Research On Aeronautical Electromagnetic Inversion Method In Frequency Domain Based On Deep Learning

Frequency domain airborne electromagnetic detection is a geophysical method for collecting different frequency electromagnetic signals from electromagnetic sensors mounted on an airborne mobile platform,and then processing data and studying the distribution of geological anomalies in underground space.It has the advantages of high efficiency,not limited by topographic and climatic conditions,and is one of the important technical means for geo-exploration in desert,gobi,mountainous areas and other areas with poor geomorphological conditions.It is widely used in mineral survey,environmental monitoring,engineering survey,geological mapping and other fields.Especially in the past 20 years,more frequency domain aero-electromagnetic exploration systems have been developed at home and abroad.These technological innovations have greatly improved the depth,accuracy and efficiency of geophysical data acquisition,thus bringing a new round of "revolution" in exploration technology.However,the large amount of geophysical multi-source data obtained for new technology and equipment,the inaccuracy of existing data processing algorithms,as well as the bottlenecks of high hardware requirements and slow computing speed,seriously restrict its application.Therefore,a deep learning based 1D inversion method of frequency domain aeroelectromagnetic is presented.In-depth learning,besides inversion prediction itself using the trained network model,mainly includes two important contents: sample dataset construction and network model design.(1)In the aspect of data set construction,a traversal and ordered geoelectric model based on Markov decision-making process is proposed first,and its electromagnetic response is calculated forward.Finally,a sufficient and typical sample data pair is synthesized,which provides necessary data support for deep learning training.(2)In network model building,this paper establishes a hybrid neural network model(CNN-LSTM)with convolutional neural network(CNN)nested long-term memory network(LSTM),which is used to establish the complex nonlinear relationship between frequency domain airborne electromagnetic response and resistivity model.It mainly uses the CNN network to extract the characteristics of the frequency domain airborne electromagnetic response as the encoding layer of the hybrid network,and uses LSTM network to further process the output information of the CNN network as the decoding layer of the hybrid network.In this paper,the mixed model completed by training is tested with theoretical and experimental data,and the results show that:(1)The method presented in this paper has high computational accuracy,good robustness and high computational efficiency,usually less than 1ms,can achieve realtime intelligent inversion.(2)The CNN-LSTM hybrid network model proposed in this paper not only has better inversion results than the traditional CNN model,but also has better anti-noise performance than the traditional CNN network model.(3)The test results of real data show that the CNN-LSTM hybrid network model proposed in this paper is more consistent with the inversion results of the traditional Occam method than that of the typical CNN network model,which indicates that the methods presented in this paper have higher reliability.By using CNN-LSTM model inversion imaging,we find that the results of Occam inversion method,which is commonly used in 1D models,are less different and have the possibility to be closer to actual geology than Occam inversion method.Secondly,CNN-LSTM is computationally efficient and can be used for inversion imaging of underground resistivity in real time.In addition,because sufficient and typical sample data are used to train the CNN-LSTM model,this network model is not limited to a specific area,and it has a certain universal applicability.