Complex Data Prediction Based On Variational Mode Decomposition And Deep Forest

Complex,nonlinear and non-stationary data analysis has become a hot topic in the field of modern statistics and data science.It is of great theoretical significance and application value to explore adaptive data decomposition,variable selection and statistical learning methods for complex time series data.How to mine the key information of data and establish efficient prediction methods to achieve accurate inference is an important and challenging research.The paper takes nonlinear,non-stationary and high-dimensional time series data as the research object,and takes adaptive decomposition and machine learning theory as the core to build a new deep forest method,which provides a strong theoretical support for enriching and developing data science methods.The main research contents of this article are as follows:(1)The regularized variational mode decomposition method optimized by the improved Harris Hawk Optimization algorithm is proposed.The ridge regression penalty is used to construct variational equation in variational mode decomposition.The accuracy of reconstruction after decomposition is not good at the cost of giving up the unbias and reducing the precision.In order to solve the issue,a regularized variational mode decomposition method is proposed.Based on the L2 penalty in the Wiener filtering process of variational mode decomposition,the L1 penalty term is introduced to improve the efficiency and accuracy of the decomposition.In this paper,the decomposition balance factor determined by the entropy of envelope spectrum and the reconstruction error is used as the fitness function of the Harris Hawk algorithm to establish the optimal variational mode decomposition.(2)The variable selection method of adaptive elastic network with minimum common redundancy and maximum relevance is constructed.In view of the nonlinear and highdimensional attributes of the data and the dependence between variables,the paper combined the coefficient compression and mutual information theory,proposed the adaptive elastic network method of minimum common redundancy maximum relevance,the penalty term of the adaptive elastic network is weighted.The minimum common redundancy maximum relevance criterion is used to calculate weight.The objective is to explore a variable selection method based on data-driven model-free hypothesis.The method fully considers the redundant information between candidate variables,selected variables and target variables,and achieves the purpose of controlling redundant variables and selecting related variables.(3)A two-stage weighted deep forest method is proposed.It is proposed to solve the problems of weak heterogeneity and the same weight in the training process of deep forest algorithm.In the multi-grained scanning stage,the enhanced features generated by windows with different granularity are weighted to reflect the contribution of different scanning granularity and enhance the feature representation ability.In the cascade forest stage,different learners are constructed from the perspective of increasing the diversity of classifiers.At the same time,weights are assigned to the cascade classifiers to reduce the negative impact of learners with poor fitting performance and enhance the role of strong learners.This work takes complex,nonlinear,non-stationary and high-dimensional time series data as the research object.The regularized variational mode decomposition method is optimized by the improved Harris Hawk algorithm,the variable selection method of adaptive elastic network based on the minimum common redundancy maximum relevance criterion,and the two-stage weighted deep forest classification algorithm to establish a complete data prediction framework.In the process of each algorithm,simulation data is used to verify the effectiveness of the proposed method from different aspects.The method is applied to the biomedical epilepsy EEG recognition and the rolling bearing fault diagnosis cases in the field of mechanical engineering.In the empirical analysis,the comparison experiment with the traditional model is established.The superiority of the proposed method is proved by statistical test.The proposed model can be applied to the analysis of time series data in other fields.The work of the paper provides theoretical support for the analysis of complex,nonlinear and non-stationary time series data.