| Magnetic Anomaly Detection(MAD)is an important technology in geological exploration,marine survey,and defense military.In the aeromagnetic anomaly detection project,the aeromagnetic measurement platform contains a large amount of ferromagnetic material,which will generate a large interference magnetic signal during aviation operations,and the interference magnetic signal can submerge the target magnetic anomaly signal easily.Therefore,the key to detecting the target magnetic anomaly signal accurately is whether the platform interference magnetic field in the signal is compensated accurately.This process is known as aeromagnetic compensation technology.The aeromagnetic compensation technology can be divided into two categories: “hard compensation” and “soft compensation”.Among them: “hard compensation” is the use of electronic hardware devices such as coils to generate magnetic signals that are opposite to the interference magnetic field for Elimination;“soft compensation” is based on computer technology,where signal processing algorithms are used to calculate the interference magnetic field to achieve compensation.At present,the "soft compensation" technology with high precision,flexibility,programmable and automation is gradually replacing the "hard compensation" technology with poor versatility and high cost,and has become the first choice for aeromagnetic compensation.The aeromagnetic compensation technology studied in this paper is a modeling method based on the measured signal,attitude information and other data of the platform to perform mathematical regression on the interference magnetic field,which belongs to the category of “soft compensation”.The main research work of this paper is as follows:Firstly,the mathematical expression of the maneuvering magnetic field is derived in the flight platform coordinate system based on the Tolles-Lawson model.The improvement of the geomagnetic time-invariant hypothesis in the traditional compensation method is proposed: the autoregressive moving average of the background geomagnetic signal is made.The ARMA(Autoregressive moving average)model of geomagnetic field is modeled to obtain time series features,and then the Kalman filter is used to separate the geomagnetic signals from the total magnetic field to construct the basis function.From the experimental results,the background magnetic field obtained by Kalman filtering is more accurate than the total field average,and the compensation effect is better.Secondly,this paper proposes a calibration scheme for the low accuracy of the threeaxis fluxgate in magnetic compensation.Firstly,the types of error factors are analyzed,and the ideal and non-ideal three-axis coordinate systems are established.The mathematical expressions of non-orthogonality,sensitivity and zero bias are derived.The measurement data using high-precision optical pump magnetometer is combined with the confidence domain value.The optimization method is used for calibration;the robustness of the calibration algorithm and the compensation effect before and after the correction are verified by the experimental data,and the influence of the size of each element of the error coefficient matrix on the compensation effect before and after the correction is analyzed.Thirdly,a deep learning neural network is designed to train the magnetic compensation model.Firstly,using the normal Dense Neural Network(DNN),the threecomponent magnetic field and the total magnetic field are used as input data,the interference magnetic field is used as the model applicability of the output data,and the method of adding Gaussian noise to expand the data size is proposed for the insufficient data of the calibration flight acquisition.The effectiveness of the method is verified by experimental comparison.Then,the deep Recurrent Neural Network is selected in the construction of the compensation model for long time span,and the time variable function is constructed to explain the cyclical trend of the geomagnetic field.Finally,the LSTM and GRU network training indicators are analyzed and compared.The final experimental results also demonstrate the long-term generalization ability of deep Recurrent Neural Networks. |
PDF Full Text Request