Machine Learning-based Noise-separation And Inversionimaging Method In Distributed Full-waveform Induced Polarization Exploration

Induced polarization(IP)is a geophysical branch method for detecting conductivity and chargeability of geological bodies.In recent years,the distributed full-waveform electrical exploration instruments and equipment have been developed in China and abroad,and the efficiency of IP data acquisition has been developed rapidly.This paper is aiming to establish a set of preliminarily intelligent noise-separation and inversion imaging methods for large-scale data generated by distributed IP exploration,so as to improve the data quality and application effect.To improve the accuracy and efficiency of distributed IP data processing,this paper develops an antiinterference technique based on de-noising library and statistical decision.Firstly,the forwarding of IP fullwaveform response in 3D medium is realized.By analyzing the characteristics of simulated theoretical IP signals waveform and various interferences,eight statistical parameters in time and frequency domain that can best represent the types of time series are extracted.Then we generate a IP signal library and a noise library,and realize the machine's judgment and recognition of different noise interferences in time series through the support vector machine(SVM)classification algorithm.Then,by learning the knowledge in the field of signal processing,we selected and improved five targeted signal processing technology,including: empirical mode decomposition,waveform matching,robust estimation,principal component analysis and wavelet analysis.They are integrated into a de-nosing library and the decision tree automatically selects the appropriate technology to suppress noise interference.The above method is an automatic anti-interference algorithm based on statistical analysis and signal processing knowledge.To solve the problems of the conventional quasi-linear optimization algorithm including initial model dependencies and low resolution,this paper improves two machine learning algorithms,the samplecompressed artificial neural network algorithm and the adaptive clustering analysis algorithm.They are respectively applied to the IP inversion and boundary demarcation.Firstly,the random medium model is used to generate electrical model samples,and the theoretical responses are generated by distributed computing.Then,the neural network model is trained using the theoretical models and responses to predict the new data.To reduce data redundancy,this paper combines data compression technology with artificial neural network to reduce the input and output sample dimension and improve the speed and accuracy of neural network inversion.Additionally,it is also an important work in data interpretation to identify the boundaries of abnormal bodies according to the inversion results.In this paper,clustering analysis is improved to automatically determine the number of clusters according to the distribution characteristics and robust statistics of the original data.The adaptive clustering is used for attribute clustering,boundary demarcation and anomaly center positioning.The above methods further improve the accuracy and automation level of IP inversion imaging.Finally,the methods presented in this paper are applied to the practical IP data of a lead-zinc polymetallic mining area in southwest China,and the high quality IP data of over 5000 measuring points are obtained.The anti-interference effect of different electrode spaces and different frequencies are analyzed.The electrical survey and sounding data before and after anti-interference processing are compared.The new inversion imaging algorithms are also used to carry out the inversion of planar IP parameters,two-dimensional electrical sounding and three-dimensional multi-profiles based on the measured waveform data.The boundary demarcation and attribute clustering of the inversion results are also carried out.The data processing results reflect the abnormal characteristics of the resistivity and polarizability of the underground medium in the survey area.Combining with the geological data of the survey area,the favorable areas for mineralization are deduced and the algorithms are verified.In summary,in order to improve the data quality and application effect of distributed full-waveform IP exploration,two comprehensive algorithms are studied in this paper,including the anti-interference algorithm based on de-noising library and statistical decision,and the inversion imaging algorithm based on samplecompressed neural network and adaptive clustering.The testing results of simulated and practical data show that the new algorithm can effectively improve the quality of IP data,enhance the ability of observation data to detect the underground abnormal body,and improve the accuracy and automation level of data processing and interpretation.The framework and algorithm in this paper can be further migrated to other artificial source electromagnetic exploration methods.