Investigations On Magnetic Anomaly Detection Based On Neural Network

Since the Second World War,magnetic anomaly detection(MAD)has been applied in the military field.In recent years,with the progress of the weak magnetic sensor technology,the MAD technology has been further emphasized and developed in the military application.The magnetic anomaly signal detection technology is one of the most critical technologies.Its essence is to judge whether there is magnetic anomaly signal from the complex measured magnetic signal.In this paper,neural network is proposed to be used as a classifier in MAD.The detection model of magnetic anomalies based on neural network is established,and the application of neural network in MAD is studied from two aspects: automatic feature extraction and artificial feature extraction.In this way,the detection theory of MAD based on neural network is preliminarily established,which will provide some theoretical and methodological support to the application of MAD in underwater target detection in the future.The main work and innovations are as follows:Firstly,the generation of magnetic anomaly signal is introduced,and the timedomain and frequency-domain characteristics of the anomaly signal are analyzed theoretically,which lays the foundation for the later design of the neural network model suitable for MAD.The essence of MAD is to judge whether there is magnetic anomaly signal from the complex measured signal,which is a binary classification problem.Therefore,the neural network is used as a classifier to identify whether there is the anomaly signal in the measured signal.The detection method of magnetic anomaly signal based on neural network is proposed,and the basic process of model training and signal detection in neural network is introduced.Then,a method of MAD based on automatic feature extraction is proposed,that is,the original data of the measured signal is used as the input of neural network directly,so that the network can automatically learn the data features from the original data that are conducive to classification.This paper introduces the data preprocessing methods such as sample set making and data normalization,as well as the optimization methods of neural network from three aspects: output layer’s threshold,neural network’s topology and activation function.The results of simulation and experiment show that the detection probability of the neural network detector based on automatic feature extraction(FCN-400)is higher than that of OBF detector under the colored Gaussian noise and the measured noise.For example,the detection probability of FCN-400 detector is about 20%higher than that of OBF detector when the magnetic moment direction is 120°and the CPA is 450 m under colored Gaussian noise with noise index of 0.5.Finally,to solve the problem that the features learned by neural network from raw data are not enough to distinguish the magnetic signal from noise,a neural network model based on artificial feature extraction is proposed.That is,the signal’s time-frequency features,statistical features and magnetic moment’s features extracted from the original data are taken as the input of neural network.Then,based on different feature fusion methods,we design two network structures: fully connected and partly connected network,to detect the magnetic anomaly signal.The results of simulation and experiment show that the detection probability of the fully connected(FCN-42)and the partly connected(PCN-42)neural network detector based on artificial feature extraction is higher than that of OBF under the condition of multiple groups of colored Gaussian noise and measured noise,which realizes more effective detection of magnetic anomaly signal.For example,under the colored Gaussian noise with noise index of 0.5,when the magnetic moment direction is 120 ° and CPA is 450 m,the detection probability of FCN-42 detector is 40% higher than that of OBF detector,and that of PCN-42 detector is 45%higher than that of OBF detector.