Study On Slope Displacement Prediction Based On Multi-Source Domain Transfer Learning For Small Samples

The occurrence of mine landslide will seriously affect the development of national economy and the safety of people’s lives and property.The deformation variable of slope displacement can well characterize the development process of slope deformation,and can also make an early warning for the occurrence of landslide disasters.In practical slope engineering,the displacement deformation of slope is influenced by many factors,and the displacement monitoring equipment has some problems,such as regular maintenance,small sample size of displacement data;and poor prediction accuracy of deep learning network.Therefore,the transfer learning method is used to establish a slope displacement prediction model to make early warning for landslide disasters,and the proposed network is verified and evaluated.Taking the data provided by a mine slope monitoring project in Shijiazhuang,Hebei Province as an example,the study is carried out.The main research contents are as follows:(1)The displacement and deformation of the slope are influenced by many factors.Mutual information(MI)and Pearson correlation coefficient(PC)are used to determine the key influencing factors of slope displacement.Compared with decomposition methods such as wavelet transform and empirical mode decomposition,the variational mode decomposition method can accurately match the physical meaning of each component.After preprocessing,the displacement data is decomposed into trend,periodic,and random components using an improved variational mode decomposition method.Auto-regression models are used to predict the trend displacement component,and a long-short term memory network model is employed to predict the periodic displacement component.Due to the difficulty in predicting and analyzing the random displacement component that is affected by many external factors,a generated adversarial network(WGAN)based on Wasserstein distance is utilized for prediction.The experimental comparison results of the random displacement component show that the WGAN network model is superior to traditional prediction algorithms and has strong robustness.(2)In practical mining slope monitoring,data loss occurs when monitoring equipment performs long-term data collection,and the prediction effect of deep learning networks is poor when the sample size is small.To solve this problem,a multi-source domain transfer learning method is introduced,and a prediction model named MDTF-WGAN is proposed to combine multi-source domain transfer learning and WAGN network models,which effectively avoids the “ negative transfer ”phenomenon of single-source domain transfer learning.The experimental comparison results show that the MDTF-WGAN prediction model can better fit the real values,and the trend of the predicted random displacement component is similar to that of the real values.(3)In practical engineering applications,the data collected in different scenarios can provide more abundant generalization knowledge,but the difference between the source domain and the target domain is too large,which leads to poor transfer performance.In this thesis,local maximum mean difference(LMMD)is introduced to match the distance of sub-domains from the perspective of local adaptation,so as to realize more accurate transfer of knowledge and effectively improve the transfer performance of multi-source domain transfer learning.Compared with the prediction model combined with multi-source domain transfer learning and WGAN network,the prediction accuracy of random term displacement is slightly improved after domain adaptive matching,which has certain practical significance for landslide displacement prediction.