Research On Target Localization And Identification Method Based On Magnetic Gradient Tensor

The detection of targets using distortion potential fields generated around magnetic sources is an important technical means in geophysics,which is widely used in underwater and subsurface magnetic anomaly exploration with its advantages of wireless passive,low power consumption,strong concealment,and non-line-of-sight detection.As the latest generation of magnetic measurement technology,the magnetic gradient tensor(MGT)characterizes the minimum spatial change rate of the magnetic vector field along the orthogonal axis of the Cartesian coordinate system,which objectively and comprehensively characterizes all elements of the magnetic anomaly and has unique advantages in physical property inversion such as the positioning and boundary outline of anomaly target.With the deepening of geophysical exploration,the refined detection of small magnetic targets has become one of the frontiers of international research due to its commercial and military importance.However,this technique currently suffers from the superimposed coupling of platform magnetic interference,the heavy reliance on priori information about the number of targets,and the limited ability of manual feature expression and identification,which seriously limits its application in related fields.To solve the above problems,with the support of National Key Research & Development Program of the 13 th Five-Year Plan,this paper independently developed a set of High-Tc Superconducting Quantum full tensor magnetic gradient instrument system for the refined detection of underwater or underground small targets.The research work focuses on high-precision magnetic gradient tensor interference compensation,target position and parameter solution under blind environment,and target identification based on Two-stage Fine-grained(TSFG)instance segmentation network.These core technologies construct a relatively complete target magnetic sensing system from top to bottom.Firstly,magnetic compensation technology is used to suppress the interference magnetic field irrelevant to the target in the measured gradient tensor data of system.Then,basic attributes such as position and magnetic moment that reflect the magnetic characteristics of targets are obtained through analysis.On this basis,the depth characteristics of the target signal are mined based on Convolutional Neural Network(CNN)framework,and finally locate and identify targets.The main research content in this paper are as follows:(1)Aiming at the low accuracy of tensor measurement caused by the interference magnetic field associated with the system during measurement,a method of identifying and compensating the interference parameters of MGT based on rotation invariant constraint was proposed.Through in-depth analysis of the source,composition and its spatial distribution characteristics of the interference field,this paper introduces the Frobenius norm of gradient tensor as the constraint principle to establish an integrated compensation model for the interference error of the differential MGT.In order to solve the problem of low accuracy in parameter solution and poor compensation effect of this model,a group optimization estimation strategy combining Levy flight search mechanism and conversion probability is proposed.Both the simulation and field experiment results show that the proposed method overcomes the shortcomings of the conventional method that is sensitive to the initial iteration and easy to fall into premature.It can effectively suppress the interference magnetic field and provide reliable data for the follow-up magnetic target detection task.(2)Aiming at the difficulty of directly solving target localization in high-dimensional and strongly nonlinear parameter spaces in blind detection scenarios,a multi-target localization method based on unsupervised pre-detection mechanism was proposed.This paper constructs the mathematical model of gradient tensor observation array based on the equivalent magnetic dipole.The solvability condition of the equation and its explicit expression of general solution are re-derived,and Clustering-based i Forest is proposed as "pre-detector" to estimate the solution of targets.The theoretical model and the measured results show that compared to the conventional single fixed model,the proposed method avoids the error caused by human initialization,can automatically discover new targets and terminate vanishing targets under unsupervised conditions,and has a strong anti-noise ability.(3)Aiming at the problem that conventional methods have limited expression and recognition ability for targets,a recognition method based on R-CNN convolutional neural network depth feature representation is proposed.In this paper,Res Net residual block and transversally connected Top-Down convolution core of different sizes are used to improve the network’s learning ability to the shallow,deep and multi-scale features of targets.To solve the fuzzy mask label in high-frequency region,Point Rend rendering is used here.Finally,we design a Two-stage Fine-grained(TSFG)instance segmentation network model suitable for MGT target recognition.Through the verification of different types of target models,TSFG proposed in this paper achieves good results in the recognition and edge detection of seven typical targets.Compared with the popular Mask R-CNN benchmark model,the class average accuracy of TSFG is improved by 5.3%,F1 score is improved by 3.3%,and the predicted contour matches better with the actual situation of targets.