探潜是现代海洋军事中的重要主题，航载磁探潜是现今探潜问题中的研究和应用热点。航载磁探潜中，由于探测距离较远、相对移动速度较快和潜艇磁隐身技术的发展，探测器接收的潜艇磁信号具备“微弱”和“瞬态”特征。因此，航载磁探潜的核心科学问题是微弱瞬态潜艇磁信号的检测问题。潜艇磁信号可分为磁异信号和极低频磁信号，潜艇处于不同状态时可对外表现不同的磁信号特征，并且其信号模型会出现差异和未知参量。因此，针对不同种类且信号模型差异的微弱瞬态潜艇磁信号和不同的应用要求，进行4方面研究，实现对微弱瞬态潜艇磁信号的快速、全面、可靠检测，以及目标磁信号定位。对于微弱瞬态潜艇磁异信号的检测。由于探测器与潜艇的相对位置未知，潜艇磁异信号模型中出现一个未知的隐变量，它直接影响信号的估计与检测。因此，提出非概率分布的期望最大化（Non-Probability Distribution Expectation Maximization，NPD-EM）算法对潜艇磁异信号进行估计，以较优量替代最优量，克服传统期望最大化方法中后验概率未知的问题，并提出自带噪声补偿的NPD-EM检测器，实现对含隐变量微弱瞬态潜艇磁异信号的高性能估计与检测，信噪比可低至-34dB。对于微弱瞬态潜艇极低频磁信号的检测。由于该信号的多样性，其幅值、频率和相位时变且未知。因此，提出基于线性绝对值核函数支持向量机（Linear-absolute-value Support Vector Machine, L-abs SVM）的检测器——L-abs SVM检测器，从“微弱瞬态潜艇极低频磁信号始终是一定约束条件下的短时正弦信号”这一观点出发，以机器学习和核函数投影等方法，实现对微弱瞬态潜艇极低频磁信号的恒定低虚警检测。信噪比低至-11dB时，虚警概率保持为0，检测概率达到1。进一步提高探测器的通用性和快速性，提出基于正弦高斯混合模型的累积和（Sine-Gauss-Noise Cumulative Sum, SinGN-CUSUM）检测器，以实现对微弱瞬态潜艇磁异信号和极低频磁信号的通用快速检测。该方法以正弦高斯混合模型对潜艇磁信号进行通用建模，以累积和检测方法构建检测器，实现对微弱瞬态潜艇磁信号的通用快速检测，信噪比可低至-15.5dB，计算量相比于传统检测方法降低4个量级。最后，分析以上所提各方法的特点，构建基于多检测方法性能互补的联合检测系统。可同时发挥NPD-EM检测器的可靠性、L-abs SVM检测器的恒定低虚警特性和SinGN-CUSUM检测器的快速性，在保证最优检测性能不变的前提下，大幅降低计算量，实现实时检测和目标磁信号的定位。

Submarine detection is very important in the modern navy. The airborne magnetic submarine detection is the focus of recend researches and applications. In the airborne magnetic submarine detection, the magnetic signal of the submarine received by the detector is weak and transient, because of the long detection distance, the high relative speed between the detector and the submarine, and the development of the magnetic silencing technique of the submarine. So, the essential scientific problem in the airborne magnetic submarine detection is the detection of weak transient magnetic signals from submarines.The magnetic signals from submarines can be devided into two varieties according to their generating mechanism and their frequence characteristic, that are the magnetic abnormal signal and the extremely low frequency magnetic signal. The submarine would appear different magnetic signals on different conditions. And the models of the magnetic submarine signals would be different and appear some unknown parameters. Therefore, there are four researches in this thesis to detect different weak transient magnetic signals from submarines and to deal with different requirements in applications. Then, the fast overall reliable detection of weak transient magnetic signals from submarines and the signal location can be realized.In the detection of weak transient magnetic abnormal signals from submarines, there is an unknown latent variable in the signal model, because of the unknown relative position between the detector and the submarine. The latent variable would impact the estimation and detection of the magnetic abnormal signal from a submarine. Because of this, a non-probability distribution expectation maximization (NPD-EM) algorithm is proposed to estimate the weak transient magnetic abnormal signal from a submarine, in which the optimal value is relaced by a sub-optimal value and the problem caused by the latent variable in the signal model is overcome. A NPD-EM detector is also proposed to detect the weak transient magnetic abnormal signal from a submarine, which is based on the NPD-EM algorithm and has a noise compensation. The NPD-EM algorithm and the NPD-EM detector can work with excellent performance when the signal-to-noise ratio is only -34 dB.In the detection of weak transient extremely low frequency magnetic signals from submarines, the extremely low frequency magnetic signals from submarines are various, that the amplitude, frequency and phase are time-varying and unknown. Because of this, a support vector machine based on the linear-absolute-value kernel function (L-abs SVM) and a detector based on the L-abs SVM (L-abs SVM detector) are proposed. The L-abs SVM detector is based on the idea that every weak transient extremely low frequency magnetic signals from submarines must be the short-time sinusoidal signal with some constrain conditions. With the theories and techniques of machine learning, kernel function and hyperplane, the L-abs SVM detector can detect the weak transient extremely low frequency magnetic signal from a submarine at a low constant false alarm rate with the signal’s amplitude, frequency and phase unknown and time-varying. When the signal-to-noise ratio is over -11 dB, the detection probability can be 1, and the false alarm ratio is 0.In order to improve the universality and rapidity of the detection system, a sine-Gauss-noise cumulative sum (SinGN-CUSUM) detector is proposed. The SinGN-CUSUM detector is constructed by the cumulative sum method of the sequential detection theory, and a sine-Gauss mixture model is built for both the magnetic abnormal signal and the extremely low frequency magnetic signal from a submarine, which realizes the adaptive modeling for the magntic submarine signals. The SinGN-CUSUM detector can realize the universal fast detection for both the magnetic abnormal signal and the extremely low frequency magnetic signal from a submarine when the signal-to-noise rate is -15 dB, and the calculation is reduced 4 orders compared with the traditional detection method.Then, a united detection system with the proposed detectors is constructed. The united detection system can detect signals reliably as the NPD-EM detector, be a low constant false alarm rate as the L-abs SVM detector and detect signals rapidly as the SinGN-CUSUM detector. The united detection system reduces the calculation but gurantees the detection performance. The real-time detection and signal location are realized.