CN116304561A - Ocean target magnetic anomaly signal denoising and detecting method based on deep migration learning - Google Patents
Ocean target magnetic anomaly signal denoising and detecting method based on deep migration learning Download PDFInfo
- Publication number
- CN116304561A CN116304561A CN202310106011.7A CN202310106011A CN116304561A CN 116304561 A CN116304561 A CN 116304561A CN 202310106011 A CN202310106011 A CN 202310106011A CN 116304561 A CN116304561 A CN 116304561A
- Authority
- CN
- China
- Prior art keywords
- data
- denoising
- magnetic field
- magnetic
- ocean
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R33/00—Arrangements or instruments for measuring magnetic variables
- G01R33/0023—Electronic aspects, e.g. circuits for stimulation, evaluation, control; Treating the measured signals; calibration
- G01R33/0029—Treating the measured signals, e.g. removing offset or noise
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R33/00—Arrangements or instruments for measuring magnetic variables
- G01R33/02—Measuring direction or magnitude of magnetic fields or magnetic flux
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/30—Assessment of water resources
Abstract
The application discloses a marine target magnetic anomaly signal denoising and detecting method based on deep transfer learning, which comprises the following steps: constructing a simulated marine target magnetic anomaly signal data set by adopting a magnetic dipole model; establishing an actually measured ocean magnetic field noise data set; adding the data in the actually measured ocean magnetic field noise data set into the simulated ocean target magnetic anomaly signal data set to form a simulated noise-containing ocean target magnetic anomaly data set; training a denoising self-encoder network by adopting a simulated noisy marine target magnetic anomaly data set; training a fully connected classifier network by adopting denoising magnetic field data; and acquiring magnetic field data to be detected, and denoising and detecting by using the trained denoising self-encoder network and the fully-connected classifier network to obtain a final magnetic anomaly signal detection result. According to the method, the deep neural network model can be trained under the condition of a small amount of measured data, and compared with a traditional signal processing method, the method can realize faster denoising of magnetic field data and detection of magnetic abnormal signals.
Description
Technical Field
The application relates to the technical field of signal and information processing, in particular to a method for denoising and detecting a magnetic anomaly signal of an ocean target based on deep migration learning.
Background
The substance with ferromagnetic material is affected by the combination of permanent magnetism and the magnetization of the surrounding geomagnetic field, so that the geomagnetic field around the substance is locally distorted, and a quasi-stationary magnetic phenomenon, called a magnetic anomaly signal, is generated. The magnetic abnormal signal is used as an important physical field source of the ocean target, is not limited by conditions such as ocean topography, ocean environment and the like, has high sensitivity and resolution, can be compatible with various electronic instruments, and can provide a valuable signal source for ship navigation, direction identification and the like.
The current detection methods for the magnetic anomaly signals of the ocean targets are mainly divided into two types: one is to perform theoretical modeling on a marine target magnetic anomaly signal, analyze the characteristics of the signal and the energy domain of the signal, and then judge whether the target magnetic anomaly signal exists or not in the characteristic domain; the other is to take the signal characteristic research of the background magnetic field as an entry point, analyze the statistical distribution characteristic of the noise of the ocean background magnetic field, mine the characteristic difference between the noise and the target magnetic abnormal signal, and enhance and detect the target magnetic abnormal signal by utilizing noise filtering and inhibiting means. Common magnetic anomaly signal detection methods include orthogonal basis functions (Orthonormal Basis Function, OBF), minimum entropy detectors (Minimum Entropy Detector, MED), wavelet transform (Wavelet Transform, WT) denoising, and the like.
OBF is a detection method based on the characteristics of target magnetic anomaly signals. According to a physical model of a target magnetic anomaly signal, the magnetic anomaly signal sequence can be represented by using three mutually incoherent basis functions through linear calculation, and the basis functions can be converted into OBF after orthogonalization, normalization and other treatments. Based on the above, the corresponding magnetic anomaly signal representation coefficient is calculated according to the property of the basis function, and the existence of the target magnetic anomaly signal in the time sequence is judged according to the calculated magnetic anomaly signal representation coefficient. MED is a magnetic anomaly signal detection method based on information entropy features. Compared with the pure ocean environment magnetic field noise, the target magnetic anomaly signal generally has higher order, and the information entropy is an index for measuring the order strength of the signal. The unordered distribution of the ocean magnetic field noise is affected by the occurrence of the target magnetic anomaly signals, so that the information entropy of the time sequence is correspondingly changed. Therefore, the measured ocean magnetic field time sequence can be converted into an information entropy sequence corresponding to the ocean magnetic field time sequence, and the change trend of the information entropy is observed to set a threshold value according to the information entropy, so that detection of the target magnetic abnormal signal is realized. The WT is a signal analysis method which has the characteristics of removing correlation, being applicable to various resolutions, being capable of coping with local time-frequency changes and the like. The key of the denoising effect of the wavelet threshold algorithm lies in the setting of a threshold value and the construction of a threshold function, the wavelet transform-based method is widely applied in the denoising field, and the denoised data is easier to judge whether magnetic abnormal signals are contained according to the signal characteristics of the denoised data.
The magnetic anomaly signal detection method has dependence on priori knowledge, and ideal assumption is made on signal characteristics, so that the application effect of the magnetic anomaly signal detection method in practical problems is limited. The OBF detector can only work normally in Gaussian white noise environment, and the application of the OBF detector in complex ocean noise environment is severely restricted. MED distinguishes the magnetic anomaly of the target based on the information entropy difference of the signal and the noise, but the distinguishing capability is limited when the signal-to-noise ratio is low, and the remote detection requirement of the magnetic anomaly of the target cannot be met. WTs rely on different characteristics of signal and noise in the wavelet function space for noise suppression, and their implementation involves artificial model design and parameter tuning, lacking in adaptation to specific problems. In summary, for research on the magnetic anomaly signal detection algorithm, the conventional method is generally poor in universality due to the limitation of specific conditions, and is difficult to effectively apply to complex and changeable marine environments. Therefore, development of a general ocean target magnetic anomaly signal detection model and algorithm is needed to realize suppression of ocean stray magnetic field noise and detection of weak magnetic anomaly signals, and an effective solution is provided for accurate perception of magnetic targets in a complex ocean environment.
In recent years, deep neural networks are receiving attention and application in various fields due to flexible model structures and strong learning capabilities, and have achieved superior performance compared with the conventional methods. The deep learning technology is hopeful to provide a general solution for the detection of the magnetic anomaly signals of the ocean targets, an integrated network model structure combining the denoising and detection functions is formed, and the comprehensive sensing capability of the magnetic anomaly signals under the condition of low signal-to-noise ratio is further improved. However, the measured ocean magnetic field data has the problems of sample scarcity and class imbalance, and the learning and training of a magnetic anomaly perception model based on a deep neural network are severely restricted.
Disclosure of Invention
The embodiment of the application provides a deep migration learning-based ocean target magnetic anomaly signal denoising and detecting method, which is used for solving the problems caused by sample scarcity and class unbalance of actually measured ocean magnetic field data in the prior art.
On one hand, the embodiment of the application provides a marine target magnetic anomaly signal denoising and detecting method based on deep migration learning, which comprises the following steps:
constructing a simulated marine target magnetic anomaly signal data set by adopting a magnetic dipole physical field model;
collecting actual measurement ocean magnetic field noise data, and establishing an actual measurement ocean magnetic field noise data set;
adding the data in the actually measured ocean magnetic field noise data set into the simulated ocean target magnetic anomaly signal data set to form a simulated noise-containing ocean target magnetic anomaly data set;
constructing a denoising self-encoder network;
denoising the simulated noise-containing marine target magnetic anomaly data set by adopting a denoising self-encoder network to obtain denoising magnetic field data, and adjusting parameters of the denoising self-encoder network by utilizing the denoising magnetic field data;
constructing a fully connected classifier network;
classifying the denoising magnetic field data by adopting a full-connection classifier network to obtain a classification result, and adjusting parameters of the full-connection classifier network by utilizing the classification result;
and acquiring magnetic field data to be detected, inputting the magnetic field data to be detected into the adjusted denoising self-encoder network to perform denoising processing to obtain the denoising data to be detected, inputting the denoising data to be detected into the adjusted full-connection classifier network to perform classification processing to obtain a final magnetic anomaly signal detection result.
The ocean target magnetic anomaly signal denoising and detecting method based on deep migration learning has the following advantages:
1. aiming at the problems that the acquisition cost of the magnetic anomaly signals of the ocean targets is high and a complete ocean magnetic field data set is not yet used for deep learning, the method combines a magnetic dipole physical field model, and when the distance between the ocean targets and the magnetic sensor is greater than twice the maximum size of the targets, the simulated generated magnetic field passing characteristic curve and the real magnetic field passing characteristic curve have high similarity. On the basis, the method and the device provide that a training data set is constructed by using simulation magnetic field data for combined denoising and detection model learning, and the extracted hierarchical features are transferred to real ocean magnetic field data through a deep transfer learning technology, so that the extraction and mining of target features in the ocean magnetic field data are realized, and the deep learning bottleneck problem of small-sample ocean magnetic field data is well solved.
2. Aiming at the problems of low efficiency, poor robustness and serious dependence on manual design and parameter selection of the traditional denoising method, the application provides a marine magnetic field noise suppression model and algorithm based on a convolution denoising self-encoder, which can realize effective modeling and learning of noise with different distributions, fully excavate multi-scale time correlation characteristics of marine magnetic field data, and has higher signal-to-noise ratio lifting amplitude and faster data processing speed compared with the traditional denoising method. The convolution denoising self-encoder has strong universality, and after the model is trained by a large amount of actually measured ocean magnetic field noise data, the model can be used for an automatic denoising task under a complex interference environment, so that the problem of performance limitation existing in the traditional denoising method is effectively solved.
3. Aiming at the problems that the traditional magnetic anomaly denoising and detection method relies on manual waveform interpretation and priori parameter information of a detection scene is required to be known in advance, the method can realize intelligent perception of the target magnetic anomaly signal in the ocean magnetic field data by adopting a neural network-based ocean target magnetic anomaly signal detection model and algorithm. The method provided by the application can automatically judge whether the data contains the target magnetic anomaly signal without manual intervention or acquiring expensive scene parameter information, effectively overcomes the defect that the traditional method needs to judge manually according to waveform characteristics, and improves the detection precision and efficiency of the marine target magnetic anomaly signal under the condition of low signal-to-noise ratio.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present application, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flowchart of a method for denoising and detecting a magnetic anomaly signal of an ocean target based on deep transfer learning according to an embodiment of the present application;
fig. 2 is a schematic diagram of deep migration learning using simulation data according to an embodiment of the present application;
FIG. 3 is a schematic diagram of a scenario for constructing simulation data using a magnetic dipole physical field model according to an embodiment of the present disclosure;
fig. 4 is a schematic structural diagram of a denoising self-encoder network and a fully-connected classifier network according to an embodiment of the present application;
FIG. 5 is a diagram of denoising results of a marine target magnetic anomaly signal provided in an embodiment of the present application; wherein (a) is an original noise-free ocean target magnetic anomaly signal, (b) is an ocean target magnetic anomaly signal after noise is added, and (c) is a result graph after noise-containing data is denoised by using the method;
fig. 6 is a comparison chart of denoising results of marine target magnetic anomaly signals after four-level decomposition by using a wavelet transform denoising method according to an embodiment of the present application.
Detailed Description
The following description of the embodiments of the present application will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are only some, but not all, of the embodiments of the present application. All other embodiments, which can be made by one of ordinary skill in the art without undue burden from the present disclosure, are within the scope of the present disclosure.
Fig. 1 is a flowchart of a method for denoising and detecting a magnetic anomaly signal of an ocean target based on deep migration learning according to an embodiment of the present application. The embodiment of the application provides a marine target magnetic anomaly signal denoising and detecting method based on deep migration learning, which comprises the following steps:
s100, constructing a simulated marine target magnetic anomaly signal data set by adopting a magnetic dipole physical field model.
Illustratively, S100 specifically includes: according to a magnetic dipole physical field model, the ocean target is equivalent to a magnetic dipole, and as shown in fig. 3, a calculation expression of an induced magnetic field vector generated by the ocean target at a magnetic sensor measuring point is obtained; changing the motion speed of the ocean target and the positive transverse distance and the vertical distance between the ocean target and the magnetic sensor, and determining an induced magnetic field vector time sequence by the magnetic sensor according to a calculation expression of the induced magnetic field vector; and carrying out normalization processing on the induced magnetic field vector time sequence to obtain a simulated marine target magnetic anomaly signal data set.
Before determining the induced magnetic field vector calculation expression, a spatial three-dimensional coordinate system O-xyz for measuring the magnetic field of the magnetic sensor marine target is also established, and under the spatial three-dimensional coordinate system, the induced magnetic field vector is calculated
The calculated expression of (2) is:
wherein mu is 0 Is the magnetic permeability of the vacuum and is equal to the magnetic permeability of the vacuum,
is the displacement vector of the ocean object and the magnetic sensor in the space three-dimensional coordinate system,
is the magnetic moment vector of the ocean target to be measured. The space intersection situation of the ocean target and the magnetic sensor is set by changing the motion speed v of the ocean target and the positive transverse distance a and the vertical distance h between the ocean target and the magnetic sensor. Sampling frequency f along with magnetic sensor s Generating a displacement vector +.>
Then according to the calculation expression of the induced magnetic field vector +.>
Further carrying out normalization processing on the time sequence of the simulation ocean target magnetic anomaly signal data set to construct the simulation ocean target magnetic anomaly signal data set
S110, collecting actually measured ocean magnetic field noise data and establishing an actually measured ocean magnetic field noise data set.
Illustratively, S110 specifically includes: collecting actual measurement ocean magnetic field noise data; carrying out subsection interception and augmentation expansion on actually measured ocean magnetic field noise data; normalization processing is carried out on the actually measured ocean magnetic field noise data obtained after segmentation interception and augmentation expansion, and an actually measured ocean magnetic field noise data set is obtained.
Normalization processing is carried out on the actually measured ocean magnetic field noise data obtained after segmentation interception and augmentation expansion, and then an actually measured ocean magnetic field noise data set can be constructed
And S120, adding the data in the actually measured ocean magnetic field noise data set into the simulated ocean target magnetic anomaly signal data set to form a simulated noise-containing ocean target magnetic anomaly data set.
Illustratively, S120 specifically includes: and adding the data in the actually measured ocean magnetic field noise data set into the simulated ocean target magnetic anomaly signal data set in an amplitude modulation manner according to different signal to noise ratios, and cutting off the added data set to obtain the simulated noise-containing ocean target magnetic anomaly data set, as shown in fig. 2.
The actually measured ocean magnetic field noise data set is added into the simulated ocean target magnetic anomaly signal data set S, and the simulated noise-containing ocean target magnetic anomaly data set is obtained after the truncation processing
y (j-1)*q+k ＝f(s j +αn k ),j＝1,2,…,p；k＝1,2,…,q
Wherein, alpha is noise amplitude modulation factor, which is used for controlling the simulated noise-containing marine target magnetic anomaly signal generating specific signal-to-noise ratio, and f (·) is a truncated function with the following form:
s130, constructing a denoising self-encoder network.
Illustratively, the denoising self-encoder network includes a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, a first deconvolution layer, a second deconvolution layer, and a third convolutional layer, which are sequentially arranged, as shown in fig. 4.
And S140, denoising the simulated noise-containing marine target magnetic anomaly data set by adopting a denoising self-encoder network to obtain denoising magnetic field data, and adjusting parameters of the denoising self-encoder network by utilizing the denoising magnetic field data.
Illustratively, when denoising the simulated noisy marine target magnetic anomaly data set by adopting the denoising self-encoder network, inputting the simulated noisy marine target magnetic anomaly data set into a first convolution layer, and performing dot multiplication and addition operation on input data by each convolution check in the first convolution layer to obtain first convolution output data; inputting the first convolution output data into a first maximum pooling layer to obtain first pooled output data; inputting the first pooled output data into a second convolution layer, and performing dot multiplication and addition operation on the input data by each convolution check in the second convolution layer to obtain second convolution output data; inputting the second convolution output data into a second maximum pooling layer to obtain second pooled output data; inputting the second pooled output data into a first deconvolution layer, and respectively carrying out deconvolution operation on the input data by each deconvolution core in the first deconvolution layer to obtain first deconvolution output data; inputting the first deconvolution output data into a second deconvolution layer, and respectively carrying out deconvolution operation on the input data by each deconvolution core in the second deconvolution layer to obtain second deconvolution output data; and inputting the second deconvoluted output data into a third convolution layer, and performing dot multiplication and addition operation on the input data by each convolution check in the third convolution layer to obtain denoising magnetic field data.
Specifically, S140 includes the following sub-steps:
1) Simulating noise-containing marine target magnetic anomaly signals
Input to the first convolution layer (convolution kernel C1, offset b 1 i ) Each convolution kernel performs dot multiplication and phase on the input dataThe adding operation, the first convolution output data generated is:
wherein y is k To simulate the kth element of the noisy marine target magnetic anomaly signal,
the (k-i+1) th element of the ith convolution kernel, and the step size of the convolution operation is selected to be 1. For the final convolution data, 0 elements are filled back when it is less than m+1 elements, and then x is generated by ReLU (Rectified LinearUnit) activation function i ＝max(0,x i ). Performing a convolution operation using 32 convolution kernels to form first convolution output data +.>
2) The data after the operation of the first convolution layer is sent to a first 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 ) Filling back 0 elements when the data entering the window is less than 4 elements to form a first pooled output data
3) Pooling the first output data
Through a second convolution layer, where each channel of the input data corresponds to 1 convolution kernel (the convolution kernel is C2 i Offset b 2_i ) Each convolution check performs dot multiplication and addition operation on input data, and new data is generated as follows:
wherein x is k For the kth element of the first pooled output data,
the (k-i+1) th element of the ith convolution kernel, and the step size of the convolution operation is selected to be 1. For the final convolution data, 0 elements are filled backwards when the convolution data is less than m+1 elements, and then new data x is generated through a ReLU activation function i ＝max(0,x i ) To form a second convolved output data
4) Outputting the second convolution output data
Through the second maximum pooling layer, the size of the pooling window is 4, and the operation performed can be expressed as x i ＝max(x 4*i-3 ,x 4*i-2 ,x 4*i-1 ,x 4*i ) Filling 0 elements backwards when the data entering the window is less than 4 elements to form second pooled output data +.>
5) Pooling the second pooled output data
Through the first deconvolution layer (ith deconvolution kernel +.>
The offset is +.>
) A total of 32 deconvolution kernels are respectively pair +.>
Is deconvoluted for 32 channels with a shift step of 4, < >>
A new data amount of the superimposed value is generated for the deconvolution kernel to move a step size smaller than the deconvolution kernel. The data generated by the deconvolution operation can be expressed as +.>
Then the output data generated by the ReLU activation function is x i ＝max(0,x i ) The first deconvoluted output data after passing through the first deconvolution layer may be expressed as +.>
6) Outputting the first deconvolution data
Through the second deconvolution layer (ith deconvolution kernel is
The offset is +.>
) A total of 32 deconvolution kernels are respectively pair +.>
Is deconvoluted for 32 channels with a shift step of 4, < >>
A new data amount of the superimposed value is generated for the deconvolution kernel to move a step size smaller than the deconvolution kernel. The data generated by the deconvolution operation can be expressed as +.>
Then the output data generated by the ReLU activation function is x i ＝max(0,x i ) The second deconvoluted output data after passing through the second deconvolution layer can be expressed as +.>
7) Outputting the second deconvolution data
Through the third convolution layer (convolution kernel C3, offset b 3_i ) Each channel of the X is subjected to dot multiplication and addition operation by 1 convolution kernel, and the generated new data is:
wherein x is k Outputting the kth element of the data for the second deconvolution, C3 k-i+1 Is the k-i+ 1 element of the third convolution layer, and the step size of the convolution operation is selected to be 1. For the final convolution data, 0 elements are filled backwards when the final convolution data is less than m+1 elements, and then the generated 32-channel data corresponding points are added to obtain x i ＝x 1_i +x 2_i +…+x 32_i Finally, the data generated through the sigmoid activation function is that
The denoised magnetic field data after passing through the third convolution layer can be expressed as + ->
Further, when parameters of the denoising self-encoder network are adjusted by using the denoising magnetic field data, taking the mean square error of the denoising magnetic field data and the data in the simulation noise-containing ocean target magnetic anomaly data set as a loss function, and determining the parameters of the denoising self-encoder network at the next moment based on the loss function.
The denoising magnetic field data x output by the denoising self-encoder network is consistent with the dimension y of the input simulation noise-free ocean target magnetic anomaly signal, a Adam (Adaptive Moment Estimation) optimizer is used for learning and updating the convolution kernel and offset (represented by theta) of the network, and the adopted loss function
Mean square error for x and y:
solving for a loss function
For the network parameter theta at the time t t Partial derivative g of (2) t Obtaining:
next, the first and second moment estimates m of the gradient in momentum form at time t are obtained t And v t ：
On the basis, first-order and second-order moment estimation after correction of the deviation at the moment t is obtained
And->
Then get the network parameter theta at the time t+ 1 t+1 The updated formula of (c) is as follows:
where η and ε are the iteration steps and a small number. The convolution kernel and the offset of the denoising self-encoder network are continuously adjusted and updated through the iterative training process, so that the structure of the denoised magnetic field data after network operation is approximate to the structure of the denoised magnetic field data, and the training of the denoising self-encoder network is completed.
S150, constructing a full-connection classifier network.
The fully-connected classifier network illustratively comprises a fully-connected layer, a ReLU activation function and a Softmax layer which are sequentially arranged, as shown in FIG. 4, wherein the input weights of the fully-connected layer are as follows
The offset is +.>
S160, classifying the denoising magnetic field data by using a full-connection classifier network to obtain a classification result, and adjusting parameters of the full-connection classifier network by using the classification result.
Illustratively, when the denoising magnetic field data is classified by adopting a fully-connected classifier network, the denoising magnetic field data is input into a fully-connected layer to obtain intermediate output, and the intermediate output is subjected to a ReLU activation function to obtain output data of the fully-connected layer; and inputting the output data of the full-connection layer into the Softmax layer, obtaining the output probability value of each node in the Softmax layer, and taking the class corresponding to the node with the largest output probability value as the classification result of the denoising magnetic field data.
The intermediate output may be expressed as:
will output the middle
Generating an output y=max (0, h) of the full connection layer through the ReLU activation function, and having
Finally, the output of the full-connection layer is->
Obtaining an output vector through a Softmax layer with the node number of C
And thereby generating an output probability value for each node:
wherein p is i ∈[0,1]An output probability value for the ith node, and has
And selecting a class corresponding to the node with the largest output probability value as a classification result of the denoising magnetic field data, and judging whether a target magnetic abnormal signal exists in the data sample or not according to the classification result.
Further, when the parameters of the fully connected classifier network are adjusted by using the classification result, taking the cross entropy between the output probability value of each node in the Softmax layer and the corresponding real class mark as a loss function, and determining the parameters of the fully connected classifier network at the next moment based on the loss function.
In training of the full-connection classifier network, the output probability value of the node is calculated
With true category labels
Comparing, and constructing a cross entropy loss function L for full-connection classifier network parameter learning:
and then, according to the same training step as that of the denoising self-encoder network in the step S140, iteratively updating parameters of the full-connection classifier network, thereby completing training of the full-connection classification detector network.
S170, collecting magnetic field data to be detected, inputting the magnetic field data to be detected into an adjusted denoising self-encoder network to perform denoising processing to obtain denoising data to be detected, inputting the denoising data to be detected into an adjusted full-connection classifier network to perform classification processing to obtain a final magnetic anomaly signal detection result.
The effect of the present application is further illustrated by the following simulation experiments:
(1) Experimental simulation conditions:
the real data adopted in the experiment are the magnetic field data of the water surface ship and the ocean environment collected in three ways, the sampling frequency of the fluxgate sensor is 200Hz, each section of selected data sample contains 80000 magnetic field sampling points, the simulated magnetic field data are generated according to the same sampling frequency and data length as the real data, and the detailed composition condition of the ocean magnetic field data set constructed in the experiment is shown in table 1. The experiment is based on Intel (R) Core (TM) i7-8750H CPU running with a main frequency of 2.20GHz and a memory of 16GB, MATLAB R2021a software is adopted on a Windows10 operating system to simulate magnetic field data, and Pycharm2021 software, python3.8 programming language and TensorFlow2.2 deep learning framework are used for combined denoising and construction and learning training of a detection network model.
Details of the composition of the ocean magnetic field data set constructed in Table 1
(2) Data denoising performance evaluation criterion:
(2a) And (4) improving the signal-to-noise ratio before and after data denoising: signal to noise ratio improvement (dB)
When the signal-to-noise ratio of the data before and after denoising is calculated, the data is normalized. Assume that
Is normalized raw noise-free data, < >>
Is normalized noisy data, +.>
Is the normalized denoising data, the signal to noise ratio improves the SNR improve Is defined as follows:
signal to noise ratio improving SNR improve The larger the value of (c) is, the better the performance of the denoising algorithm is.
(2b) The operating efficiency of the denoising algorithm: time(s)
The data denoising process starts from the time when the noisy data is put into the memory, and the difference value of the internal timer of the system until the denoised data is generated is calculated. Under the same calculation configuration condition, the shorter the denoising processing time is, the higher the operation efficiency of the denoising algorithm is.
(3) Magnetic anomaly detection performance evaluation criteria:
(3a) Detection rate: p (P) d (％)
Detection rate P d Is defined as follows:
P d ＝n tt /n t
wherein n is tt Is the number of positive samples, n t Is the positive sample total.
(3b) False alarm rate: p (P) f (％)
False alarm rate P f Is defined as follows:
P f ＝n ct /n c
wherein n is ct The negative samples are judged to be the number of positive samples, n c Is the negative sample total.
(3c) Accuracy rate: acc (%)
The definition of the accuracy Acc is as follows:
Acc＝n r /n
wherein n is r For the number of correctly classified samples, n is the total number of samples.
(3d) The operation efficiency of the detection algorithm: time(s)
The time for detecting the magnetic anomaly is from the time when the denoised data is input into the memory, and the difference value of the internal timer of the system is calculated until the data classification result is generated. Under the same calculation configuration condition, the shorter the magnetic anomaly detection time is, the higher the operation efficiency of the detection algorithm is.
(4) The experimental contents are as follows:
experiment one
After the constructed ocean magnetic field data set is used for carrying out noise adding treatment on the data, the denoising self-encoder network and the traditional wavelet transformation method are respectively used for denoising, and experimental results are respectively shown in fig. 5 and 6. In addition, the denoising performance of the method (denoted by CDAE) and the wavelet transform method (denoted by WT) on the test data set were compared, and the denoising effect and efficiency were evaluated by using signal-to-noise ratio improvement (in dB) and average running time (in ms) indexes, and the specific statistical results are shown in table 2.
Table 2 comparison of denoising Performance of different denoising methods on test dataset
| Denoising device | Signal to noise ratio improvement (dB) | Average run time (ms) |
| WT | 15 | 29 |
| CDAE | 19 | 22 |
As can be seen from the results in Table 2, compared with the traditional wavelet transform method, the denoising method provided by the application has higher signal-to-noise ratio lifting amplitude and faster denoising processing speed, and is beneficial to realizing rapid detection of magnetic abnormal signals under the condition of low signal-to-noise ratio.
Experiment two
The full-connection classifier network (expressed by FCC) is used for detecting the magnetic anomaly signals in the test data set, and comparing the performance of the traditional OBF and MED magnetic anomaly denoising and detecting methods, and respectively obtaining the detection rate P d False alarm rate P f The detection effect and efficiency were evaluated by the accuracy Acc and the average running time (in ms) index, and the specific statistical results are shown in table 3.
TABLE 3 comparison of detection Performance of different magnetic anomaly denoising and detection methods on test data sets
As can be seen from the results in Table 3, the denoising and detecting method provided by the application is obviously superior to the traditional OBF and MED methods in terms of detection rate, false alarm rate, accuracy and running time index, and has the advantages of higher detection accuracy and detection rate, extremely low false alarm rate and faster detection running time.
Therefore, the magnetic anomaly signal denoising and detecting method provided by the application adopts the intelligent and integrated deep neural network architecture driven by data and the learning sensing strategy, can greatly reduce the dependence on scene priori knowledge and artificial rule design, effectively improves the detection capability of a magnetic measurement system on the magnetic anomaly signal of a target with low signal to noise ratio, and has wide application prospect in the field of remote magnetic detection of the marine target under the complex interference environment.
While preferred embodiments of the present application have been described, additional variations and modifications in those embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. It is therefore intended that the following claims be interpreted as including the preferred embodiments and all such alterations and modifications as fall within the scope of the application.
It will be apparent to those skilled in the art that various modifications and variations can be made in the present application without departing from the spirit or scope of the application. Thus, if such modifications and variations of the present application fall within the scope of the claims and the equivalents thereof, the present application is intended to cover such modifications and variations.
Claims (8)
1. The method for denoising and detecting the magnetic anomaly signal of the marine target based on deep migration learning is characterized by comprising the following steps:
constructing a simulated marine target magnetic anomaly signal data set by adopting a magnetic dipole physical field model;
collecting actual measurement ocean magnetic field noise data, and establishing an actual measurement ocean magnetic field noise data set;
adding the data in the actually measured ocean magnetic field noise data set into the simulated ocean target magnetic anomaly signal data set to form a simulated noise-containing ocean target magnetic anomaly data set;
constructing a denoising self-encoder network;
denoising the simulated noisy marine target magnetic anomaly data set by adopting the denoising self-encoder network to obtain denoising magnetic field data, and adjusting parameters of the denoising self-encoder network by utilizing the denoising magnetic field data;
constructing a fully connected classifier network;
classifying the denoising magnetic field data by adopting the fully-connected classifier network to obtain a classification result, and adjusting parameters of the fully-connected classifier network by utilizing the classification result;
and acquiring magnetic field data to be detected, inputting the magnetic field data to be detected into the adjusted denoising self-encoder network to perform denoising processing to obtain denoising data to be detected, inputting the denoising data to be detected into the adjusted full-connection classifier network to perform classification processing to obtain a final magnetic anomaly signal detection result.
2. The method for denoising and detecting the magnetic anomaly signal of the marine target based on the deep transfer learning according to claim 1, wherein the constructing the simulated magnetic anomaly signal data set of the marine target by using the magnetic dipole physical field model comprises the following steps:
according to a magnetic dipole physical field model, the ocean target is equivalent to a magnetic dipole, and a calculation expression of an induced magnetic field vector generated by the ocean target at a magnetic sensor measuring point is obtained;
changing the motion speed of the ocean target and the positive transverse distance and the vertical distance between the ocean target and the magnetic sensor, and determining an induced magnetic field vector time sequence by the magnetic sensor according to the calculated expression of the induced magnetic field vector;
and carrying out normalization processing on the induced magnetic field vector time sequence to obtain the simulated marine target magnetic anomaly signal data set.
3. The method for denoising and detecting the magnetic anomaly signal of the marine target based on deep transfer learning according to claim 1, wherein the collecting the measured marine magnetic field noise data and establishing the measured marine magnetic field noise data set comprises:
collecting actual measurement ocean magnetic field noise data;
intercepting and expanding the actually measured ocean magnetic field noise data in a segmented way;
and carrying out normalization processing on the actually measured ocean magnetic field noise data subjected to segmentation interception and augmentation expansion to obtain the actually measured ocean magnetic field noise data set.
4. The method for denoising and detecting the magnetic anomaly signal of the marine target based on deep transfer learning according to claim 1, wherein the adding the data in the measured marine magnetic field noise data set into the simulated marine target magnetic anomaly signal data set to form the simulated noisy marine target magnetic anomaly data set comprises:
and adding the data in the actually measured ocean magnetic field noise data set into the simulated ocean target magnetic anomaly signal data set in an amplitude modulation manner according to different signal to noise ratios, and obtaining the simulated noise-containing ocean target magnetic anomaly data set after cutting off the added data set.
5. The method for denoising and detecting the magnetic anomaly signal of the marine target based on the deep transfer learning according to claim 1, wherein the denoising self-encoder network comprises a first convolution layer, a first maximum pooling layer, a second convolution layer, a second maximum pooling layer, a first deconvolution layer, a second deconvolution layer and a third convolution layer which are sequentially arranged, when the denoising self-encoder network is adopted to denoise the magnetic anomaly data set of the marine target with the simulation noise, the magnetic anomaly data set of the marine target with the simulation noise is input into the first convolution layer, and each convolution check input data in the first convolution layer is subjected to dot multiplication and addition operation to obtain first convolution output data; inputting the first convolution output data into the first maximum pooling layer to obtain first pooled output data; inputting the first pooled output data into the second convolution layer, and performing dot multiplication and addition operation on the input data by each convolution check in the second convolution layer to obtain second convolution output data; inputting the second convolution output data into the second maximum pooling layer to obtain second pooled output data; inputting the second pooled output data into the first deconvolution layer, and respectively carrying out deconvolution operation on the input data by each deconvolution core in the first deconvolution layer to obtain first deconvolution output data; inputting the first deconvolution output data into the second deconvolution layer, and respectively carrying out deconvolution operation on the input data by each deconvolution core in the second deconvolution layer to obtain second deconvolution output data; and inputting the second deconvolution output data into the third convolution layer, and performing dot multiplication and addition operation on the input data by each convolution check in the third convolution layer to obtain the denoising magnetic field data.
6. The method for denoising and detecting a magnetic anomaly signal of an ocean based on deep migration learning according to claim 1, wherein when the parameters of the denoising self-encoder network are adjusted by using the denoising magnetic field data, the mean square error of the denoising magnetic field data and the data in the simulated noise-containing ocean target magnetic anomaly data set is taken as a loss function, and the parameters of the next moment of the denoising self-encoder network are determined based on the loss function.
7. The deep migration learning-based ocean target magnetic anomaly signal denoising and detection method according to claim 1, wherein the fully connected classifier network comprises a fully connected layer, a ReLU activation function and a Softmax layer which are sequentially arranged, when the fully connected classifier network is adopted to classify the denoising magnetic field data, the denoising magnetic field data is input into the fully connected layer to obtain intermediate output, and the intermediate output passes through the ReLU activation function to obtain output data of the fully connected layer; and inputting the output data of the full connection layer into the Softmax layer to obtain an output probability value of each node in the Softmax layer, and taking a class corresponding to the node with the largest output probability value as a classification result of the denoising magnetic field data.
8. The deep migration learning-based ocean target magnetic anomaly signal denoising and detection method according to claim 7, wherein when parameters of the fully connected classifier network are adjusted by using the classification result, cross entropy between an output probability value of each node in the Softmax layer and a corresponding real class mark is taken as a loss function, and parameters at the next moment of the fully connected classifier network are determined based on the loss function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310106011.7A CN116304561A (en) | 2023-02-13 | 2023-02-13 | Ocean target magnetic anomaly signal denoising and detecting method based on deep migration learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310106011.7A CN116304561A (en) | 2023-02-13 | 2023-02-13 | Ocean target magnetic anomaly signal denoising and detecting method based on deep migration learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116304561A true CN116304561A (en) | 2023-06-23 |
Family
ID=86819570
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310106011.7A Pending CN116304561A (en) | 2023-02-13 | 2023-02-13 | Ocean target magnetic anomaly signal denoising and detecting method based on deep migration learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116304561A (en) |
-
2023
- 2023-02-13 CN CN202310106011.7A patent/CN116304561A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107481264B (en) | Video target tracking method with self-adaptive scale | |
| Zhu et al. | Intelligent logging lithological interpretation with convolution neural networks | |
| CN108710777B (en) | Diversified anomaly detection identification method based on multi-convolution self-coding neural network | |
| CN114564982B (en) | Automatic identification method for radar signal modulation type | |
| CN113050042B (en) | Radar signal modulation type identification method based on improved UNet3+ network | |
| CN108171119B (en) | SAR image change detection method based on residual error network | |
| CN105844279A (en) | Depth learning and SIFT feature-based SAR image change detection method | |
| CN109002848B (en) | Weak and small target detection method based on feature mapping neural network | |
| CN109522855A (en) | In conjunction with low resolution pedestrian detection method, system and the storage medium of ResNet and SENet | |
| CN114595732B (en) | Radar radiation source sorting method based on depth clustering | |
| CN113657491A (en) | Neural network design method for signal modulation type recognition | |
| CN109035300B (en) | Target tracking method based on depth feature and average peak correlation energy | |
| CN103824302B (en) | The SAR image change detection merged based on direction wave area image | |
| Li et al. | IncepTCN: A new deep temporal convolutional network combined with dictionary learning for strong cultural noise elimination of controlled-source electromagnetic data | |
| CN111612130B (en) | Frequency shift keying communication signal modulation mode identification method | |
| CN105374047A (en) | Improved bilateral filtering and clustered SAR based image change detection method | |
| CN112633147A (en) | Magnetic anomaly detection method based on multi-feature fusion and intelligent iterative optimization algorithm | |
| CN103971362B (en) | SAR image change-detection based on rectangular histogram and elite genetic algorithm for clustering | |
| CN110673208B (en) | First arrival picking method and system for high-dimensional feature constraint under machine learning framework | |
| CN110706208A (en) | Infrared dim target detection method based on tensor mean square minimum error | |
| CN105205807A (en) | Remote sensing image change detection method based on sparse automatic code machine | |
| Wang et al. | Marine target magnetic anomaly detection based on multi-task deep transfer learning | |
| CN108986083A (en) | SAR image change detection based on threshold optimization | |
| CN110503631B (en) | Remote sensing image change detection method | |
| CN116894223A (en) | Airborne underwater anomaly detection method based on self-adaptive cancellation and ResNet neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |