Marine Target Magnetic Anomaly Signal Denoising and Detection Method Based on Deep Transfer Learning

Abstract

The present application discloses a method for denoising and detecting magnetic anomaly signals of ocean targets based on deep transfer learning, including: constructing a simulated magnetic anomaly signal dataset of ocean targets using a magnetic dipole model; establishing a measured ocean magnetic field noise dataset; adding the data in the measured ocean magnetic field noise dataset to the simulated magnetic anomaly signal dataset of ocean targets to form a simulated noisy magnetic anomaly dataset of ocean targets; training a denoising autoencoder network using the simulated noisy magnetic anomaly dataset of ocean targets; training a fully connected classifier network using denoised magnetic field data; collecting magnetic field data to be detected, and performing denoising and detection using the trained denoising autoencoder network and fully connected classifier network to obtain the final magnetic anomaly signal detection result. The present application can train a deep neural network model with a small amount of measured data, and can achieve faster magnetic field data denoising and magnetic anomaly signal detection compared with traditional signal processing methods.

Description

Technical Field

[0001] The present application relates to the field of signal and information processing technology, and in particular to a method for denoising and detecting magnetic anomaly signals of marine targets based on deep transfer learning. Background Art

[0002] Materials with ferromagnetic materials will be affected by the permanent magnetism and the magnetization of the surrounding geomagnetic field, causing the geomagnetic field around the material to be locally distorted and deformed, thereby generating a quasi-stationary magnetic phenomenon, which is called a magnetic anomaly signal. As an important physical field source for marine targets, magnetic anomaly signals are not restricted by marine topography, marine environment and other conditions, and have high sensitivity and resolution. They are compatible with a variety of electronic instruments and can provide valuable signal sources for ship navigation, direction identification, etc.

[0003] At present, there are two main types of detection methods for marine target magnetic anomaly signals: one is to theoretically model marine target magnetic anomaly signals, analyze their signal and energy domain characteristics, and then determine the presence or absence of target magnetic anomaly signals in the characteristic domain; the other is to use the signal characteristics of the background magnetic field as the starting point, analyze the statistical distribution characteristics of the marine background magnetic field noise, and explore the characteristic differences between it and the target magnetic anomaly signal, and use noise filtering and suppression methods to enhance and detect the target magnetic anomaly signal. Commonly used magnetic anomaly signal detection methods include orthonormal basis function (OBF), minimum entropy detector (MED), wavelet transform (WT) denoising, etc.

[0004] OBF is a detection method based on the characteristics of target magnetic anomaly signals. According to the physical model of target magnetic anomaly signals, the magnetic anomaly signal sequence can be linearly represented by three mutually incoherent basis functions. After orthogonalization and normalization, etc., the basis functions can be transformed into OBF. On this basis, the corresponding magnetic anomaly signal representation coefficients are calculated according to the properties of the basis functions, and based on this, the existence of target magnetic anomaly signals in the time series is judged. MED is a magnetic anomaly signal detection method based on information entropy characteristics. Compared with the pure ocean environmental magnetic field noise, the target magnetic anomaly signal usually has higher orderliness, and information entropy is an index to measure the strength of signal orderliness. Since the appearance of the target magnetic anomaly signal affects the disordered distribution of the ocean magnetic field noise, the magnitude of the information entropy of the time series changes accordingly. Therefore, the measured ocean magnetic field time series can be transformed into the corresponding information entropy sequence, and the threshold is set by observing the change trend of the information entropy, so as to realize the detection of target magnetic anomaly signals. WT is a signal analysis method with the characteristics of removing correlation, being applicable to multiple resolutions, and being able to cope with time-frequency local changes, etc. The core of the denoising effect of the wavelet threshold algorithm lies in the setting of the threshold and the construction of the threshold function. The method based on wavelet transform is widely used in the field of denoising. After denoising, it is easier to judge whether the data contains magnetic anomaly signals according to its signal characteristics.

[0005] The above magnetic anomaly signal detection methods rely on prior knowledge and make ideal assumptions about signal characteristics, which restricts their application effects in practical problems. The OBF detector can only work normally in the Gaussian white noise environment, seriously restricting its application in complex ocean noise environments. MED distinguishes target magnetic anomalies based on the information entropy difference between signals and noise, but its discrimination ability is relatively limited when the signal-to-noise ratio is low, and it cannot meet the long-distance detection requirements of target magnetic anomalies. WT relies on the different characteristics of signals and noise in the wavelet function space to suppress noise. Its implementation process involves artificial model design and parameter adjustment, lacking adaptability to specific problems. In summary, for the research of magnetic anomaly signal detection algorithms, due to the limitations of specific conditions, traditional methods usually have poor generality and are difficult to effectively cope with the complex and changing ocean environment applications. Therefore, it is urgent to develop general ocean target magnetic anomaly signal detection models and algorithms to realize the suppression of ocean stray magnetic field noise and the detection of weak magnetic anomaly signals, and provide effective solutions for the accurate perception of magnetic targets in complex ocean environments.

[0006] In recent years, due to its flexible model structure and powerful learning ability, deep neural networks have received much attention and application in many fields, and have achieved more excellent performance than traditional methods. With the help of deep learning technology, it is expected to provide a general solution for the detection of magnetic anomaly signals of ocean targets, form an integrated network model structure with joint denoising and detection functions, and further improve the comprehensive perception ability of magnetic anomaly signals under low signal-to-noise ratio conditions. However, the measured ocean magnetic field data has problems of sample scarcity and class imbalance, which seriously restrict the learning and training of the magnetic anomaly perception model based on deep neural networks. Summary of the Invention

[0007] The embodiments of the present application provide a method for denoising and detecting magnetic anomaly signals of ocean targets based on deep transfer learning to solve the problems brought by sample scarcity and class imbalance in the measured ocean magnetic field data in the prior art.

[0008] On the one hand, the embodiments of the present application provide a method for denoising and detecting magnetic anomaly signals of ocean targets based on deep transfer learning, including:

[0009] Construct a simulation ocean target magnetic anomaly signal dataset by using the magnetic dipole physical field model;

[0010] Collect measured ocean magnetic field noise data and establish a measured ocean magnetic field noise dataset;

[0011] Add the data in the measured ocean magnetic field noise dataset to the simulation ocean target magnetic anomaly signal dataset to form a simulation noisy ocean target magnetic anomaly dataset;

[0012] Construct a denoising autoencoder network;

[0013] Use the denoising autoencoder network to denoise the simulation noisy ocean target magnetic anomaly dataset to obtain denoised magnetic field data, and use the denoised magnetic field data to adjust the parameters of the denoising autoencoder network;

[0014] Construct a fully connected classifier network;

[0015] Use the fully connected classifier network to classify the denoised magnetic field data to obtain a classification result, and use the classification result to adjust the parameters of the fully connected classifier network;

[0016] Collect the magnetic field data to be detected, input the magnetic field data to be detected into the adjusted denoising autoencoder network for denoising processing to obtain the denoised data to be detected, and input the denoised data to be detected into the adjusted fully connected classifier network for classification processing to obtain the final magnetic anomaly signal detection result.

[0017] The method for denoising and detecting magnetic anomaly signals of ocean targets based on deep transfer learning in the present application has the following advantages:

[0018] 1. Aiming at the problems that the acquisition cost of marine target magnetic anomaly signals is relatively high and there is still a lack of a complete marine magnetic field dataset for deep learning, this application combines the magnetic dipole physical field model. When the distance between the marine target and the magnetic sensor is greater than twice the maximum size of the target, the simulated magnetic field passing characteristic curve has a high similarity with the real magnetic field passing characteristic curve. On this basis, this application proposes to use the simulated magnetic field data to construct a training dataset for joint denoising and detection model learning, and transfer the extracted hierarchical features to the real marine magnetic field data through deep transfer learning technology to achieve the extraction and mining of target features in the marine magnetic field data, which well overcomes the deep learning bottleneck problem of small sample marine magnetic field data.

[0019] 2. Aiming at the problems of low efficiency, poor robustness, and strong dependence on manual design and parameter selection existing in traditional denoising methods, this application proposes a marine magnetic field noise suppression model and algorithm based on a convolutional denoising autoencoder, which can effectively model and learn noises with different distributions, and can fully explore the multi-scale time correlation characteristics of marine magnetic field data. Compared with traditional denoising methods, it has a higher signal-to-noise ratio improvement amplitude and a faster data processing speed. The convolutional denoising autoencoder in this application has strong versatility. After being trained with a large number of measured marine magnetic field noise data, the model can be used for automatic denoising tasks in complex interference environments, effectively breaking through the performance limitation problems existing in traditional denoising methods.

[0020] 3. Aiming at the problems that traditional magnetic anomaly denoising and detection methods rely on manual waveform interpretation and require prior parameter information of the detection scenario, this application adopts a marine target magnetic anomaly signal detection model and algorithm based on a neural network, which can realize the intelligent perception of target magnetic anomaly signals in marine magnetic field data. The method proposed in this application can automatically judge whether the data contains target magnetic anomaly signals without manual intervention or obtaining expensive scenario parameter information, effectively overcoming the drawback that traditional methods need to be judged manually according to waveform characteristics, and improving the detection accuracy and efficiency of marine target magnetic anomaly signals under low signal-to-noise ratio conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following described drawings are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0022] Figure 1 It is a flowchart of a marine target magnetic anomaly signal denoising and detection method based on deep transfer learning provided by an embodiment of the present application;

[0023] Figure 2 It is a schematic diagram of the principle of using simulation data for deep transfer learning provided by an embodiment of the present application;

[0024] Figure 3 It is a schematic diagram of the scenario for constructing simulation data using the physical field model of magnetic dipoles provided by an embodiment of the present application;

[0025] Figure 4 It is a schematic diagram of the structures of the denoising autoencoder network and the fully connected classifier network provided by an embodiment of the present application;

[0026] Figure 5 It is a graph showing the denoising results of the magnetic anomaly signal of the marine target provided by an embodiment of the present application; where (a) is the original noise-free magnetic anomaly signal of the marine target, (b) is the magnetic anomaly signal of the marine target after adding noise, and (c) is the result graph after denoising the noisy data using the method of the present application;

[0027] Figure 6 It is a comparison graph of the denoising results of the magnetic anomaly signal of the marine target after four-level decomposition using the wavelet transform denoising method provided by an embodiment of the present application. Specific embodiments

[0028] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the protection scope of the present application.

[0029] Figure 1 It is a flow chart of the method for denoising and detecting the magnetic anomaly signal of the marine target based on deep transfer learning provided by an embodiment of the present application. The embodiment of the present application provides a method for denoising and detecting the magnetic anomaly signal of the marine target based on deep transfer learning, including:

[0030] S100, constructing a simulation dataset of the magnetic anomaly signal of the marine target using the physical field model of magnetic dipoles.

[0031] Exemplarily, S100 specifically includes: according to the physical field model of magnetic dipoles, equivalent the marine target to a magnetic dipole, as Figure 3 shown, obtain the calculation expression of the induced magnetic field vector generated by the marine target at the magnetic sensor measurement point; change the movement speed of the marine target and the beam reach distance and vertical distance from the magnetic sensor, and determine the time series of the induced magnetic field vector by the magnetic sensor according to the calculation expression of the induced magnetic field vector; perform normalization processing on the time series of the induced magnetic field vector to obtain the simulation dataset of the magnetic anomaly signal of the marine target.

[0032] Before determining the calculation expression of the induced magnetic field vector, it is necessary to first establish a three-dimensional space coordinate system O-xyz for the magnetic sensor to measure the magnetic field of ocean targets. In this three-dimensional space coordinate system, the induced magnetic field vector has the following calculation expression:

[0033]

[0034] In the formula, μ0 is the magnetic permeability of vacuum, is the displacement vector between the ocean target and the magnetic sensor in the three-dimensional space coordinate system, is the magnetic moment vector of the ocean target to be measured. By changing the motion speed v of the ocean target, its beam reach distance a and vertical distance h from the magnetic sensor, the spatial intersection situation between the ocean target and the magnetic sensor is set. With the sampling frequency f of the magnetic sensor s a time series of the displacement vector is generated, and then a time series of is obtained according to the calculation expression of the induced magnetic field vector, and it is further normalized to construct a simulated ocean target magnetic anomaly signal data set

[0035] S110, collect the measured ocean magnetic field noise data and establish a measured ocean magnetic field noise data set.

[0036] Exemplarily, S110 specifically includes: collecting the measured ocean magnetic field noise data; segmenting and augmenting the measured ocean magnetic field noise data; normalizing the segmented and augmented measured ocean magnetic field noise data to obtain a measured ocean magnetic field noise data set.

[0037] Normalizing the segmented and augmented measured ocean magnetic field noise data can construct a measured ocean magnetic field noise data set

[0038] S120, add the data in the measured ocean magnetic field noise data set to the simulated ocean target magnetic anomaly signal data set to form a simulated noisy ocean target magnetic anomaly data set.

[0039] Exemplarily, S120 specifically includes: adding the data in the measured ocean magnetic field noise data set to the simulated ocean target magnetic anomaly signal data set with different signal-to-noise ratios for amplitude modulation, and truncating the added data set to obtain a simulated noisy ocean target magnetic anomaly data set, as Figure 2 shown.

[0040] Adding the measured ocean magnetic field noise data set to the simulated ocean target magnetic anomaly signal data set S, and after truncation, a simulated noisy ocean target magnetic anomaly data set is obtained

[0041] y (j-1)*q+k = f(s j + αn k ), j = 1, 2, …, p; k = 1, 2, …, q

[0042] where α is the noise amplitude modulation factor used to control the simulation of the noisy marine target magnetic anomaly signal with a specific signal-to-noise ratio, and f(·) is a truncation function of the following form:

[0043]

[0044] S130, construct a denoising autoencoder network.

[0045] Exemplarily, the denoising autoencoder network includes a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, a first deconvolutional layer, a second deconvolutional layer, and a third convolutional layer arranged in sequence, as Figure 4 shown.

[0046] S140, use the denoising autoencoder network to denoise the simulation noisy marine target magnetic anomaly data set to obtain denoised magnetic field data, and use the denoised magnetic field data to adjust the parameters of the denoising autoencoder network.

[0047] Exemplarily, when using the denoising autoencoder network to denoise the simulation noisy marine target magnetic anomaly data set, input the simulation noisy marine target magnetic anomaly data set into the first convolutional layer, and each convolutional kernel in the first convolutional layer performs dot product and addition operations on the input data to obtain the first convolutional output data; input the first convolutional output data into the first max-pooling layer to obtain the first pooling output data; input the first pooling output data into the second convolutional layer, and each convolutional kernel in the second convolutional layer performs dot product and addition operations on the input data to obtain the second convolutional output data; input the second convolutional output data into the second max-pooling layer to obtain the second pooling output data; input the second pooling output data into the first deconvolutional layer, and each deconvolutional kernel in the first deconvolutional layer performs deconvolution operations on the input data respectively to obtain the first deconvolutional output data; input the first deconvolutional output data into the second deconvolutional layer, and each deconvolutional kernel in the second deconvolutional layer performs deconvolution operations on the input data respectively to obtain the second deconvolutional output data; input the second deconvolutional output data into the third convolutional layer, and each convolutional kernel in the third convolutional layer performs dot product and addition operations on the input data to obtain the denoised magnetic field data.

[0048] Specifically, S140 includes the following sub-steps:

[0049] 1) Input the simulation noisy marine target magnetic anomaly signal into the first convolutional layer (the convolutional kernel is C1, and the bias is b1i ), each convolutional kernel performs dot multiplication and addition operations on the input data, and the generated first convolutional output data is:

[0050]

[0051] where y k is the k-th element of the simulated noisy ocean target magnetic anomaly signal, is the (k - i + 1)-th element of the i-th convolutional kernel, and the stride of the convolutional operation is selected as 1. For the final convolutional data, when it has less than m + 1 elements, 0 elements are padded at the back, and then the ReLU (Rectified Linear Unit) activation function is passed to generate x i = max(0, x i ). 32 convolutional kernels are used to perform the convolutional operation to form the first convolutional output data

[0052] 2) The data after the first convolutional layer operation is sent to the first max pooling layer. The size of the pooling window is 4, and the operation can be expressed as x i = max(x 4*i-3 , x 4*i-2 , x 4*i-1 , x 4*i ), and when the data entering the window has less than 4 elements, 0 elements are padded at the back to form the first pooling output data

[0053] 3) The first pooling output data passes through the second convolutional layer, where each channel of the input data corresponds to 1 convolutional kernel (the convolutional kernel is C2 i , and the bias is b 2_i ). Each convolutional kernel performs dot multiplication and addition operations on the input data, and the generated new data is:

[0054]

[0055] where x k is the k-th element of the first pooling output data, is the (k - i + 1)-th element of the i-th convolutional kernel, and the stride of the convolutional operation is selected as 1. For the final convolutional data, when it has less than m + 1 elements, 0 elements are padded at the back, and then the ReLU activation function is passed to generate the new data x i = max(0, x i ), to form the second convolutional output data

[0056] 4) The second convolutional output data After the second maximum pooling layer, the size of the pooling window is 4, and the operation can be expressed as x i = max(x 4*i-3 , x 4*i-2 , x 4*i-1 , x 4*i ). When the data entering the window is less than 4 elements, 0 elements are filled backward to form the second pooling output data

[0057] 5) The second pooling output data Pass through the first deconvolution layer (the i-th deconvolution kernel is The bias is ). There are 32 deconvolution kernels that perform deconvolution operations on 32 channels of respectively, with a moving step of 4, is the superposition value where the moving step of the deconvolution kernel is less than the new data volume generated by the deconvolution kernel. The data generated by the deconvolution operation can be expressed as Then the output data generated by passing through the ReLU activation function is x i = max(0, x i ). The first deconvolution output data after passing through the first deconvolution layer can be expressed as

[0058] 6) The first deconvolution output data Pass through the second deconvolution layer (the i-th deconvolution kernel is The bias is ). There are 32 deconvolution kernels that perform deconvolution operations on 32 channels of respectively, with a moving step of 4, is the superposition value where the moving step of the deconvolution kernel is less than the new data volume generated by the deconvolution kernel. The data generated by the deconvolution operation can be expressed as Then the output data generated by passing through the ReLU activation function is x i = max(0, x i ). The second deconvolution output data after passing through the second deconvolution layer can be expressed as

[0059] 7) The second deconvolution output data Pass through the third convolution layer (the convolution kernel is C3, the bias is b 3_i ). One convolution kernel performs dot multiplication and addition operations on each channel of X, and the new data generated is:

[0060]

[0061] Among them, x k is the k-th element of the second deconvolution output data, C3 k-i+1is the (k - i + 1)-th element of the third convolutional layer, and the stride of the convolutional operation is selected as 1. For the final convolutional data, when it has less than m + 1 elements, 0 elements are padded at the back, and then the corresponding points of the generated 32-channel data are added together to obtain x i = x 1_i + x 2_i + … + x 32_i , and finally the data generated through the sigmoid activation function is The denoised magnetic field data after passing through the third convolutional layer can be expressed as

[0062] Furthermore, when adjusting the parameters of the denoising autoencoder network using the denoised magnetic field data, the mean square error between the denoised magnetic field data and the data in the simulated noisy marine target magnetic anomaly dataset is used as the loss function, and based on this loss function, the parameters of the denoising autoencoder network at the next moment are determined.

[0063] The denoised magnetic field data x output by the denoising autoencoder network has the same dimension as the input simulated noise-free marine target magnetic anomaly signal y. The Adam (Adaptive Moment Estimation) optimizer is used to learn and update the convolutional kernels and biases (denoted by θ) of the network, and the loss function adopted is the mean square error between x and y:[[]]END]]

[0064]

[0065] Find the loss function The partial derivative g t of the network parameter θ t at time t is obtained as:[[]]END]]

[0066]

[0067] Next, the first-order and second-order moment estimates m t and v t of the gradient in momentum form at time t are calculated:[[]]END]]

[0068]

[0069] On this basis, the first-order and second-order moment estimates and

[0070]

[0071] after bias correction at time t are obtained, and then t+1 the update formula for the network parameter θ at time t + 1 is as follows:[[]]END]]

[0072]

[0073] Among them, η and ε are the iteration step size and a very small number. Through the above iterative training process, the convolution kernels and bias amounts of the denoising autoencoder network are continuously adjusted and updated, so that the structure of the magnetic field data after adding noise approaches the data without adding noise after passing through the network operation, thereby completing the training of the denoising autoencoder network.

[0074] S150, construct a fully connected classifier network.

[0075] Exemplarily, the fully connected classifier network includes a fully connected layer, a ReLU activation function, and a Softmax layer arranged in sequence, as Figure 4 shown, where the input weights of the fully connected layer are and the bias amount is

[0076] S160, use the fully connected classifier network to classify the denoised magnetic field data, obtain the classification result, and use the classification result to adjust the parameters of the fully connected classifier network.

[0077] Exemplarily, when using the fully connected classifier network to classify the denoised magnetic field data, input the denoised magnetic field data into the fully connected layer to obtain an intermediate output, pass the intermediate output through the ReLU activation function to obtain the output data of the fully connected layer; input the output data of the fully connected layer into the Softmax layer to obtain the output probability value of each node in the Softmax layer, and take the category corresponding to the node with the largest output probability value as the classification result of the denoised magnetic field data.

[0078] The intermediate output can be expressed as:

[0079]

[0080] Pass the intermediate output through the ReLU activation function to generate the output y = max(0, h) of the fully connected layer, and there is Finally, pass the output of the fully connected layer through the Softmax layer with the number of nodes C to obtain the output vector and generate the output probability value of each node accordingly:

[0081]

[0082] where p i ∈[0, 1] is the output probability value of the i-th node, and there is Select the category corresponding to the node with the largest output probability value as the classification result of the denoised magnetic field data, and use this to judge whether there is a target magnetic anomaly signal in the data sample.

[0083] Further, when adjusting the parameters of the fully-connected classifier network using the classification results, the cross-entropy between the output probability value of each node in the Softmax layer and the corresponding true class label is used as the loss function, and the parameters of the fully-connected classifier network at the next moment are determined based on this loss function.

[0084] In the training of the fully-connected classifier network, the output probability value of the node is compared with the true class label and the cross-entropy loss function L for learning the parameters of the fully-connected classifier network is constructed:

[0085]

[0086] Then, following the same training steps as in S140 for the denoising autoencoder network, the parameters of the fully-connected classifier network are iteratively updated to complete the training of the fully-connected classification detector network.

[0087] S170, collect the magnetic field data to be detected, input the magnetic field data to be detected into the adjusted denoising autoencoder network for denoising processing to obtain the denoised data to be detected, and input the denoised data to be detected into the adjusted fully-connected classifier network for classification processing to obtain the final magnetic anomaly signal detection result.

[0088] The effects of this application are further illustrated by the following simulation experiments:

[0089] (1) Experimental simulation conditions:

[0090] The real data used in this experiment are the magnetic field data of surface ships and the ocean environment collected in Sanya. The sampling frequency of the fluxgate sensor is 200 Hz. Each selected data sample contains 80,000 magnetic field sampling points. The simulated magnetic field data are generated according to the same sampling frequency and data length as the real data. The detailed composition of the ocean magnetic field dataset constructed in the experiment is shown in Table 1. This experiment is run on an Intel(R) Core(TM) i7-8750H CPU with a main frequency of 2.20 GHz and a memory of 16 GB. Magnetic field data simulation is carried out using MATLAB R2021a software on the Windows 10 operating system. The joint denoising and detection network models are built and learned and trained using Pycharm2021 software, Python3.8 programming language, and TensorFlow2.2 deep learning framework.

[0091] Table 1 Detailed composition of the constructed ocean magnetic field dataset

[0092]

[0093] (2) Data denoising performance evaluation criteria:

[0094] (2a) SNR improvement before and after data denoising: SNR improvement (dB)

[0095] When calculating the SNR improvement before and after data denoising, the data needs to be normalized first. Assume that is the original noise-free data after normalization, is the noisy data after normalization, is the denoised data after normalization, then the SNR improvement SNR improve is defined as follows:

[0096]

[0097] SNR improvement SNR improve The larger the value, the better the performance of the denoising algorithm.

[0098] (2b) Operating efficiency of the denoising algorithm: Time (s)

[0099] The time for data denoising processing starts from when the noisy data is put into memory and calculates the difference of the system internal timer until the denoised data is generated. Under the same computing configuration conditions, the shorter the denoising processing time, the higher the operating efficiency of the denoising algorithm.

[0100] (3) Magnetic anomaly detection performance evaluation criteria:

[0101] (3a) Detection rate: P d (%)

[0102] The detection rate P d is defined as follows:

[0103] P d = n tt / n t

[0104] where, n tt is the number of positive samples judged as positive samples, n t is the total number of positive samples.

[0105] (3b) False alarm rate: P f (%)

[0106] The false alarm rate P f is defined as follows:

[0107] P f = n ct / n c

[0108] where, n ct is the number of negative samples judged as positive samples, n cis the total number of negative samples.

[0109] (3c) Accuracy: Acc(%)

[0110] The definition of accuracy Acc is as follows:

[0111] Acc = n r / n

[0112] where n r is the number of correctly classified samples, and n is the total number of samples.

[0113] (3d) Operating efficiency of the detection algorithm: Time (s)

[0114] The time for magnetic anomaly detection starts from when the denoised data is input into the memory and calculates the difference of the system internal timer until the data classification result is generated. Under the same calculation configuration conditions, the shorter the time for magnetic anomaly detection, the higher the operating efficiency of the detection algorithm.

[0115] (4) Experimental content:

[0116] Experiment 1

[0117] Using the constructed marine magnetic field dataset, after adding noise to the data, the denoising autoencoder network of this application and the traditional wavelet transform method are respectively used for denoising, and the experimental results are respectively as Figure 5 and Figure 6 shown. In addition, the denoising performances of the method of this application (denoted by CDAE) and the wavelet transform method (denoted by WT) on the test dataset are compared, and the signal-to-noise ratio improvement (unit: dB) and average running time (unit: ms) indicators are used to evaluate the denoising effect and efficiency. The specific statistical results are shown in Table 2.

[0118] Table 2 Comparison of denoising performances of different denoising methods on the test dataset

[0119] Denoiser SNR improvement (dB) Average running time (ms) WT 15 29 CDAE 19 22

[0120] As can be seen from the results in Table 2, the denoising method proposed in this application has a higher signal-to-noise ratio improvement amplitude and a faster denoising processing speed compared with the traditional wavelet transform method, which helps to achieve the rapid detection of magnetic anomaly signals under low signal-to-noise ratio conditions.

[0121] Experiment 2

[0122] Use the fully connected classifier network (denoted by FCC) of this application to detect the magnetic anomaly signals in the test dataset, and compare with the performances of the traditional OBF and MED magnetic anomaly denoising and detection methods, and respectively obtain the detection rate P d and the false alarm rate P f, the detection effect and efficiency are evaluated by the accuracy Acc and the average running time (unit: ms). The specific statistical results are shown in Table 3.

[0123] Table 3 Comparison of Detection Performance of Different Magnetic Anomaly Denoising and Detection Methods on the Test Dataset

[0124]

[0125] From the results in Table 3, it can be seen that the denoising and detection methods proposed in this application are significantly superior to the traditional OBF and MED methods in terms of detection rate, false alarm rate, accuracy, and running time, with high detection accuracy and rate, extremely low false alarm rate, and fast detection running time.

[0126] It can be seen that the magnetic anomaly signal denoising and detection method proposed in this application adopts a data-driven intelligent and integrated deep neural network architecture and learning perception strategy, which can greatly reduce the dependence on scene prior knowledge and artificial rule design, effectively improve the detection ability of the magnetic measurement system for low signal-to-noise ratio target magnetic anomaly signals, and has broad application prospects in the field of long-distance magnetic detection of marine targets in complex interference environments.

[0127] Although the preferred embodiments of this application have been described, those skilled in the art can make additional changes and modifications once they know the basic creative concept. Therefore, the appended claims are intended to be interpreted to include the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0128] Obviously, those skilled in the art can make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalent technologies, this application is also intended to include these changes and modifications.

Claims

1. A method for denoising and detecting magnetic anomaly signals of ocean targets based on deep transfer learning, characterized in that, Including: Constructing a simulated marine target magnetic anomaly signal data set by using a magnetic dipole physical field model; Collecting measured marine magnetic field noise data and establishing a measured marine magnetic field noise data set; Adding the data in the measured marine magnetic field noise data set to the simulated marine target magnetic anomaly signal data set to form a simulated noisy marine target magnetic anomaly data set; Constructing a denoising autoencoder network; Using the denoising autoencoder network to perform denoising processing on the simulated noisy marine target magnetic anomaly data set to obtain denoised magnetic field data, and using the denoised magnetic field data to adjust the parameters of the denoising autoencoder network; Constructing a fully connected classifier network; Using the fully connected classifier network to classify the denoised magnetic field data to obtain a classification result, and using the classification result to adjust the parameters of the fully connected classifier network; Collecting magnetic field data to be detected, inputting the magnetic field data to be detected into the adjusted denoising autoencoder network for denoising processing to obtain denoised data to be detected, and inputting the denoised data to be detected into the adjusted fully connected classifier network for classification processing to obtain a final magnetic anomaly signal detection result.

2. The method for denoising and detecting marine target magnetic anomaly signals based on deep transfer learning according to claim 1, characterized in that The constructing a simulated marine target magnetic anomaly signal data set by using a magnetic dipole physical field model includes: According to the magnetic dipole physical field model, equivalent the marine target to a magnetic dipole to obtain a calculation expression of the induced magnetic field vector generated by the marine target at the magnetic sensor measurement point; Changing the motion speed of the marine target and the beam distance and vertical distance from the magnetic sensor, and determining the induced magnetic field vector time series by the magnetic sensor according to the calculation expression of the induced magnetic field vector; Performing normalization processing on the induced magnetic field vector time series to obtain the simulated marine target magnetic anomaly signal data set.

3. The method for denoising and detecting marine target magnetic anomaly signals based on deep transfer learning according to claim 1, characterized in that The collecting measured marine magnetic field noise data and establishing a measured marine magnetic field noise data set includes: Collecting measured marine magnetic field noise data; Performing segmented interception and augmentation on the measured marine magnetic field noise data; Performing normalization processing on the segmented intercepted and augmented measured marine magnetic field noise data to obtain the measured marine magnetic field noise data set.

4. The method for denoising and detecting marine target magnetic anomaly signals based on deep transfer learning according to claim 1, wherein The adding the data in the measured marine magnetic field noise data set to the simulated marine target magnetic anomaly signal data set to form a simulated noisy marine target magnetic anomaly data set includes: Adding the data in the measured marine magnetic field noise data set to the simulated marine target magnetic anomaly signal data set by amplitude modulation at different signal-to-noise ratios, and performing truncation processing on the added data set to obtain the simulated noisy marine target magnetic anomaly data set.

5. The method for denoising and detecting marine target magnetic anomaly signals based on deep transfer learning according to claim 1, characterized in that, The denoising autoencoder network includes a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a first deconvolutional layer, a second deconvolutional layer, and a third convolutional layer arranged in sequence. When using the denoising autoencoder network to denoise the simulated noisy marine target magnetic anomaly dataset, the simulated noisy marine target magnetic anomaly dataset is input into the first convolutional layer, and each convolutional kernel in the first convolutional layer performs dot multiplication and addition operations on the input data to obtain the first convolutional output data; the first convolutional output data is input into the first max pooling layer to obtain the first pooling output data; The first pooling output data is input into the second convolutional layer, and each convolutional kernel in the second convolutional layer performs dot multiplication and addition operations on the input data to obtain the second convolutional output data; the second convolutional output data is input into the second max pooling layer to obtain the second pooling output data; The second pooling output data is input into the first deconvolutional layer, and each deconvolutional kernel in the first deconvolutional layer performs deconvolution operations on the input data respectively to obtain the first deconvolutional output data; The first deconvolutional output data is input into the second deconvolutional layer, and each deconvolutional kernel in the second deconvolutional layer performs deconvolution operations on the input data respectively to obtain the second deconvolutional output data; The second deconvolutional output data is input into the third convolutional layer, and each convolutional kernel in the third convolutional layer performs dot multiplication and addition operations on the input data to obtain the denoised magnetic field data.

6. The method for denoising and detecting marine target magnetic anomaly signals based on deep transfer learning according to claim 1, wherein When using the denoised magnetic field data to adjust the parameters of the denoising autoencoder network, the mean square error between the denoised magnetic field data and the data in the simulated noisy marine target magnetic anomaly dataset is used as the loss function, and based on this loss function, the parameters of the denoising autoencoder network at the next moment are determined.

7. The method for denoising and detecting marine target magnetic anomaly signals based on deep transfer learning according to claim 1, characterized in that The fully connected classifier network includes a fully connected layer, a ReLU activation function, and a Softmax layer arranged in sequence. When using the fully connected classifier network to classify the denoised magnetic field data, the denoised magnetic field data is input into the fully connected layer to obtain an intermediate output, and the intermediate output passes through the ReLU activation function to obtain the output data of the fully connected layer; The output data of the fully connected layer is input into the Softmax layer to obtain the output probability value of each node in the Softmax layer, and the category corresponding to the node with the largest output probability value is used as the classification result of the denoised magnetic field data.

8. The method for denoising and detecting marine target magnetic anomaly signals based on deep transfer learning according to claim 7, characterized in that, When using the classification result to adjust the parameters of the fully connected classifier network, the cross entropy between the output probability value of each node in the Softmax layer and the corresponding true class label is used as the loss function, and based on this loss function, the parameters of the fully connected classifier network at the next moment are determined.

Images