对数正态分布双变点的DL-MCMC混合检测框架OA

Change-point detection in lognormal distributions holds significant application value in fields such as finance,industrial monitoring,and biomedicine.However,traditional Markov Chain Monte Carlo(MCMC)methods suffer from issues such as sensitivity to parameter initialization and slow convergence,which limit their practical performance.This paper proposes an innovative hybrid detection framework—the Deep Learning-MCMC(DL-MCMC)hybrid detection framework.It combines the temporal feature extraction capabilities of one-dimensional convolutional neural networks(CNN)and bidirectional long short-term memory networks(BiLSTM)with the statistical inference advantages of MCMC.In the optimization stage,a Metropolis-Hastings(M-H)algorithm with adaptive window width is employed to dynamically adjust the sampling range,enabling efficient and accurate detection of double change-points.Experimental results demonstrate that the proposed framework effectively addresses the sensitivity of traditional MCMC to initial parameter values,providing a new approach for multiple change-point analysis in complex time-series data.