Method and device for suppressing noise in aviation shaft-frequency magnetic field detection

Abstract

The present invention provides a method and device for suppressing noise in aviation shaft-frequency magnetic field detection, comprising: step 1, obtaining a standard shaft-frequency magnetic field signal through forward modeling of an alternating electric dipole equivalent source; step 2, obtaining noise interference samples through experimental observation and mathematical simulation to construct a noisy shaft-frequency magnetic field signal; step 3, designing and training multiple groups of one-dimensional noise reduction autoencoder deep learning network models using the noisy shaft-frequency magnetic field signal as input and the standard shaft-frequency magnetic field signal as the expected output; step 4, performing principal component analysis on the multiple groups of noise reduction autoencoder network weight coefficients obtained through training to obtain the optimal network coefficient and build an optimized network model; step 5, testing and further optimizing the optimized network model using measured shaft-frequency magnetic field data to obtain the optimal network model; step 6, inputting the shaft-frequency magnetic field data to be optimized into the optimal network model to complete noise suppression in aviation shaft-frequency magnetic field detection. The technical solution of the present invention is applied to solve the technical problem of difficulty in suppressing environmental noise interference in existing aviation shaft-frequency magnetic field detection of underwater targets.

Description

Technical Field

[0001] The present invention relates to the field of geophysics and underwater target detection technology, and in particular to a method and device for suppressing noise in aviation axial frequency magnetic field detection. Background Art

[0002] Shaft-frequency electromagnetic fields are extremely low-frequency alternating electromagnetic fields generated by underwater targets in motion. Their signals exhibit distinct line spectrum characteristics in the frequency domain, with the fundamental frequency and harmonics significantly more powerful than other frequencies. Underwater target detection technology based on shaft-frequency electromagnetic fields is at the forefront of research both domestically and internationally. However, shaft-frequency electromagnetic field detection is subject to various environmental noise interferences and limitations in practical applications. Currently, research on shaft-frequency electric field signal processing techniques is relatively mature, with researchers proposing algorithms such as adaptive filtering, high-order spectral analysis, wavelet transforms, and empirical mode decomposition, which have been effectively applied. This is primarily because shaft-frequency electric fields detect underwater targets passing through them using electric field sensors installed on the seafloor. The detection platform is fixed in position, resulting in long signal observation times and relatively stable environmental noise interference. In the early days, shaft-frequency magnetic field detection also used inductive magnetic sensors fixed to the seafloor, employing similar signal processing methods for noise suppression. With the development of aerial detection platforms such as drones, aerial shaft-frequency magnetic field detection technology has become a research hotspot due to its advantages of high detection speed, high maneuverability, and ability to continuously track moving targets. Since the distance between the aerial platform and the underwater target is farther and the observation time is shorter, the signal is weaker and the existing signal processing method has limited noise suppression capabilities. Summary of the Invention

[0003] The present invention provides a method and device for suppressing noise in aviation shaft-frequency magnetic field detection, which can solve the technical problem of difficulty in suppressing environmental noise interference in existing aviation shaft-frequency magnetic field detection of underwater targets.

[0004] According to one aspect of the present invention, a method for suppressing noise in aviation shaft-frequency magnetic field detection based on deep learning of an improved denoising autoencoder is provided. The method for suppressing noise in aviation shaft-frequency magnetic field detection comprises: step one, obtaining a standard shaft-frequency magnetic field signal through forward modeling of an alternating electric dipole equivalent source; step two, obtaining noise interference samples through experimental observation and mathematical simulation, and constructing a noisy shaft-frequency magnetic field signal; step three, taking the noisy shaft-frequency magnetic field signal as input and the standard shaft-frequency magnetic field signal as the expected output, designing and training multiple groups of one-dimensional denoising autoencoder deep learning network models; step four, performing principal component analysis on the multiple groups of denoising autoencoder network weight coefficients obtained through training, obtaining the optimal network coefficient, and building an optimized network model; step five, testing and further optimizing the optimized network model through measured shaft-frequency magnetic field data, and obtaining the optimal network model; step six, inputting the shaft-frequency magnetic field data to be optimized into the optimal network model, and completing the noise suppression in aviation shaft-frequency magnetic field detection.

[0005] Furthermore, step one specifically includes: according to the type of underwater target and the number of propellers of the hull, the field source is equivalent to the accumulation of multiple electric dipole field sources; according to the flight altitude, speed and heading and sensor sampling rate parameters of the airborne magnetic detection platform, the coordinate position and data sampling time of each measuring point are determined; a layered medium model of the survey area is established, and the first layer of the layered medium model of the survey area is the air layer, the second layer is the seawater layer, and the third layer is the crustal rock layer; based on the accumulation of multiple electric dipole field sources, the coordinate position and data sampling time of each measuring point and the layered medium model of the survey area, the magnetic potential and magnetic field control equations are established according to the Maxwell equations, and the three components of the magnetic field and the total magnetic field frequency domain response of each measuring point in the air layer are solved by using the fast Hankel integral transform; the frequency domain response of each measuring point is converted into a complete time series by using the inverse Fourier transform; the time series of each measuring point is sampled along the survey line according to the passage of time, and the complete axis frequency magnetic field signal is obtained by combination.

[0006] Furthermore, step two specifically includes: obtaining different types of noise interference samples through experimental observation; simulating common types of noise interference through mathematical functions based on the time-frequency domain characteristics of the noise interference; generating multi-source mixed noise interference samples by randomly combining and superimposing experimental observation noise and mathematical simulation noise; and adding the noise interference samples to the standard shaft-frequency magnetic field signal generated in step one to construct a noisy shaft-frequency magnetic field signal.

[0007] Furthermore, the simulation of common types of noise interference through mathematical functions specifically includes: simulating low-frequency electromagnetic interference through the superposition of sinusoidal functions, multi-time functions, linear functions and exponential functions; simulating background random noise of different intensities through Gaussian signals of different amplitudes that appear continuously throughout the entire time period; simulating spike pulse interference through outliers whose partial amplitudes are greater than the set threshold of the normal signal and appear sporadically throughout the entire time period; simulating short-term burst strong interference through Gaussian signals whose overall amplitude is greater than the set threshold of the normal signal but only appears in a certain set time period.

[0008] Furthermore, step three specifically includes: simplifying the traditional denoising autoencoder model and converting the weight coefficients of each layer of the network into a one-dimensional form; taking the noisy axial frequency magnetic field signal as input and the standard axial frequency magnetic field signal as the expected output, randomly dividing the input and output sample data sets into N groups, and simultaneously training N groups of one-dimensional denoising autoencoder deep learning network models through parallel computing.

[0009] Furthermore, N groups of one-dimensional denoising autoencoder deep learning network models are trained simultaneously through parallel computing. Specifically, the following steps are used: the standard axial frequency magnetic field signal obtained by forward modeling is used as the pure signal sample x; multi-source mixed noise interference is added to the standard axial frequency magnetic field signal to form a disturbance sample x cn ; Encoding convolutional layer f in denoising autoencoder neural network e For the perturbation sample xcn Compression is performed to obtain low-dimensional implicit coding c; decoding convolution layer f d The low-dimensional implicit code c is decompressed and reconstructed into the output sample y again; the network coefficients of the encoding convolution layer and the decoding convolution layer are iteratively updated by minimizing the mean square error E(x,y) between the actual output sample y and the expected output x of the network model to obtain the final converged network weight coefficient.

[0010] Furthermore, the low-dimensional implicit code c is c=f e (x cn )=ReLU(Wx cn +b), the output sample y is y=f d (c) = ReLU(W'c+b'), where ReLU is the activation function, W is the weight coefficient matrix, b is the offset vector, and x cn is the perturbation sample, W' is the weight coefficient matrix corresponding to the decoding convolution layer, and b' is the offset vector corresponding to the decoding convolution layer.

[0011] Furthermore, step 4 specifically includes: performing singular value decomposition on the weight coefficients of the multiple sets of denoising autoencoder networks obtained through training; using the processed eigenvalue matrix S p , eigenvector U and eigenvector V reconstruct the matrix to obtain the common component θ of multiple groups of network weight coefficients p As the optimal weight coefficient; based on the optimal weight coefficient, build an optimized denoising autoencoder network model.

[0012] Furthermore, step five specifically includes: inputting the measured shaft-frequency magnetic field data into the optimized network model for noise reduction processing to obtain the shaft-frequency magnetic field signal after noise reduction; obtaining the line spectrum of the shaft-frequency magnetic field signal after noise reduction through spectrum analysis, determining the characteristic line spectrum corresponding to the fundamental frequency and the harmonics through conventional line spectrum detection methods, and calculating the characteristic line spectrum signal-to-noise ratio; if the characteristic line spectrum signal-to-noise ratio is improved, it means that the network model is available; if the characteristic line spectrum signal-to-noise ratio is not improved, it means that the network model is still unavailable, further expand the standard signal samples and noise interference samples according to steps one and two, and perform training and testing again according to steps three to five until the network model is available.

[0013] According to another aspect of the present invention, an aviation shaft-frequency magnetic field detection noise suppression system based on deep learning of an improved noise reduction autoencoder is provided. The aviation shaft-frequency magnetic field detection noise suppression system based on deep learning of an improved noise reduction autoencoder uses the aviation shaft-frequency magnetic field detection noise suppression method based on deep learning of an improved noise reduction autoencoder as described above to perform noise suppression.

[0014] According to another aspect of the present invention, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-mentioned method for noise suppression in aviation axial frequency magnetic field detection based on deep learning of an improved denoising autoencoder.

[0015] According to another aspect of the present invention, a computer-readable storage medium is provided, which stores a computer program, characterized in that when the computer program is executed by a processor, it implements the steps of the above-mentioned method for suppressing noise in aviation axial frequency magnetic field detection based on deep learning of an improved denoising autoencoder.

[0016] The technical solution of the present invention provides a method for suppressing noise in aviation shaft-frequency magnetic field detection based on deep learning of an improved denoising autoencoder. The method obtains a standard shaft-frequency magnetic field signal through forward modeling of an alternating electric dipole equivalent source; obtains noise interference samples through experimental observation and mathematical simulation to construct a noisy shaft-frequency magnetic field signal; uses the noisy shaft-frequency magnetic field signal as input and the standard shaft-frequency magnetic field signal as the expected output to design and train multiple sets of one-dimensional denoising autoencoder deep learning network models; performs principal component analysis on the weight coefficients of the multiple sets of denoising autoencoder networks obtained through training to obtain the optimal network coefficient and build the optimal network model; tests and further optimizes the optimal network model of the denoising autoencoder through measured shaft-frequency magnetic field data; finally, inputs the shaft-frequency magnetic field data to be optimized into the optimal network model to complete the noise suppression of aviation shaft-frequency magnetic field detection. The technical solution of the present invention can effectively suppress the noise of aviation shaft-frequency magnetic field detection and solve the technical problem of the difficulty in suppressing noise interference in the existing aviation shaft-frequency magnetic field detection of underwater targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings are included to provide a further understanding of the embodiments of the present invention, constitute a part of the specification, illustrate the embodiments of the present invention, and together with the description, explain the principles of the present invention. Obviously, the drawings described below are only some embodiments of the present invention, and those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0018] Figure 1 A flow chart of a method for suppressing noise in aviation shaft-frequency magnetic field detection based on deep learning of an improved noise reduction autoencoder is shown according to a specific embodiment of the present invention;

[0019] Figure 2 A flow chart of standard frequency magnetic field signal modeling based on an aviation detection platform is shown according to a specific embodiment of the present invention;

[0020] Figure 3A flow chart of generating a noisy shaft-frequency magnetic field signal sample according to a specific embodiment of the present invention is shown;

[0021] Figure 4 A schematic diagram of a deep learning neural network structure of a denoising autoencoder and a change in signal sample length is shown according to a specific embodiment of the present invention;

[0022] Figure 5 A flowchart of a network model test based on measured shaft frequency magnetic field data is shown according to a specific embodiment of the present invention. DETAILED DESCRIPTION

[0023] It should be noted that, in the absence of conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is in no way intended to limit the present invention and its application or use. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0024] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0025] Unless otherwise specifically stated, the relative arrangement of the parts and steps, the numerical expressions and the numerical values set forth in these embodiments do not limit the scope of the present invention. At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship. The techniques, methods and equipment known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, the techniques, methods and equipment should be considered as part of the authorization specification. In all examples shown and discussed here, any specific values should be interpreted as being merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values. It should be noted that similar numbers and letters represent similar items in the following figures, and therefore, once an item is defined in one figure, it does not need to be further discussed in subsequent figures.

[0026] like Figures 1 to 5As shown, according to a specific embodiment of the present invention, a method for suppressing noise in aviation shaft-frequency magnetic field detection based on deep learning of an improved denoising autoencoder is provided, and the method for suppressing noise in aviation shaft-frequency magnetic field detection includes: step one, obtaining a standard shaft-frequency magnetic field signal through forward modeling of an alternating electric dipole equivalent source; step two, obtaining noise interference samples through experimental observation and mathematical simulation, and constructing a noisy shaft-frequency magnetic field signal; step three, taking the noisy shaft-frequency magnetic field signal as input and the standard shaft-frequency magnetic field signal as the expected output, designing and training multiple groups of one-dimensional denoising autoencoder deep learning network models; step four, performing principal component analysis on the multiple groups of denoising autoencoder network weight coefficients obtained through training, obtaining the optimal network coefficient, and building an optimized network model; step five, testing and further optimizing the optimized network model through measured shaft-frequency magnetic field data to obtain the optimal network model; step six, inputting the shaft-frequency magnetic field data to be optimized into the optimal network model to complete the noise suppression in aviation shaft-frequency magnetic field detection.

[0027] By applying this configuration, a method for suppressing noise in aviation shaft-frequency magnetic field detection based on deep learning of an improved denoising autoencoder is provided. The method obtains a standard shaft-frequency magnetic field signal through forward modeling of an alternating electric dipole equivalent source; obtains noise interference samples through experimental observation and mathematical simulation to construct a noisy shaft-frequency magnetic field signal; takes the noisy shaft-frequency magnetic field signal as input and the standard shaft-frequency magnetic field signal as the expected output, designs and trains multiple sets of one-dimensional denoising autoencoder deep learning network models; performs principal component analysis on the multiple sets of denoising autoencoder network weight coefficients obtained through training to obtain the optimal network coefficient and build the optimal network model; tests and further optimizes the optimal network model of the denoising autoencoder through measured shaft-frequency magnetic field data; finally, inputs the shaft-frequency magnetic field data to be optimized into the optimal network model to complete the noise suppression of aviation shaft-frequency magnetic field detection. By applying the technical solution of the present invention, effective suppression of noise in aviation shaft-frequency magnetic field detection can be achieved, solving the technical problem of difficulty in suppressing noise interference in existing underwater target aviation shaft-frequency magnetic field detection.

[0028] Specifically, in the present invention, in order to achieve noise suppression in aviation shaft-frequency magnetic field detection, it is first necessary to obtain a standard shaft-frequency magnetic field signal through alternating electric dipole equivalent source forward modeling. In the present invention, step one specifically includes: according to the type of underwater target and the number of propellers of the hull, the field source is equivalent to the accumulation of multiple electric dipole field sources; according to the flight altitude, speed and heading of the airborne magnetic detection platform and the sensor sampling rate parameters, the coordinate position and data sampling time of each measuring point are determined; a layered medium model of the survey area is established, and the first layer of the layered medium model of the survey area is the air layer, the second layer is the seawater layer, and the third layer is the crustal rock layer; based on the accumulation of multiple electric dipole field sources, the coordinate position and data sampling time of each measuring point and the layered medium model of the survey area, the magnetic potential and magnetic field control equations are established according to the Maxwell equations, and the three components of the magnetic field and the total magnetic field frequency domain response of each measuring point in the air layer are solved by using the fast Hankel integral transform; the frequency domain response of each measuring point is converted into a complete time series by using the inverse Fourier transform; the time series of each measuring point is sampled along the survey line according to the passage of time, and the complete axial frequency magnetic field signal is obtained by combination.

[0029] As a specific embodiment of the present invention, the underwater target shaft frequency magnetic field is generated by the corrosion / anti-corrosion current modulated by the rotation of the main shaft such as the propeller, and has a certain time-frequency domain characteristic. The field source can be equivalent to a horizontal alternating electric dipole. Based on the equivalent source, forward modeling can obtain a standard electromagnetic signal without noise interference. The modeling process is as follows: Figure 2 As shown in the figure, the field source is first calculated based on the type of underwater target and the number of propellers on the hull, equivalent to the accumulation of multiple electric dipoles. The coordinates of each measuring point and the data sampling time are then determined based on parameters such as the flight altitude, speed and heading, and sensor sampling rate of the aerial magnetic exploration platform (UAV) used in the actual detection mission. A layered medium model is then established for the survey area, with the first layer being air, the second being seawater, and the third being crustal rock. The magnetic potential and magnetic field governing equations are established based on Maxwell's equations, and the fast Hankel integral transform is used to solve them to obtain the three components of the magnetic field and the total magnetic field frequency response at each measuring point in the air layer. The inverse Fourier transform is used to convert the frequency domain response of each measuring point into a complete time series. Finally, the time series of each measuring point is sampled over time along the survey line and combined to obtain a complete shaft-frequency magnetic field signal. By varying the target field source intensity and the parameters of the layered medium model (such as the thickness and conductivity of each layer), shaft-frequency magnetic field signals can be obtained under different conditions. Parallel computing can rapidly generate tens of thousands of signal sets, thus establishing a standard signal sample library.

[0030] After obtaining the standard shaft-frequency magnetic field signal, noise interference samples can be obtained through experimental observation and mathematical simulation to construct a noisy shaft-frequency magnetic field signal. In the present invention, step two specifically includes: obtaining different types of noise interference samples through experimental observation; simulating common types of noise interference through mathematical functions according to the time-frequency domain characteristics of noise interference; generating multi-source mixed noise interference samples by randomly combining and superimposing experimental observation noise and mathematical simulation noise; adding the noise interference samples to the standard shaft-frequency magnetic field signal generated in step one to construct a noisy shaft-frequency magnetic field signal. Among them, simulating common types of noise interference through mathematical functions specifically includes: simulating low-frequency electromagnetic interference through the superposition of sinusoidal functions, multi-time functions, linear functions and exponential functions; simulating background random noise of different intensities through Gaussian signals of different amplitudes that appear continuously throughout the entire period; simulating spike pulse interference through outliers whose partial amplitudes are greater than the normal signal set threshold and appear sporadically throughout the entire period; simulating short-term burst strong interference through Gaussian signals whose overall amplitude is greater than the normal signal set threshold but only appears in a certain set time period.

[0031] As a specific embodiment of the present invention, common noise interference sources in airborne shaft-frequency magnetic field detection include high-altitude interference sources such as magnetic storms and lightning, marine interference sources such as waves and storms, electromagnetic interference sources from water-based facilities such as ships and drilling rigs, and persistent Gaussian-like random noise in the marine environment. These various types of noise interference are complex and variable. First, different types of noise interference samples need to be obtained through experimental observation. The parameters used for noise observation, such as the aircraft platform, flight altitude, and data sampling rate, are the same as those used for airborne shaft-frequency magnetic field detection of underwater targets. Based on the experimental observations, mathematical functions are used to simulate common types of noise interference based on the time-frequency domain characteristics of the noise interference. Low-frequency electromagnetic interference is simulated by superimposing sinusoidal, multi-order, linear, and exponential functions. Background random noise of varying intensities is simulated by continuously appearing Gaussian signals of varying amplitudes throughout the entire time period. Spike interference is simulated by sporadic outliers with amplitudes significantly greater than the normal signal. Short-duration, bursty strong interference is simulated by Gaussian signals with amplitudes significantly greater than the normal signal but occurring only during a specific time period. Finally, a multi-source mixed noise interference sample is generated by randomly combining experimental observation noise and mathematical simulation noise. Then the noise interference sample is added to the standard shaft frequency magnetic field signal generated in step 1 to construct a noisy shaft frequency magnetic field signal sample, such as Figure 3 shown.

[0032] Furthermore, after constructing the noisy shaft-frequency magnetic field signal, the noisy shaft-frequency magnetic field signal can be used as input and the standard shaft-frequency magnetic field signal as the expected output to design and train multiple groups of one-dimensional noise reduction autoencoder deep learning network models. In the present invention, step three specifically includes: simplifying the traditional noise reduction autoencoder model and converting the weight coefficients of each layer of the network into a one-dimensional form; using the noisy shaft-frequency magnetic field signal as input and the standard shaft-frequency magnetic field signal as the expected output, randomly dividing the input and output sample data sets into N groups, and simultaneously training N groups of one-dimensional noise reduction autoencoder deep learning network models through parallel computing. Among them, simultaneously training N groups of one-dimensional noise reduction autoencoder deep learning network models through parallel computing specifically includes: using the standard shaft-frequency magnetic field signal obtained by forward modeling as a pure signal sample x; adding multi-source mixed noise interference to the standard shaft-frequency magnetic field signal to form a disturbance sample x cn ; Encoding convolutional layer f in denoising autoencoder neural network e For the perturbation sample x cn Compression is performed to obtain low-dimensional implicit coding c; decoding convolution layer f d The low-dimensional implicit code c is decompressed and reconstructed into the output sample y; the network coefficients of the encoding convolution layer and the decoding convolution layer are iteratively updated to obtain the final converged network weight coefficient by minimizing the mean square error E(x,y) between the actual output sample y and the expected output x of the network model. The low-dimensional implicit code c is c = f e (x cn )=ReLU(Wx cn +b), the output sample y is y=f d (c) = ReLU(W'c+b'), where ReLU is the activation function, W is the weight coefficient matrix, b is the offset vector, and x cn is the perturbation sample, W' is the weight coefficient matrix corresponding to the decoding convolution layer, and b' is the offset vector corresponding to the decoding convolution layer.

[0033] As a specific embodiment of the present invention, the deep learning method has higher accuracy and efficiency in noise suppression, which helps to improve the accuracy of target signal extraction. This provides a new idea for suppressing environmental noise interference in underwater target aviation shaft-frequency magnetic field detection. Based on this, after obtaining the standard shaft-frequency magnetic field signal and the noisy shaft-frequency magnetic field signal, a denoising autoencoder deep learning neural network model is established with the noisy shaft-frequency magnetic field signal as the input sample and the standard shaft-frequency magnetic field signal as the expected output sample for noise suppression. Traditional denoising autoencoders are mainly used for image data processing. The weight coefficients of each network layer are in two-dimensional form, the network scale is large, and the computational efficiency is low. In addition, traditional methods usually only train one set of deep learning network models, which are prone to problems such as poor model stability and easy to fall into overfitting. This patent takes into account that the original signals observed in aviation axial-frequency magnetic field detection missions are one-dimensional data. To improve efficiency, this method first simplifies the traditional denoising autoencoder model and converts the weight coefficients of each layer of the network into a one-dimensional form. Then, the input and output sample data sets are randomly divided into N groups, and N groups of one-dimensional denoising autoencoder models are trained simultaneously through parallel computing. Furthermore, principal component analysis is used to extract the common components of the weight coefficients of the N groups of neural network models as the optimal coefficients to build the optimal network model. Compared with the traditional denoising autoencoder method, this improved method can improve the model training efficiency while enhancing the model's generalization ability.

[0034] The one-dimensional denoising autoencoder network structure is as follows Figure 4 As shown, the network consists of an encoding convolutional layer and a decoding convolutional layer. The encoding convolutional layer, consisting of a convolutional layer and a pooling layer, reduces the dimensionality of the input data by downsampling, compressing noisy input data into a low-dimensional implicit code. The decoding layer, consisting of a transposed convolutional layer, reconstructs the low-dimensional implicit code into standard output data free of noise by upsampling. Assuming the original signal length is L, the denoising autoencoder performs three compressions and three reconstructions on the signal, preserving the main components of the original signal while suppressing random noise interference. Both the encoding and decoding convolutional layers consist of network coefficients, which are the model parameters to be optimized. After establishing the denoising autoencoder network model, these model parameters are randomly initialized. Then, the network coefficients are optimized by minimizing the mean squared error between the model output and the expected output.

[0035] The calculation and training process of each group of one-dimensional denoising autoencoder network models is as follows. After obtaining the standard axial frequency magnetic field signal through forward modeling, it is used as the pure signal sample x, and then multi-source mixed noise interference is added to it to form a perturbation sample x cn The encoding convolutional layer f in the denoising autoencoder neural network e The perturbation samples are compressed to obtain the low-dimensional implicit code c, which is calculated as follows:

[0036] c=f e (xcn )=ReLU(Wx cn +b), (1)

[0037] Among them, ReLU is the activation function, W is the weight coefficient matrix, b is the offset vector, W and b are collectively referred to as network coefficients. Finally, the decoding convolution layer f d The low-dimensional implicit code c is decompressed and reconstructed into the output sample y again. The calculation formula is:

[0038] y=f d (c) = ReLU(W'c+b'), (2)

[0039] Where W' and b' are the weight coefficient matrix and offset vector corresponding to the decoding convolution layer. The network coefficients of the encoding convolution layer and the decoding convolution layer are then iteratively updated by minimizing the mean square error E(x,y) between the actual output y of the network model and the expected output (i.e., the clean signal sample) x:

[0040] E(x,y)=||xy|| 2 =||x-ReLU{W'[ReLU(Wx cn +b)]+b'}|| 2 , (3)

[0041] During the iterative optimization process, the adaptive momentum estimation (Adam) algorithm is used to iteratively calculate the network coefficient θ = [W, b, W', b'], and the formula is as follows:

[0042]

[0043] Among them, k is the iteration sequence number, α is the learning rate, which controls the update speed of the network coefficient; ε is a very small variable to avoid the denominator being 0; m k+1 and v k+1 are two momentum coefficients, and the formula is as follows:

[0044]

[0045] in, is the gradient of the objective function E with respect to the network coefficient matrix θ, β1 and β2 are two non-negative parameters less than 1, usually 0.9. In the above formula, θ k , m k , v k The initial value of is randomly set, and the final converged network weight coefficient is obtained through iteration.

[0046] Furthermore, after designing and training multiple sets of one-dimensional denoising autoencoder deep learning network models, principal component analysis can be performed on the multiple sets of denoising autoencoder network weight coefficients obtained through training to obtain the optimal network coefficients and build an optimized network model. In the present invention, step 4 specifically includes: performing singular value decomposition on the multiple sets of denoising autoencoder network weight coefficients obtained through training; using the processed eigenvalue matrix S p , eigenvector U and eigenvector V reconstruct the matrix to obtain the common component θ of multiple groups of network weight coefficients p As the optimal weight coefficient; based on the optimal weight coefficient, build an optimized denoising autoencoder network model.

[0047] As a specific embodiment of the present invention, since the sample data and initial values of weight coefficients used in training different network models are not exactly the same, the training results obtained in each group are the same in principle, but there are differences in details. This patent uses principal component analysis to extract the main components of the network coefficients and improve the generalization ability of the model. There are N groups of network coefficients obtained through training, and each group has M coefficients. group It is a matrix composed of network coefficients of each group, with a dimension of M×N. First, it is subjected to singular value decomposition. The calculation formula is as follows:

[0048]

[0049] Among them, SVD is the singular value decomposition operator, U and V are M×M and N×N matrices respectively, representing the covariance matrix and The eigenvector of , S is an M×N matrix, except for the main diagonal, all other elements are 0, and each element on the main diagonal is a singular value. Through singular value decomposition, the network coefficient matrix θ group Decompose into different subspace vectors. The first singular value represents the common component with the strongest correlation in the data matrix, and the other singular values mainly represent the components with weaker correlation in the data matrix. At this time, only the first largest singular value is retained, and the other singular values are set to 0.

[0050] Then use the processed eigenvalue matrix S p By reconstructing the matrix with the eigenvectors U and V, we can obtain the common components θ of multiple groups of network weight coefficients p , which is the optimal weight coefficient θ p , the calculation formula is as follows:

[0051] θ p =U*S p *V T (7)

[0052] Based on the optimal weight coefficient θ p , build a denoising autoencoder network, which is the optimized network model.

[0053] After constructing the optimized network model, in order to further improve the accuracy of the noise suppression of the network model, the optimized network model can be tested and further optimized by measuring the shaft frequency magnetic field data to obtain the optimal network model. In the present invention, step five specifically includes: inputting the measured shaft frequency magnetic field data into the optimized network model for noise reduction processing to obtain the shaft frequency magnetic field signal after noise reduction; obtaining the line spectrum of the shaft frequency magnetic field signal after noise reduction through spectrum analysis, determining the characteristic line spectrum corresponding to the fundamental frequency and the harmonic frequency through the conventional line spectrum detection method, and calculating the signal-to-noise ratio of the characteristic line spectrum; if the characteristic line spectrum signal-to-noise ratio is improved, it means that the network model is available, if the characteristic line spectrum signal-to-noise ratio is not improved, it means that the network model is still unavailable, further expand the standard signal samples and noise interference samples according to steps one and two, and perform training and testing again according to steps three to five until the network model is available.

[0054] As a specific embodiment of the present invention, after the denoising autoencoder neural network model is trained, the measured noisy axial frequency magnetic field signal is used as the signal to be denoised for testing, such as Figure 5 As shown. First, the signal to be denoised is input into the optimized network model obtained in step 4 for denoising to obtain the denoised signal. Then, the line spectrum of the axial frequency magnetic field signal after denoising is obtained through spectrum analysis. The characteristic line spectrum corresponding to the fundamental frequency and the harmonics is determined by the conventional line spectrum detection method. The signal-to-noise ratio of the characteristic line spectrum (the ratio of the fundamental frequency and harmonics line spectrum energy to the sum of all line spectrum energies) is calculated. If the characteristic line spectrum signal-to-noise ratio is improved, it means that the network model is applicable. If the characteristic line spectrum signal-to-noise ratio is not improved, it means that the network model is still unavailable and it is necessary to further expand the standard signal samples and noise interference samples according to steps 1 and 2, and perform training and testing again according to steps 3 to 5.

[0055] According to another aspect of the present invention, an aviation shaft-frequency magnetic field detection noise suppression system based on improved noise reduction autoencoder deep learning is provided. The aviation shaft-frequency magnetic field detection noise suppression system based on improved noise reduction autoencoder deep learning uses the aviation shaft-frequency magnetic field detection noise suppression method based on improved noise reduction autoencoder deep learning as described above to perform noise suppression.

[0056] By applying this configuration, a noise suppression system for aviation shaft-frequency magnetic field detection based on deep learning of an improved noise reduction autoencoder is provided. The system obtains a standard shaft-frequency magnetic field signal through forward modeling of an alternating electric dipole equivalent source; obtains noise interference samples through experimental observation and mathematical simulation to construct a noisy shaft-frequency magnetic field signal; takes the noisy shaft-frequency magnetic field signal as input and the standard shaft-frequency magnetic field signal as the expected output, designs and trains multiple sets of one-dimensional noise reduction autoencoder deep learning network models; performs principal component analysis on the multiple sets of noise reduction autoencoder network weight coefficients obtained by training, obtains the optimal network coefficient, and builds the optimal network model; tests and further optimizes the optimal network model of the noise reduction autoencoder through measured shaft-frequency magnetic field data; finally, inputs the shaft-frequency magnetic field data to be optimized into the optimal network model to complete the noise suppression of aviation shaft-frequency magnetic field detection. By applying the technical solution of the present invention, effective suppression of aviation shaft-frequency magnetic field detection noise can be achieved, solving the technical problem of difficulty in suppressing noise interference in existing underwater target aviation shaft-frequency magnetic field detection.

[0057] According to another aspect of the present invention, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-mentioned method for noise suppression in aviation axial frequency magnetic field detection based on deep learning of an improved denoising autoencoder.

[0058] According to another aspect of the present invention, a computer-readable storage medium is provided, which stores a computer program, characterized in that when the computer program is executed by a processor, it implements the steps of a method for suppressing noise in aviation axial frequency magnetic field detection based on deep learning of an improved denoising autoencoder as shown above.

[0059] In summary, the present invention provides a method for suppressing noise in aviation shaft-frequency magnetic field detection based on deep learning of an improved denoising autoencoder. The method obtains a standard shaft-frequency magnetic field signal through forward modeling of an alternating electric dipole equivalent source; obtains noise interference samples through experimental observation and mathematical simulation, and constructs a noisy shaft-frequency magnetic field signal; uses the noisy shaft-frequency magnetic field signal as input and the standard shaft-frequency magnetic field signal as the expected output to design and train multiple groups of one-dimensional denoising autoencoder deep learning network models; performs principal component analysis on the multiple groups of denoising autoencoder network weight coefficients obtained through training to obtain the optimal network coefficient and build the optimal network model; tests and further optimizes the optimal network model of the denoising autoencoder through measured shaft-frequency magnetic field data; finally, inputs the shaft-frequency magnetic field data to be optimized into the optimal network model to complete the noise suppression of aviation shaft-frequency magnetic field detection. By applying the technical solution of the present invention, effective suppression of aviation shaft-frequency magnetic field detection noise can be achieved, solving the technical problem of difficulty in suppressing noise interference in existing underwater target aviation shaft-frequency magnetic field detection.

[0060] For ease of description, spatially relative terms such as "above", "above", "on the upper surface of", "above", etc. may be used herein to describe the spatial positional relationship of a device or feature to other devices or features as shown in the figures. It should be understood that spatially relative terms are intended to include different orientations of the device in use or operation in addition to the orientation described in the figures. For example, if the device in the drawings is inverted, the device described as "above other devices or structures" or "above other devices or structures" will be positioned as "below other devices or structures" or "below other devices or structures". Thus, the exemplary term "above" can include both "above" and "below". The device can also be positioned in other different ways (rotated 90 degrees or in other orientations), and the spatially relative descriptions used here are interpreted accordingly.

[0061] In addition, it should be noted that the use of terms such as "first" and "second" to limit components is only for the convenience of distinguishing the corresponding components. Unless otherwise stated, the above terms have no special meaning and therefore cannot be understood as limiting the scope of protection of the present invention.

[0062] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A noise suppression method for aviation shaft frequency magnetic field detection based on deep learning of an improved noise reduction autoencoder, characterized in that: The method for suppressing aviation shaft-frequency magnetic field detection noise includes: Step 1: Obtain the standard shaft frequency magnetic field signal through the forward modeling of the alternating electric dipole equivalent source; Step 2: Obtain noise interference samples through experimental observation and mathematical simulation to construct a noisy shaft frequency magnetic field signal; Step 3: Using the noisy shaft-frequency magnetic field signal as input and the standard shaft-frequency magnetic field signal as expected output, design and train multiple groups of one-dimensional noise reduction autoencoder deep learning network models; Step 4: Perform principal component analysis on the weight coefficients of the multiple sets of denoising autoencoder networks obtained through training to obtain the optimal network coefficients and build an optimized network model. Step 5: Testing and further optimizing the optimized network model by measuring shaft frequency magnetic field data to obtain the optimal network model; Step 6: Inputting the shaft frequency magnetic field data to be optimized into the optimal network model to complete the noise suppression of aviation shaft frequency magnetic field detection. Step 3 specifically includes: The traditional denoising autoencoder model is simplified, and the weight coefficients of each layer of the network are converted into a one-dimensional form. The noisy shaft-frequency magnetic field signal is used as input and the standard shaft-frequency magnetic field signal is used as the expected output. The input and output sample data sets are randomly divided into N groups. N groups of one-dimensional denoising autoencoder deep learning network models are trained simultaneously through parallel computing. The specific training of N groups of one-dimensional denoising autoencoder deep learning network models through parallel computing includes: The standard axial frequency magnetic field signal obtained by forward modeling is used as the pure signal sample x; Add multi-source mixed noise interference to the standard shaft frequency magnetic field signal to form a disturbance sample x cn ; The encoding convolutional layer f in the denoising autoencoder neural network e For the perturbation sample x cn Compression is performed to obtain a low-dimensional implicit code c; Decoding convolutional layer f d Decompress the low-dimensional implicit code c again and reconstruct it into the output sample y; The network coefficients of the encoding convolution layer and the decoding convolution layer are iteratively updated by minimizing the mean square error E(x, y) between the actual output sample y and the expected output x of the network model to obtain the final converged network weight coefficients.

2. The method for suppressing noise in aviation shaft-frequency magnetic field detection based on deep learning of an improved noise reduction autoencoder according to claim 1 is characterized in that: The step 1 specifically includes: According to the type of underwater target and the number of propellers on the hull, the field source is equivalent to the accumulation of multiple electric dipole field sources; Determine the coordinate position and data sampling time of each measuring point based on the flight altitude, speed and heading of the aeromagnetic detection platform and the sensor sampling rate parameters; Establishing a layered medium model of the survey area, wherein the first layer of the layered medium model of the survey area is an air layer, the second layer is a seawater layer, and the third layer is a crustal rock layer; Based on the accumulation of the multi-electric dipole field sources, the coordinate positions and data sampling time of each measuring point, and the layered medium model of the measuring area, the magnetic potential and magnetic field control equations are established according to the Maxwell equations, and the three components of the magnetic field at each measuring point in the air layer and the frequency domain response of the total magnetic field are obtained by using the fast Hankel integral transform. The frequency domain response of each measuring point is converted into a complete time series using inverse Fourier transform; The time series of each measuring point is sampled over time along the measuring line and combined to obtain a complete shaft frequency magnetic field signal.

3. The method for suppressing noise in aviation shaft-frequency magnetic field detection based on deep learning of an improved noise reduction autoencoder according to claim 1 is characterized in that: The second step specifically includes: Obtain different types of noise interference samples through experimental observation; According to the time-frequency domain characteristics of noise interference, common types of noise interference are simulated through mathematical functions; The multi-source mixed noise interference samples are generated by randomly combining experimental observation noise and mathematical simulation noise; The noise interference sample is added to the standard shaft-frequency magnetic field signal generated in step 1 to construct a noisy shaft-frequency magnetic field signal.

4. The method for suppressing noise in aviation shaft-frequency magnetic field detection based on deep learning of an improved noise reduction autoencoder according to claim 3 is characterized in that: The simulation of common types of noise interference through mathematical functions specifically includes: simulating low-frequency electromagnetic interference through the superposition of sinusoidal functions, multiple functions, linear functions and exponential functions; simulating background random noise of different intensities through Gaussian signals of different amplitudes that appear continuously throughout the entire time period; simulating spike pulse interference through outliers whose partial amplitudes are greater than the set threshold of normal signals and appear sporadically throughout the entire time period; simulating short-term burst strong interference through Gaussian signals whose overall amplitude is greater than the set threshold of normal signals but only appears in a certain set time period.

5. The method for suppressing noise in aviation shaft-frequency magnetic field detection based on deep learning of an improved noise reduction autoencoder according to claim 1, characterized in that: The low-dimensional implicit code c is c=f e (x cn )=ReLU(Wx cn +b), the output sample y is y=f d (c) = ReLU(W'c+b'), where ReLU is the activation function, W is the weight coefficient matrix, b is the offset vector, and x cn is the perturbation sample, W' is the weight coefficient matrix corresponding to the decoding convolution layer, and b' is the offset vector corresponding to the decoding convolution layer.

6. The method for suppressing noise in aviation shaft-frequency magnetic field detection based on deep learning of an improved noise reduction autoencoder according to any one of claims 1 to 5, characterized in that: The step 4 specifically includes: Perform singular value decomposition on multiple sets of denoising autoencoder network weight coefficients obtained through training; Use the processed eigenvalue matrix S p , eigenvector U and eigenvector V reconstruct the matrix to obtain the common component θ of multiple groups of network weight coefficients p As the optimal weight coefficient; Based on the optimal weight coefficients, an optimized denoising autoencoder network model is built.

7. The method for suppressing noise in aviation shaft-frequency magnetic field detection based on deep learning of an improved noise reduction autoencoder according to claim 6, characterized in that: The step five specifically includes: Inputting the measured shaft frequency magnetic field data into the optimized network model for noise reduction processing to obtain a noise-reduced shaft frequency magnetic field signal; The line spectrum of the axial frequency magnetic field signal after noise reduction is obtained by spectrum analysis, the characteristic line spectrum corresponding to the fundamental frequency and the harmonic frequency is determined by conventional line spectrum detection method, and the signal-to-noise ratio of the characteristic line spectrum is calculated; If the characteristic line spectrum signal-to-noise ratio is improved, it means that the network model is available. If the characteristic line spectrum signal-to-noise ratio is not improved, it means that the network model is still unavailable. According to step one and step two, further expand the standard signal samples and noise interference samples, and perform training and testing again according to steps three to five until the network model is available.

8. An aviation shaft frequency magnetic field detection noise suppression system based on deep learning of an improved noise reduction autoencoder, characterized in that: The aviation shaft-frequency magnetic field detection noise suppression system based on improved noise reduction autoencoder deep learning uses the aviation shaft-frequency magnetic field detection noise suppression method based on improved noise reduction autoencoder deep learning as described in any one of claims 1 to 7 to perform noise suppression.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: The processor executes the computer program to implement the steps of the method for suppressing noise in aviation axial frequency magnetic field detection based on deep learning of an improved noise reduction autoencoder according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of a method for suppressing noise in aviation axial frequency magnetic field detection based on deep learning of an improved noise reduction autoencoder as described in any one of claims 1 to 7 are implemented.

Images