Research On Magnetic Target Recognition Algorithm Based On Magnetic Gradient Tensor

Magnetic anomaly detection is a very important method in geophysics.Because magnetic gradient tensor detection has more advantages than traditional detection methods,it has gradually become a research hotspot in the field of magnetic anomaly detection technology in recent years.Although magnetic gradient tensor detection has been widely used in the detection of large amphibious magnetic anomalous targets,it is not accurate enough to detect the shape and attitude of small magnetic targets such as ferromagnetic pipes and unexploded ordnance.And in the process of interpreting the magnetic target,it is necessary to select a suitable positioning method according to different shape and posture information.For example,the wrong choice of the shape of the target will directly cause the positioning accuracy to drop.For the past few years,machine learning has been widely used in magnetic anomaly recognition because they can reduce the dependence on signal accuracy and avoid a large number of calculations and derivation processes for interpreting magnetic data.In summary,this paper studies a set of magnetic target recognition algorithms based on a magnetic gradient tensor system built by four vector magnetic sensors for the shape recognition of small underground magnetic targets.Firstly,the basic principles of magnetic gradient tensor is introduced briefly in this paper.The measurement system of the magnetic gradient tensor with different structures is simulated and compared,then uses the best one to design a measurement system based on four orthogonal magnetic sensors with three-axis is designed to measure the disturbance signal generated by the magnetic target.The gradient tensor measurement system measures the disturbance signal generated by the magnetic target.Then,the measurement error of a single magnetic sensor is analyzed and the error mathematical model is established.And the linear neural network method is used for correction.The recognition of detected features based on intelligent learning algorithms can be divided into two processes,feature dimensionality reduction and pattern recognition.First,principal component analysis(PCA)and stability competitive adaptive reweighted sampling(SCARS)are used to reduce the dimensionality of the original magnetic anomaly data.In this paper,an identification extreme learning machine method is proposed and a probabilistic neural network method is introduced to intelligently identify the features after dimensionality reduction.In the end,PCA-DELM achieved an accuracy of 36.67%,PCA-PNN is 42.67%,and SCARS-DELM is 93.33 %,SCARS-PNN is 89.33%.This paper also introduces the deep learning theory,with the help of convolution and pooling operations,realizes the deep mining of the original magnetic anomaly data characteristics,and uses the fully connected layer to achieve decision classification.In order to reduce the amount of calculations and parameters in the convolution process,the Ultra-Lightweight Subspace Attention Mechanism(ULSAM)is introduced to realize the channel convolution of attention,and then the design can perform the original feature the ULSAM-CNN framework structure,which is deeply excavated and recognized with high precision,finally obtains 97.33% recognition results under the deep learning theory.