Background Noise Magnetic Anomaly Detection Method Based on Sparse Denoising Autoencoder

Abstract

This application discloses a method for detecting magnetic anomalies in background noise based on a sparse denoising autoencoder, belonging to the field of intelligent signal learning and perception technology. The method includes: acquiring noisy magnetic field data; inputting the data into a sparse denoising autoencoder for sparse denoising processing to obtain denoised magnetic field data; establishing a loss function; training the sparse denoising autoencoder using the loss function; acquiring magnetic field data to be detected; inputting the magnetic field data to be detected into the trained sparse denoising autoencoder to obtain the corresponding real-time reconstruction error; comparing the real-time reconstruction error with a set reconstruction error threshold to determine whether the magnetic field data to be detected contains magnetic anomaly signals. Compared with existing machine learning methods, the method in this application does not require manual annotation of massive amounts of data; it only requires training using unlabeled ocean magnetic field noise data collected in the ocean.

Description

Technical Field

[0001] This application relates to the field of intelligent signal learning and perception technology, and in particular to a method for detecting background noise magnetic anomalies based on a sparse denoising autoencoder. Background Technology

[0002] The Earth is a vast magnetic field. Ferromagnetic objects placed within this field become magnetized, causing distortion and deformation of the surrounding magnetic field—a phenomenon known as magnetic anomaly. Because detecting magnetic anomaly signals does not require actively emitting electromagnetic waves, it is possible to detect, locate, and navigate marine targets in a highly covert manner by detecting magnetic anomaly signals in the ocean. However, the actual ocean magnetic field environment is complex, containing various types of noise interference, and lacks an accurate model of the ocean's background magnetic field to serve as a benchmark for data analysis. Accurately detecting weak target magnetic anomaly signals in the complex and variable ocean noise environment remains a significant challenge.

[0003] To date, methods for detecting magnetic anomalies can be broadly categorized into two types. One approach transforms signal features, mapping complex time-domain magnetic signals mathematically to specific feature spaces, emphasizing and amplifying anomalous components in the magnetic data, and then using a set feature threshold to determine the presence or absence of a target. The other approach extracts multiple features from the magnetic field signal and employs machine learning methods for automatic target classification. For the first type of method, commonly used models include orthogonal basis decomposition, minimum entropy, wavelet transform, and Hilbert-Huang transform. Orthogonal basis decomposition projects or unfolds a complex signal onto an orthogonal basis, obtaining its coordinate representation on that basis. This eliminates redundant information, effectively extracts features, and highlights anomalous components. Minimum entropy directly reflects signal uncertainty by utilizing the statistical properties of entropy, detecting magnetic anomalies by analyzing entropy changes. Wavelet transform decomposes the signal into waveform coefficients with different time and frequency resolutions through multi-scale decomposition, detecting magnetic anomaly components by utilizing abrupt changes in the waveform coefficients at potential scales and times. The Hilbert-Huang transform method maps the signal to a time-frequency plane to generate a time-frequency graph. Based on the abrupt changes in the time-frequency characteristics of the signal, it can intuitively and effectively identify and detect the magnetic anomaly components in the signal.

[0004] Generally, orthogonal basis decomposition methods are only suitable for magnetic field environments with Gaussian white noise as the background, and are difficult to effectively handle marine environmental noise with complex distributions. Furthermore, the difficulty in estimating the shortest distance and relative motion state between the target and the moving platform limits the practical application of this method. Minimum entropy methods lack adaptability to background noise; significant variations in marine noise over region and time can render this method ineffective. Wavelet transform methods rely on empirical selection of wavelet bases, resulting in unstable detection performance and a high false alarm rate. Hilbert-Huang transform suffers from reduced analytical performance due to background noise contaminating the entire time-frequency graph, and also has high computational cost and poor real-time performance.

[0005] The second category of methods includes representative models such as Support Vector Machines, Isolation Forests, and one-dimensional convolutional deep neural networks. These methods combine magnetic anomaly detection with machine learning and deep learning techniques to achieve intelligent and automated detection of magnetic anomaly signals. These methods extract and transform the features of the target, and then use feature pattern classification to discriminate the magnetic anomaly signal. Deep learning and machine learning models require large amounts of data for training, and simulation data cannot perfectly reflect real-world conditions. However, ocean magnetic field data collected in real-world environments suffers from sample scarcity and imbalance. Acquiring real target magnetic field signals in the ocean is costly, difficult, and limited in quantity, requiring manual labeling of the dataset. However, a sufficient amount of inexpensive and unlabeled noise samples remain unutilized, hindering the practical development of this technology. To meet the application needs of real and complex marine environments, there is an urgent need to develop a magnetic anomaly detection algorithm capable of self-supervised learning from collected unlabeled magnetic field data. Summary of the Invention

[0006] This application provides a method for detecting background noise magnetic anomalies based on a sparse denoising autoencoder, which solves the problem in the prior art that a large amount of labeled target magnetic field data is required when detecting marine targets.

[0007] On one hand, embodiments of this application provide a method for detecting background noise magnetic anomalies based on a sparse denoising autoencoder, including:

[0008] Collect noise magnetic field data, which are magnetic field data in the ocean that only contain background noise magnetic field;

[0009] The noisy magnetic field data is input into a sparse denoising autoencoder for sparse denoising processing to obtain denoised magnetic field data.

[0010] A loss function is established by utilizing the activation level of neurons in the hidden layer and the difference between the denoised magnetic field data and the desired magnetic field data.

[0011] The sparse denoising autoencoder is trained using a loss function to obtain a trained sparse denoising autoencoder.

[0012] Collect data on the magnetic field to be detected;

[0013] Input the magnetic field data to be detected into the trained sparse denoising autoencoder to obtain the corresponding real-time reconstruction error.

[0014] The real-time reconstruction error is compared with the set reconstruction error threshold, and the result of the comparison determines whether the magnetic field data to be detected contains magnetic anomaly signals.

[0015] On the other hand, embodiments of this application also provide a marine target detection system, including:

[0016] The first data acquisition module is used to acquire noise magnetic field data, which is magnetic field data in the ocean that only contains background noise magnetic field.

[0017] The noise reduction module is used to input the noisy magnetic field data into the sparse noise reduction autoencoder for sparse noise reduction processing to obtain the noise-reduced magnetic field data.

[0018] The function creation module is used to establish a loss function by utilizing the activation level of neurons in the hidden layer and the difference between the denoised magnetic field data and the expected magnetic field data.

[0019] The autoencoder training module is used to train the sparse denoising autoencoder using a loss function to obtain the trained sparse denoising autoencoder.

[0020] The second data acquisition module is used to acquire the magnetic field data to be detected.

[0021] The error calculation module is used to input the magnetic field data to be detected into the trained sparse denoising autoencoder to obtain the corresponding real-time reconstruction error.

[0022] The anomaly detection module is used to compare the real-time reconstruction error with the set reconstruction error threshold, and determine whether the magnetic field data to be detected contains a magnetic anomaly signal based on the comparison result.

[0023] On the other hand, embodiments of this application also provide a computer storage medium storing a plurality of computer instructions for causing a computer to execute the above-described method.

[0024] The background noise magnetic anomaly detection method based on a sparse denoising autoencoder in this application has the following advantages:

[0025] 1. A self-supervised sparse denoising autoencoder is proposed, and an automatic threshold segmentation method based on reconstruction error is designed to achieve adaptive detection of marine magnetic anomaly signals. Compared with existing machine learning methods, the proposed method does not require manual annotation of massive amounts of data, but only needs to be trained using unlabeled marine magnetic field noise data collected in the ocean.

[0026] 2. By capturing hierarchical signal features through multi-scale, multi-channel convolutional kernels, the distribution characteristics and dependencies of marine magnetic field noise data can be accurately learned, providing a new approach for the characterization and modeling of magnetic field noise data in complex marine environments. Simultaneously, sparse learning of background noise can effectively mine a compact latent space, enabling the extraction of discriminative features and efficient differentiation of magnetic anomaly signals.

[0027] 3. By training with self-supervised learning on unlabeled ocean noise, a magnetic anomaly denoising and detection model with stronger generalization ability can be obtained. Through end-to-end learning optimization, the proposed sparse denoising autoencoder can capture the signal variation patterns in real ocean environments, achieving effective suppression and elimination of unknown random noise. The end-to-end training method overcomes the limitation of traditional methods that require manual feature extraction, bringing a new solution to the problem of magnetic field information perception in the field of ocean exploration. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 A flowchart of a background noise magnetic anomaly detection method based on a sparse denoising autoencoder provided in this application embodiment;

[0030] Figure 2 A scene diagram illustrating the acquisition of noise magnetic field data and target magnetic field data in an experiment using the method described in this application;

[0031] Figure 3 A schematic diagram illustrating the composition of a sparse noise reduction autoencoder provided in an embodiment of this application;

[0032] Figure 4 This is a denoising result diagram of the noisy magnetic field data provided in the embodiments of this application;

[0033] Figure 5 The output result diagram of the hidden layer in the ablation experiment provided in the embodiment of this application is shown. Detailed Implementation

[0034] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] Figure 1 A flowchart illustrating the background noise magnetic anomaly detection method based on a sparse denoising autoencoder provided in this application embodiment. This application embodiment provides a background noise magnetic anomaly detection method based on a sparse denoising autoencoder, including:

[0036] S100 collects noise magnetic field data, which is magnetic field data in the ocean that only contains background noise magnetic field.

[0037] For example, one can refer to Figure 2 The scene shown is used to collect noise magnetic field data.

[0038] S110 inputs the noisy magnetic field data into the sparse denoising autoencoder for sparse denoising processing to obtain denoised magnetic field data.

[0039] For example, S110 specifically includes: inputting noisy magnetic field data into a first convolutional layer, extracting features from the noisy magnetic field data, and obtaining first encoded data; inputting the first encoded data into a second convolutional layer, extracting features from the first encoded data, and obtaining second encoded data; inputting the second encoded data into a third convolutional layer, extracting features from the second encoded data, and obtaining third encoded data; inputting the third encoded data into a first deconvolutional layer, and obtaining first decoded data; inputting the first decoded data into a second deconvolutional layer, and obtaining second decoded data; inputting the second decoded data into a third deconvolutional layer, and obtaining denoised magnetic field data.

[0040] After extracting features from the noisy magnetic field data, the first encoded data, and the second encoded data, the first convolutional layer, the second convolutional layer, and the third convolutional layer sequentially perform padding and activation processing on the extracted feature data. After deconvolving the third encoded data, the first decoded data, and the second decoded data, the first deconvolutional layer, the second deconvolutional layer, and the third deconvolutional layer sequentially perform shifting, stacking, and activation processing on the deconvolved data.

[0041] Specifically, such as Figure 3 As shown, the sparse denoising autoencoder in this application includes an encoder and a decoder. The encoder contains convolutional layers and hidden layers, while the decoder contains deconvolutional layers and hidden layers. The hidden layer with sparse constraints is connected between the third convolutional layer and the first deconvolutional layer. (Noisy magnetic field data) Entering the first convolutional layer, the size of the convolutional kernel C1 is 16×5×1 (16 convolutional kernels, each kernel size is 5, and each kernel has 1 channel). Using multi-channel convolutional kernels allows for the capture of signal features in multiple dimensions. The bias is b1. Each convolutional kernel performs dot product and summation operations on the input data to generate new data:

[0042]

[0043] in, This is the k-th element in the noisy magnetic field data. Let m be the (k-i+1)th element of the i-th convolutional kernel, and let 2 be the stride of the convolution operation. After the convolution calculation, data with fewer than m+1 elements is padded with zeros, and then the data is activated by the ReLU (Rectified Linear Unit) function to generate new data. After performing convolution operations on each convolution kernel according to this method, the first encoded data is formed.

[0044] First encoded data After the second convolutional layer, since the input data is multi-channel, the number of channels in the corresponding convolutional kernel is the same as the number of channels in the input data. The convolutional kernel of the second convolutional layer is C2, with a size of 8×5×16 (8 convolutional kernels, each kernel size of 5, and each kernel having 16 channels). This reduces the number of channels in the output data and fuses the extracted features. The bias is b2. Each convolutional kernel performs dot product and summation operations on the input data to generate new data.

[0045]

[0046] in, For the k-th element in the first encoded data, Let x be the (k-i+1)th element of the i-th convolution kernel, with a stride of 2. After the convolution calculation, data with fewer than m+1 elements is padded with zeros. The corresponding points of the 16-channel data are summed to x. 2i =x 1_i +x 2_i ...+x 16_i Then, after passing through the ReLU activation function, new data is generated. After performing convolution operations on each convolution kernel, the second encoded data is formed.

[0047] The second encoded data After the third convolutional layer, since the input data is multi-channel, the number of channels in the corresponding convolutional kernel is the same as the number of channels in the input data. The convolutional kernel of the third convolutional layer is C3, with a size of 4×3×8 (4 convolutional kernels, each kernel size of 3, and each kernel having 8 channels), and a bias of b3. Each convolutional kernel performs dot product and summation operations on the input data to generate new data:

[0048]

[0049] in, For the k-th element in the second encoded data, Let x be the (k-i+1)th element of the i-th convolution kernel, with a stride of 2. After the convolution calculation, data with fewer than m+1 elements is padded with zeros. The corresponding points of the 16-channel data are summed to x. 3i =x 1_i +x 2_i ...+x 16_i Then, after passing through the ReLU activation function, new data is generated. A convolution operation is performed on each convolution kernel to form the third encoded data. Third-code data These are the data features ultimately extracted by the encoder.

[0050] The extracted third-encoded data After the first deconvolution layer, the decoder maps the data features extracted by the encoder back to the original data space. trans1 C trans1 The size is 8×3×4 (8 convolutional kernels, each kernel size is 3, and each kernel has 4 channels), and there are 8 deconvolutional kernels for each kernel. Perform deconvolution operation. This is the bias. Each convolutional kernel performs a dot product operation on the input data to produce a set of sums. Given data of the same size, shift the data by a step of 2 to generate a new set of data. Add the overlapping portions of the new data to the original data. The new data can be represented as:

[0051]

[0052] in, This is the k-th element in the third encoded data. The sum of the corresponding points from the four channels is x. 4i =x 1_i +x 2_i ...+x 4_i Then, after passing through the ReLU activation function, new data is generated. The first decoded data generated after the deconvolution operation is represented as follows:

[0053] First decoded data After the second deconvolution layer, the deconvolution kernel is C. trans2 C trans2 The size is 16×5×8 (16 convolutional kernels, each kernel size is 5, and each kernel has 8 channels), and there are 16 deconvolutional kernels for each kernel. Perform deconvolution operation. This is the bias. Each convolutional kernel performs a dot product operation on the input data to produce a set of sums. Given data of the same size, shift the data by a step of 2 to generate a new set of data. Add the overlapping portions of the new data to the original data. The new data can be represented as:

[0054]

[0055] in, This is the k-th element in the first decoded data. The sum of the corresponding points from all 8 channels is x. 5i =x 1_i +x 2_i ...+x 8_i Then, after passing through the ReLU activation function, new data is generated. The second decoded data generated after the deconvolution operation is represented as follows:

[0056] The second decoded data After the third deconvolution layer, the deconvolution kernel is C. trans3 C trans3 The size is 1×5×16 (1 convolution kernel, each kernel size is 5, each kernel has 16 channels), and the deconvolution kernel pairs... Perform deconvolution operation. This is the bias. Each convolutional kernel performs a dot product operation on the input data to produce a set of sums. Given data of the same size, shift the data by a step of 2 to generate a new set of data. Add the overlapping portions of the new data to the original data. The new data can be represented as:

[0057]

[0058] in, This is the k-th element of the second decoded data. The sum of the corresponding points from all 16 channels is x. 6i =x 1_i +x 2_i ...+x 16_i Finally, after passing through the Sigmoid activation function, new data is generated. The obtained denoised magnetic field data is

[0059] Furthermore, the noisy magnetic field data is normalized before being input into the first convolutional layer.

[0060] Specifically, for noisy magnetic field data Each element x after processing i The value is:

[0061]

[0062] Where, x i Let i be the i-th element of the input noisy magnetic field data. The normalized data is: Feeding data into a sparse denoising autoencoder can enhance the stability and comparability of the data, thereby improving the training efficiency and performance of the model.

[0063] S120, a loss function is established by utilizing the activation level of neurons in the hidden layer and the difference between the denoised magnetic field data and the desired magnetic field data.

[0064] For example, the method for establishing the loss function is as follows: determine the activation level of each neuron in the hidden layer; determine the distance between the activation level and the expected activation level, and use the distance as a penalty term for the hidden layer; determine the difference term based on the difference between the denoised magnetic field data and the expected magnetic field data; and perform a weighted summation of the penalty term and the difference term to obtain the loss function.

[0065] Specifically, noisy magnetic field data collected from the ocean contains many redundant and repetitive features. This is because many sources of ocean electromagnetic noise are inherently random, such as noise caused by wave motion, electromagnetic interference noise from cables and other equipment, etc. If noisy magnetic field data is modeled directly, the model easily captures and repeatedly fits these redundant random noise patterns, thus failing to learn meaningful and discriminative features, leading to severe overfitting. Sparse constraints, however, force each feature extraction unit to be sensitive only to a subset of the data. Therefore, the rich redundancy in noisy magnetic field data provides an incentive to incorporate sparsity regularization, which allows the model to focus on true patterns and prevents overfitting to repetitive random noise. Using a i (x (j) The function determines the activation level of each neuron in the hidden layer as follows:

[0066]

[0067] Among them, a i (x (j)The expression () represents the activation level of the i-th neuron in response to the j-th sample. The average activation level of the neurons in the hidden layer is obtained by averaging the activation levels of all neurons.

[0068]

[0069] Set the sparsity ratio ρ of the hidden layer (the expected activation level of the hidden layer), and use KL (Kullback-Leibler) divergence to measure the distance between the current average activation level of the hidden layer and the set activation level:

[0070]

[0071] The KL divergence value, which measures the distance between the current activation level and the expected activation level, is used as a penalty term to constrain the activation level of neurons in the hidden layer.

[0072]

[0073] The mean square error is used to measure the noise reduction of the magnetic field data X. out With expected magnetic field data Y n×1 Differences:

[0074]

[0075] in, Let y be the i-th element of the denoised magnetic field data. i Let be the i-th element of the desired magnetic field data, and n be the number of elements in a sample. The mean magnetic field of a targetless region over a certain time period can be represented as the value of the desired magnetic field data, so the desired output is:

[0076]

[0077] Use a hyperparameter λ sparse By controlling the weights of the sparsity penalty term, and combining the penalty term and the difference term, we obtain the loss function with overall sparsity constraints:

[0078] J(W,b)=J MSE (W,b)+λ sparse ·J sparse (W,b)

[0079] S130, the sparse denoising autoencoder is trained using the loss function to obtain the trained sparse denoising autoencoder.

[0080] For example, when training a sparse denoising autoencoder, the Adam optimizer is used to update the convolution kernel and bias parameters of the sparse denoising autoencoder.

[0081] Find the loss function J(W,b) with respect to the network parameters θ at time t. tpartial derivative g t have to:

[0082]

[0083] Next, the first and second moment estimates m of the gradient in momentum form at time t are obtained. t and v t :

[0084]

[0085] Based on this, the first and second moment estimates after bias correction at time t are obtained. and

[0086]

[0087] Then the network parameters θ at time t+1 are obtained. t+1 The update formula is as follows:

[0088]

[0089] Here, η and ε are the iteration step size and a very small number, respectively. Through the above iterative training process, the convolution kernel and bias of the network are continuously adjusted and updated, so that the desired magnetic field data approximates the geomagnetic field data after removing environmental noise after a series of operations.

[0090] S140, collect the magnetic field data to be detected.

[0091] S150: Input the magnetic field data to be detected into the trained sparse denoising autoencoder to obtain the corresponding real-time reconstruction error.

[0092] S160 compares the real-time reconstruction error with the set reconstruction error threshold, and determines whether the magnetic field data to be detected contains a magnetic anomaly signal based on the comparison result.

[0093] For example, the reconstruction error threshold is determined by the following method: collecting target magnetic field data, which is magnetic field data in the ocean superimposed with the target magnetic field and the background noise magnetic field; inputting both the target magnetic field data and the noise magnetic field data into a trained sparse denoising autoencoder to obtain the target reconstruction error and the noise reconstruction error, respectively; and determining the reconstruction error threshold based on the maximum and minimum values ​​of the target reconstruction error and the noise reconstruction error.

[0094] Furthermore, before inputting the target magnetic field data into the trained sparse denoising autoencoder, the target magnetic field data is also normalized.

[0095] Specifically, training a sparse denoising autoencoder to remove real ocean noise allows the autoencoder to learn the distribution characteristics and dependencies of the noise, but it does not learn the distribution characteristics of the target signal. Therefore, the reconstruction error obtained from the input noisy magnetic field data should be smaller than the reconstruction error of the input target magnetic field data. Only a threshold needs to be set to distinguish them, thus achieving the detection of the target magnetic anomaly signal. Furthermore, this method is better adapted to real ocean environmental noise and has greater practicality. The mean square error is used to measure the reconstruction error.

[0096]

[0097] A test dataset can be used to determine the reconstruction error threshold. This dataset contains target magnetic field data with superimposed background noise and target magnetic fields, as well as noise magnetic field data with only background noise. The target magnetic field data and noise magnetic field data are input into a trained sparse denoising autoencoder, respectively. Each sample yields a reconstruction error. The reconstruction errors of the samples are sorted by magnitude, and the two samples with the largest differences, x0 and x1, are identified. The threshold T = (L(x0) + L(x1)) / 2 is then calculated and used to detect magnetic anomaly signals.

[0098] The detection magnetic field data is divided into two cases using the reconstruction error threshold T:

[0099]

[0100] Where X0 represents no magnetic anomaly signal and X1 represents the presence of magnetic anomaly signal.

[0101] This application also provides a system for detecting background noise magnetic anomalies based on a sparse denoising autoencoder, including:

[0102] The first data acquisition module is used to acquire noise magnetic field data, which is magnetic field data in the ocean that only contains background noise magnetic field.

[0103] The noise reduction module is used to input the noisy magnetic field data into the sparse noise reduction autoencoder for sparse noise reduction processing to obtain the noise-reduced magnetic field data.

[0104] The function creation module is used to establish a loss function by utilizing the activation level of neurons in the hidden layer and the difference between the denoised magnetic field data and the expected magnetic field data.

[0105] The autoencoder training module is used to train the sparse denoising autoencoder using a loss function to obtain the trained sparse denoising autoencoder.

[0106] The second data acquisition module is used to acquire the magnetic field data to be detected.

[0107] The error calculation module is used to input the magnetic field data to be detected into the trained sparse denoising autoencoder to obtain the corresponding real-time reconstruction error.

[0108] The anomaly detection module is used to compare the real-time reconstruction error with the set reconstruction error threshold, and determine whether the magnetic field data to be detected contains a magnetic anomaly signal based on the comparison result.

[0109] This application also provides a computer storage medium storing a plurality of computer instructions for causing a computer to execute the above-described method.

[0110] The following experiment will further illustrate this point.

[0111] (1) Experimental Conditions. To verify the effectiveness of the proposed method in detecting magnetic anomaly signals containing real noise, a field experiment was conducted near Ximaozhou Island. Two identical fluxgate sensors were placed 100m apart on the seabed parallel to the shipping route, recording real-time magnetic field changes in the area at a frequency of 200Hz. There were no other ferromagnetic objects interfering with the measurement area during the experiment. The experimental vessel was 54m long, 13.2m wide, 1200 tons, and had a draft of 2.8m. The vessel passed the two bottom sensors at different transverse distances to collect magnetic anomaly signals. In addition, a large amount of real ocean noise data was collected when there were no targets. A schematic diagram of the experimental scenario is shown below. Figure 2 As shown.

[0112] Based on this, a marine magnetic dataset was established, which includes noisy magnetic field data for training the sparse denoising autoencoder and target magnetic field data and noisy magnetic field data for verifying detection performance. The collected noisy magnetic field data and target magnetic field data were used to train the sparse denoising autoencoder and to verify the performance of the reconstruction loss threshold segmentation algorithm. The sample types and quantities of each dataset are shown in Table 1. The CPU processor used in the experiment was an Intel i7-13750H with 16GB DDR5 RAM (4800MHz), and the network model was built using the PyTorch 2.0.1 framework.

[0113] The composition of the constructed ocean magnetic field dataset is shown in Table 1.

[0114]

[0115] (2) Evaluation criteria for magnetic anomaly detection performance:

[0116] (2a) Detection rate: P d (%)

[0117] Detection rate P d The definition is as follows:

[0118] P d =n tt / n t

[0119] Where, n tt n represents the number of positive samples that are judged as positive samples. t This represents the total number of positive samples.

[0120] (2b) False alarm rate: P f (%)

[0121] False alarm rate P f The definition is as follows:

[0122] P f =n ct / n c

[0123] Where, n ct n represents the number of negative samples that are classified as positive samples. c This represents the total number of negative samples.

[0124] (2c) Accuracy: Acc (%)

[0125] Accuracy (Acc) is defined as follows:

[0126] Acc=n r / n

[0127] Where, n r The number of correctly classified samples is n, where n is the total number of samples.

[0128] (2d) Detection algorithm runtime: time (s)

[0129] The magnetic anomaly detection time is calculated from the time the denoised data is input into memory until the data classification result is generated, using the difference between the time taken by the system's internal timer. Under the same computing configuration, a shorter magnetic anomaly detection time indicates a higher efficiency of the detection algorithm.

[0130] (3) Experiment content:

[0131] Experiment 1

[0132] The test dataset was compared using OBF (Orthogonal Basis Function), MED (Minimum Entropy Detector), DeepMAD (Deep Magnetic Anomaly Detection), SVM (Support Vector Machine), 1DCNN (One-Dimensional Convolutional Neural Network), IFOrest (Isolation Forest), and the SDAE (Sparse Denoising Autoencoder) of this application, respectively, and P was calculated for each. d (detection rate), P f(False alarm rate), Acc (accuracy), and processing time; detailed numerical results are shown in Table 2. An example of the denoising results is shown below. Figure 4 As shown. Wherein: Figure 4 (a) shows the experimental data results for September, and (b) shows the experimental data results for November.

[0133] Table 2 Performance comparison of magnetic anomaly detection methods on the test dataset

[0134] detector <![CDATA[P d (%)]]> <![CDATA[P f (%)]]> Acc(%) Average runtime (ms) OBF 83.12 / 83.12 16.59 MED 85.71 9.37 90.19 9.51 DeepMAD 84.42 8.63 90.76 5.75 SVM 80.95 9.56 89.76 4.73 1D CNN 89.61 11.26 89.85 12.01 IForest 92.21 10.00 90.19 14.18 SDAE 93.51 6.38 93.61 2.09

[0135] As shown in Table 2, the method of this application outperforms OBF, MED, and recently proposed machine learning methods in terms of detection rate, false alarm rate, accuracy, and processing time, exhibiting high accuracy and detection rate, extremely low false alarm rate, and fast detection runtime.

[0136] Experiment 2

[0137] To verify the impact of sparsity constraints on the detection performance of the autoencoder, ablation experiments were conducted on the proposed sparse denoising autoencoder. The detection rate, false alarm rate, and accuracy of DAE (Denoising Autoencoder) and SDAE (Simplified Denoising Autoencoder) were compared on the validation set. Furthermore, the results output from the hidden layer were visualized and compared experimentally. Figure 5 As shown, where: Figure 5 In the diagram, (a) shows the output of the hidden layer neurons of the DAE, and (b) shows the output of the hidden layer neurons of the SDAE.

[0138] Table 3 Ablation experimental results of the autoencoder on the validation dataset.

[0139] detector <![CDATA[P d (%)]]> <![CDATA[P f (%)]]> Acc(%) Average runtime (ms) DAE 90.91 7.25 92.59 2.28 SDAE 93.51 6.38 93.61 2.09

[0140] The ablation experiments above show that the activation level of hidden layer neurons is significantly reduced after adding sparsity constraints. This verifies the effectiveness of sparsity constraints in ocean magnetic field data modeling. The ablation experiment results indicate that sparsity regularization can significantly improve the performance of ocean magnetic noise denoising tasks, thus verifying the importance of explicitly modeling the sparsity distribution of the network's hidden layers.

[0141] In summary, the method presented in this application can capture the changing patterns of signals in real marine environments, thereby effectively suppressing and eliminating unknown magnetic field noise. This fully demonstrates the generalization ability of data-driven, deep sparse denoising autoencoder networks that model the sparsity distribution of hidden layers. Furthermore, its end-to-end training method overcomes the limitation of traditional approaches that require manual extraction of signal features, providing a new solution to the problem of magnetic field sensing and analysis in the field of marine exploration.

[0142] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0143] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A background noise magnetic anomaly detection method based on a sparse denoising autoencoder, characterized in that, include: Collect noise magnetic field data, wherein the noise magnetic field data is magnetic field data in the ocean that only contains background noise magnetic field; The noise magnetic field data is input into a sparse denoising autoencoder for sparse denoising processing to obtain denoised magnetic field data. A loss function is established using the activation level of neurons in the hidden layer and the difference between the denoised magnetic field data and the desired magnetic field data. The sparse denoising autoencoder is trained using the loss function to obtain a trained sparse denoising autoencoder. Collect data on the magnetic field to be detected; The magnetic field data to be detected is input into the trained sparse denoising autoencoder to obtain the corresponding real-time reconstruction error. The real-time reconstruction error is compared with a set reconstruction error threshold, and the magnetic field data to be detected is determined to be free of magnetic anomaly signals based on the comparison result. The reconstruction error threshold is determined by the following method: Collect target magnetic field data, which is magnetic field data in the ocean superimposed with the target magnetic field and the background noise magnetic field; Both the target magnetic field data and the noise magnetic field data are input into the trained sparse denoising autoencoder to obtain the target reconstruction error and the noise reconstruction error, respectively. The reconstruction error threshold is determined based on the maximum and minimum values ​​of the target reconstruction error and the noise reconstruction error.

2. The background noise magnetic anomaly detection method based on a sparse denoising autoencoder according to claim 1, characterized in that, The method for sparse denoising autoencoder to perform sparse denoising processing on the noisy magnetic field data is as follows: The noise magnetic field data is input into the first convolutional layer, and the features of the noise magnetic field data are extracted to obtain the first encoded data; The first encoded data is input into the second convolutional layer to extract the features of the first encoded data and obtain the second encoded data. The second encoded data is input into the third convolutional layer to extract the features of the second encoded data and obtain the third encoded data; The third encoded data is input into the first deconvolution layer to obtain the first decoded data; The first decoded data is input into the second deconvolution layer to obtain the second decoded data; The second decoded data is input into the third deconvolution layer to obtain the denoised magnetic field data.

3. The background noise magnetic anomaly detection method based on a sparse denoising autoencoder according to claim 2, characterized in that, Before inputting the noisy magnetic field data into the first convolutional layer, the noisy magnetic field data is also normalized.

4. The background noise magnetic anomaly detection method based on a sparse denoising autoencoder according to claim 2, characterized in that, After extracting the noise magnetic field data, the first coded data, and the second coded data respectively for feature extraction, the first convolutional layer, the second convolutional layer, and the third convolutional layer also perform filling and activation processing on the extracted feature data in sequence. After performing deconvolution processing on the third encoded data, the first decoded data, and the second decoded data respectively, the first deconvolution layer, the second deconvolution layer, and the third deconvolution layer also sequentially perform shifting, superposition, and activation processing on the deconvolutioned data.

5. The method for detecting background noise magnetic anomalies based on a sparse denoising autoencoder according to claim 1, characterized in that, The method for establishing the loss function is as follows: Determine the activation level of each neuron in the hidden layer; Determine the distance between the activation level and the expected activation level, and use the distance as a penalty term for the hidden layer; The difference item is determined based on the difference between the denoised magnetic field data and the desired magnetic field data; The loss function is obtained by weighted summation of the penalty term and the difference term.

6. The method for detecting background noise magnetic anomalies based on a sparse denoising autoencoder according to claim 1, characterized in that, During the training of the sparse denoising autoencoder, the Adam optimizer is used to update the convolution kernel and bias parameters of the sparse denoising autoencoder.

7. The method for detecting background noise magnetic anomalies based on a sparse denoising autoencoder according to claim 1, characterized in that, Before inputting the target magnetic field data into the trained sparse denoising autoencoder, the target magnetic field data is also normalized.

8. A system for detecting background noise magnetic anomalies based on a sparse denoising autoencoder as described in any one of claims 1-7, characterized in that, include: The first data acquisition module is used to acquire noise magnetic field data, which is magnetic field data in the ocean that only contains background noise magnetic field. The noise reduction processing module is used to input the noise magnetic field data into the sparse noise reduction autoencoder for sparse noise reduction processing to obtain noise-reduced magnetic field data. The function establishment module is used to establish a loss function by utilizing the activation level of neurons in the hidden layer and the difference between the denoised magnetic field data and the desired magnetic field data. The autoencoder training module is used to train the sparse denoising autoencoder using the loss function to obtain a trained sparse denoising autoencoder. The second data acquisition module is used to acquire the magnetic field data to be detected. The error calculation module is used to input the magnetic field data to be detected into the trained sparse denoising autoencoder to obtain the corresponding real-time reconstruction error. An anomaly detection module is used to compare the real-time reconstruction error with a set reconstruction error threshold, and determine whether the magnetic field data to be detected contains a magnetic anomaly signal based on the comparison result.

9. A computer storage medium, characterized in that, The computer storage medium stores a plurality of computer instructions, which are used to cause the computer to perform the method described in any one of claims 1-7.

Images