Aeromagnetic Data Preprocessing Method Based on Deep Learning

To conduct airborne magnetic surveys, compensation flights must be carried out first. In the compensation flights, it is assumed that the regional geomagnetic intensity is constant. However, the actual recorded compensation signal will be contaminated by the regional gradient of the magnetic field and other factors. In addition, the process of solving for the compensation coefficient is actually the process of solving for the overdetermined equations. But the coefficient matrix of the 16-factor coefficient equation has serious multicollinearity. In the past decade, scholars have proposed a number of methods to deal with noise and multicollinearity problems. However, the traditional method is not satisfactory when dealing with complex data. It is hard for UAVs to make routine compensation flight like the manned aircraft do. The purpose of this study is to reduce the error caused by the redundant information in the compensation data. Therefore, we propose an unsupervised denoising and feature extraction algorithm, Denoising Autoencoder (DAE). We start with the simplest three-layer Autoencoder: established a network with the input layer, the hidden layer and the output layer (The respective dimensions are 16, 9, 16). We applied it to the rotary-wing UAV detection platform. A compensation flight was conducted at a test flight site in southern China. The Autoencoder based on deep learning can extract the useful and frequent high-order features in the compensation data to obtain the hidden layer with 9 dimensions. The neural network weights are initialized to a better distribution, and the labeled data is used for supervisory training. Finally, the data is reconstructed from the extracted high-order features to obtain output layers with 16 dimensions. Based on the reconstructed data, the least squares method is used to perform interference prediction. The results show that the DAE feature extraction can eliminate the noise and can solve the variable multi-collinearity problem. Furthermore, it can improve the prediction accuracy of the compensation algorithm. The attempt of the three-layer Autoencoder (16-9-16) were prepared for the subsequent establishment of a multi-layer Autoencoder (16-9-K-9-16, K<9).