Research On The Detection And Classification Of Magnetic Anomaly Based On Magnetic Gradient Signals

Magnetic Anomaly Detection(MAD)is a covert passive detection method.MAD is widely used in earthquake early warning,mineral detection,submarine positioning,etc.When a ferromagnetic target is detected,it is often necessary to identify the type of ferromagnetic target,so the study of magnetic anomaly classification is also very important.Magnetic anomaly classification has outstanding application requirements in urban traffic,location and identification of unexploded ordnance.Therefore,thesis conducts research on magnetic anomaly detection and classification,and the main work is as follows:(1)Established mathematical models for magnetic anomaly signals and magnetic gradient anomaly signals.Studied the time-domain and frequency-domain characteristics of magnetic anomaly signals and magnetic gradient anomaly signals.Find the appropriate filtering range based on frequency domain characteristics.Using the wavelet packet filtering method to remove the non-stationary nature of 1/ f noise and weaken the correlation between wavelet coefficient sequences,a whitening filtering method based on wavelet packets is proposed,and the kernel smoothing method is used to verify that the signal after whitening filtering is close to a Gaussian distribution.(2)Improve traditional MAD methods.Based on the idea of jointly processing magnetic gradient anomaly signals and magnetic gradient background noise,a Full Magnetic Gradient Minimum Entropy Orthogonal Basis Function(FMG-ME-OBF)detection method is proposed.The MAD method is tested using simulated magnetic gradient anomaly signals and magnetic gradient anomaly signals collected through field testing.The results show that the detection method proposed in thesis can improve the detection success rate by at least 20% and the signal-to-noise ratio(SNR)by at least 1.9in low signal-to-noise ratios.(3)Aiming at the problem of insufficient samples collected in field test,a magnetic anomaly classification method based on convolution neural network and long short-Term memory(CNN-LSTM)is proposed.CNN-LSTM inherits the excellent feature extraction ability of convolution neural network(CNN)and the memory ability of long short-term memory(LSTM).This method first whitens and filters the geomagnetic background noise collected at the detection point;Then,the kernel smoothing method is used to estimate the probability density of geomagnetic background noise,and a simulated geomagnetic background noise is generated based on this probability density;Superposition the simulated geomagnetic background noise with the simulated magnetic anomaly signal to form a training sample;Then,the simulated magnetic anomaly signal is superimposed with the real geomagnetic background noise to form a test set.The training samples are randomly divided into training set and verification set,and the convolution short-term memory neural network is trained with the samples of the training set to obtain the classification model of magnetic anomaly signals.Finally,by classifying and recognizing the samples in the test set,the accuracy rate is higher than 90%,indicating good generalization.