficiency of the CNystr?m method.Then,considering the idea of Nystr?m online gradient descent(NOGD)algorithm,a new online kernel learning framework for unbalanced data streams(SA-CNOGD)is proposed by applying the proposed CNystr?m method to kernel matrix approximation and combining adaptive weighted online kernel learning.Finally,the regret bound of the algorithm is analyzed theoretically,and the performance of the algorithm is verified experimentally.(2)An online multi-kernel boosting learning algorithm MK-SA-CNOGD based on class imbalanced data streams is proposed.This algorithm is based on the idea of online boosting algorithm,using the SA-CNOGD algorithm as the base learning algorithm.At the same time,a new kernel alignment method,named normalized kernel alignment,is defined using the geometric metric method of normalized kernel functions by introducing prior information about the two types of class distribution of samples.Applying the normalized kernel alignment method as the basis kernel function evaluation method for MK-SA-CNOGD,that is,selecting the kernel function that has the largest ratio of normalized parity values to the sum of normalized parity values of all selected kernel functions as the final selected kernel function to further improve the performance of the MK-SA-CNOGD algorithm.Experiments on UCI datasets verify the performance of the algorithm.(3)Extract six signal characteristics of microseismic signals,including kurtosis coefficient,short time zero crossing rate,wavelet packet coefficient Shannon entropy,wavelet packet coefficient energy ratio,amplitude factor,and waveform factor.Apply Spearman correlation coefficient to feature selection for these different signal characteristics,and apply MK-SA-CNOGD algorithm to establish a microseismic signal recognition model.