Study On Transient Electromagnetic Inversion Method Based On Improved Extreme Learning Machine

Transient electromagnetic method(TEM)is an important geophysical exploration method,which is widely used in the fields of groundwater exploration,oil and gas exploration,mineral resources exploration and shallow geological survey.In recent years,with the rapid development of deep learning,the traditional neural network method based on gradient learning has been widely applied in the inversion and interpretation of electrical exploration data,but it also has the defects of slow convergence speed,fall into local optimum easily and long computing time in the TEM inversion training.Therefore,a nonlinear inversion algorithm based on extreme learning machine(ELM)is proposed in this paper,which is a non-iterative learning method with fast learning speed and strong generalization performance.The ELM method can avoid the defects of the gradient learning,which can meet the needs of high-precision,high-stability,and high-efficiency inversion in geological exploration.Additionally,the theoretical research is performed on the application of ELM method in TEM inversion,and the optimization algorithm,modeling method and inversion process of ELM method are analyzed in this paper.(1)In the observation data collection,TEM data usually contain uncertain information such as noise data,which leads to the poor inversion accuracy.Therefore,a hybrid approach combining wavelet packet denoising(WPD)technique and a regularized ELM inversion model based on the leave-one-out cross-validation(ELM-LOO)is proposed in this paper.Firstly,the noise information in the TEM data is suppressed by WPD;then,the regularized ELM is introduced to further enhance the anti-interference performance of the inversion network;Furthermore,considering the key role of regularization factor in inversion,LOO methodology is used to optimize the regularization parameter in ELM to improve the inversion accuracy and generalization performance.Finally,the numerical calculation and model simulation show that the WPD-ELM-LOO method can effectively reduce the noise data and achieve high inversion accuracy and generalization,which can be regarded as an efficient and feasible TEM inversion technique that is superior to the traditional neural network and ELM methods.(2)Due to the structural limitations of the ELM algorithm that the input weights and hidden layer thresholds are randomly assigned,which leads to the instability of the network output results.Therefore,a hybrid inversion method for dimension-compressed online sequential extreme learning machine(OSELM)model based on improved chaotic differential evolution(ICDE)training is presented in this paper.Firstly,the global differential evolution(DE)algorithm is applied to optimize the initial weights and thresholds of the ELM to improve the stability of the proposed inversion network;Then,the Tent chaotic mapping technique with constraint factors is utilized to optimize the mutation factor F and the crossover factor CR to promote the global optimization ability of the DE algorithm,thereby enhancing the generalization degree of the inversion model.Moreover,the online batch learning strategy is embedded into the ELM training,and the high-dimensional input data is preprocessed with the kernel principal component analysis(KPCA)technology to enhance the compactness of the inversion network structure and obtain better inversion performance;Finally,the numerical calculation and model simulation show that the inversion performance of the proposed method is better than other improved ELM methods,which achieves better inversion accuracy,inversion stability and generalization performance.(3)In practical applications,there is usually collinearity problem between TEM data,which easily causes multicollinearity in ELM inversion and always affects the calculation efficiency of inversion.In order to solve this problem,an ELM hybrid method(F-ELM)based on fractal dimension technology is studied in this paper,which uses fractal dimension technology to retrieve the information of the hidden layer output matrix for sparsely selecting the neuron nodes of the hidden layer,which not only effectively solves the multicollinearity problem but also simplifies the ELM network structure and greatly improves the computation speed.Additionally,the introduction of fractal dimension technology is able to subtly avoid the difficulty of selecting the optimal hidden layer neuron node of ELM,which further enhances the applicability of the inversion network model.The numerical calculation and model simulation show that the F-ELM method achieves better computation speed,inversion stability and generalization in TEM inversion,especially in terms of computational efficiency.(4)The field experimental verification of nonlinear TEM inversion using the Xishan Village landslide area in Sichuan Province.Firstly,based on the previous research on inversion algorithm and combined with the needs of actual engineering exploration,an ELM comprehensive inversion method is proposed;Then,without providing any prior information,the ELM is employed to directly invert the field data,which results in a poor inversion imaging effect;Subsequently,the inversion results of the OCCAM method are used to reconstruct the training samples of the ELM,and the inversion research is conducted using the proposed method above.The experimental results show that the strategy can effectively improve the inversion quality and obtain satisfactory imaging results.The ELM inversion based on the OCCAM inversion results is able to effectively utilize the prior information that is implicated in the geoelectric model,which can incorporate the features containing actual geological data into the inversion process,thus providing an important theoretical basis for the structural information of the underground medium of the landslide body,and also proving the validity and practicability of the proposed inversion method.