An airborne underwater anomaly detection method based on adaptive cancellation and ResNet neural network

Abstract

The present invention relates to an airborne underwater anomaly detection method based on adaptive cancellation and ResNet neural network, which comprises the following steps: First, configure the distance between the airborne magnetoelectric antenna and the magnetic field sensor to be fixed; Second, the airborne magnetoelectric antenna emits low-frequency electromagnetic wave signals underwater, and the airborne fluxgate meter receives the three-axis magnetic field intensity, including the primary field of the antenna and the secondary field of the underwater metal anomaly eddy current effect; Then, use the adaptive cancellation technology to separate the secondary field in the received signal to obtain the target useful signal; Finally, use the sliding window method to complete the segmentation processing of the secondary field time series signal according to the antenna transmission signal period, and send it into the pre-trained ResNet neural network to complete the detection and identification of the presence or absence of anomaly signals. The present invention is convenient for carrying out large-area underwater anomaly metal detection. At the same time, the pre-trained model has a fast analysis and processing speed and a high accuracy rate, and is a convenient and reliable underwater anomaly detection method.

Description

Technical Field

[0001] The present invention relates to an airborne underwater anomaly detection method based on adaptive cancellation and ResNet neural network, which is mainly applied to accurately detect whether there are metal anomalies underwater and belongs to the technical field of magnetic detection and positioning. Background Art

[0002] Underwater metal anomaly detection technology is widely used in underwater mineral exploration, traffic flow monitoring, weapon detection, underwater search and rescue and other fields, and has extremely high application value. In the underwater detection application scenario, the signals received by the magnetic field sensor include the primary field signal generated by the magnetoelectric antenna emission source, the secondary field signal induced by the underwater metal anomaly object, ocean noise, clutter, etc. The strength of the secondary field signal is the main index to test the presence or absence of underwater metal anomaly objects. Therefore, the secondary field signal can be regarded as a useful signal, and the primary field signal, ocean noise and clutter can be regarded as noise signals. In addition, the primary field signal and the secondary field signal are of the same frequency and the intensity of the primary field signal is much greater than that of the secondary field signal. It is difficult for the traditional frequency domain filtering method to filter out the primary field signal. At the same time, the analytical solution form of the secondary field signal is complex and a relatively deep neural network is required. The traditional neural network method is prone to problems of gradient disappearance and gradient explosion, and the classification effect is poor. The present invention filters out the primary field signal of the magnetoelectric antenna based on the adaptive cancellation algorithm, extracts a relatively pure secondary field signal, and sends it into the ResNet neural network for high-precision classification and discrimination of the signal, so as to realize the detection of the presence or absence of underwater anomalies. Summary of the Invention

[0003] The technical problem solved by the present invention: Use an airborne magnetoelectric antenna to emit an electromagnetic wave primary field to detect the secondary field generated by underwater metal anomalies, and realize the identification of underwater non-ferromagnetic targets; Use the adaptive cancellation method to remove the geomagnetic field and the primary field, and high-precision feature extraction of the target electromagnetic signal can be realized; Use the sliding window method to realize the segmentation and truncation of data, and the signal time series information is retained; Use the ResNet neural network for classification, the number of parameters decreases, the complexity decreases, at the same time the network depth is deeper and there will be no gradient disappearance, and the accuracy of signal classification is high.

[0004] The technical solution of the present invention: An airborne underwater anomaly detection method based on adaptive cancellation and ResNet neural network, and its implementation steps are as follows:

[0005] First step, configure and train an adaptive cancellation system. The adaptive cancellation system consists of a transmitting end, a receiving end, and a data end. The transmitting end is composed of a signal generator, a voltage amplifier, and a magnetoelectric antenna. The excitation signal generated by the signal generator is power-amplified by a high-voltage amplifier and then applied as a driving voltage to the magnetoelectric antenna. Due to the magnetoelectric coupling effect, the magnetoelectric antenna generates an electromagnetic field (primary field). Since the excitation signal generated by the signal generator has the same frequency and a fixed phase difference as the primary field signal generated by the magnetoelectric antenna, and the excitation signal can be directly measured, the excitation signal generated by the signal source can be used as a reference signal to estimate the primary field signal.

[0006] The receiving end is composed of a magnetic field sensor and a preamplifier. The magnetic field sensor receives the electromagnetic signals in the surrounding environment, and the preamplifier performs band-pass filtering on the signals received by the magnetic field sensor to filter out low-frequency ocean noise and high-frequency noise.

[0007] The data end is composed of a data acquisition card and a PC. The data acquisition card acquires the excitation signal of the signal generator and the pre-filtered signal of the preamplifier, and outputs these two sets of data to the PC for adaptive cancellation.

[0008] The electromagnetic field of an underwater metal target at a relatively far distance from the distance measurement system has the following complex expression:

[0009]

[0010]

[0011] H φ =0

[0012] where, H r , H θ , H φ respectively represent the magnetic field intensities in the three coordinate axis directions of the spherical coordinate system of the magnetoelectric antenna, j represents the imaginary unit, r represents the distance from the underwater metal anomaly to the sensor, I0 represents the magnitude of the eddy current generated in the metal, θ represents the angle in the spherical coordinate system, k represents the complex wave number. Using an adaptive filter to process this sinusoidal magnetic field, its loss function can be expressed as:

[0013] e 2 (i)=(d(i)-W T (i)X(i)) 2

[0014] where, W is the parameter of the adaptive filter, d is the actual value or expected value, X is the reference signal, and i represents the i-th iteration.

[0015] In addition, the iteration of W follows the following formula:

[0016] W(i + 1) = W(i) + μe(i)X(i)

[0017] where μ is the step size of the adaptive filter.

[0018] In the training stage, the parameters of the adaptive filter can be initially set to zero, and then the parameters of the adaptive filter are iterated according to the loss function and the step size. When the change amount of the parameter is less than a given threshold, it is considered that the parameters of the adaptive filter converge. After storing the parameters, they can be directly used in the application stage for adaptive filtering.

[0019] In the second step, the secondary field is extracted using the parameters of the adaptive filter. The output signal satisfies the following form:

[0020] s = s0 + n0 - n

[0021] where s is the output signal of the adaptive filter, s0 is the target secondary field signal, n0 is the primary field and noise collected by the magnetic field sensor in the cancellation stage, and n is the filtering result of the primary field and noise signals in the learning stage by the adaptive filter. Since the learning stage makes n infinitely approach n0, the primary field and noise signals are cancelled, and a relatively pure secondary field signal can be output by the adaptive filter.

[0022] In the third step, the signal is segmented using a sliding window method. Referring to the frequency of the primary field emission signal, a sliding window with an appropriate length is selected to ensure that the sliding window length contains the entire period of the signal, and the data within the sliding window is taken as one sample; at the same time, an appropriate spacing is selected to ensure sufficient utilization of the data while ensuring an adequate amount of sample data in the training set.

[0023] In the fourth step, the pre-trained ResNet network is used to complete the classification of the target signal. In the present invention, the ResNet18 neural network architecture is selected, that is, the ResNet neural network with 18 weight layers. The entire network can be divided into a weight layer and a normalization layer according to functions. Each weight layer contains a pooling layer, a convolutional layer, and a linear fully connected layer, mainly performing convolutional operations; the normalization layer contains batch normalization of data and a pooling layer, mainly performing normalization processing of data and data volume reduction for subsequent operation processing.

[0024] The feature signal first enters the first convolutional layer to preprocess the data features, and the expression of the number of output channels satisfies the following formula:

[0025]

[0026] where n out is the number of output channels, n in is the number of input channels, p is the image padding number, k is the convolutional kernel size, and s is the step size.

[0027] The convolved signal enters the normalization layer for batch normalization and max pooling. This pooling layer does not change the number of data channels but only changes the data size.

[0028] The normalized signal data then passes through multiple weight layers. In ResNet18, there are four residual blocks in total. Introducing the residual mapping makes the output more sensitive to changes and has a better effect on weight optimization. Inside the residual block, the signal features pass through two convolutional layers in sequence, and an identity mapping is introduced to solve the gradient vanishing and network degradation problems caused by the neural network. This mapping is expressed as:

[0029] F(x) = H(x) - x

[0030] Among them, H represents the identity mapping, F represents the residual, and x is the input.

[0031] Under the condition that the input and output have the same dimension, the expression of the residual block is as follows:

[0032] y = F(x, {W i}) + x

[0033] Among them, x is the input, y is the output, F represents the network mapping relationship before summation, that is, the residual, and {W i} represents the set of weights of the network within this residual block. If the dimensions are different, it is expressed as:

[0034] y = F(x, {W i}) + W i x

[0035] After the signal passes through the residual block, it enters the average pooling layer and then passes through the fully connected layer for softmax classification. The expression is as follows:

[0036]

[0037] Among them, x i is the i-th element of the input elements of the fully connected layer.

[0038] During the pre-training process, the data collected in advance is used for data processing and classification in the same way. The MSRA method is used for the weight initialization of the convolutional neural network. The cross-entropy function is selected as the loss function, and the Adam optimizer is used to optimize the loss function. The dropout method is used to prevent overfitting. After the model finally converges, it can quickly realize the classification of the input signal.

[0039] So far, the work of the airborne underwater anomaly detection method based on adaptive cancellation and ResNet neural network proposed by the present invention ends, and a relatively high detection accuracy can be obtained.

[0040] Advantages of an airborne underwater anomaly detection method based on adaptive cancellation and ResNet neural network designed by the present invention compared with the prior art:

[0041] (1) The present invention uses an airborne magnetoelectric antenna to emit the primary electromagnetic wave field and detect the secondary field generated by underwater metal anomalies, which can not only identify underwater ferromagnetic targets but also identify underwater non-ferromagnetic targets.

[0042] (2) The present invention uses the adaptive cancellation method to remove the geomagnetic field and the primary field, which can achieve high-precision feature extraction of the target electromagnetic signal.

[0043] (3) The present invention uses a sliding window method to achieve data segmentation and truncation, retaining the signal timing information.

[0044] (4) The present invention uses a pre-trained ResNet neural network for classification, with a reduced number of parameters and complexity. At the same time, the network is deeper and does not produce gradient disappearance. The signal classification has a fast processing speed and high accuracy. Description of the Drawings

[0045] Figure 1 is a flowchart of an airborne underwater anomaly detection method based on adaptive cancellation and ResNet neural network of the present invention;

[0046] Figure 2 is a block diagram of the adaptive cancellation system designed by the present invention;

[0047] Figure 3 is the sliding window type signal data segmentation method used by the present invention;

[0048] Figure 4 is the ResNet neural network architecture diagram used by the present invention. Detailed Embodiments

[0049] As Figure 1 shown, an airborne underwater anomaly detection method based on adaptive cancellation and ResNet neural network of the present invention includes the following steps: First, iterate the parameters of the adaptive filter in waters without abnormal objects until convergence, and store the converged filter parameters; Second, use the trained adaptive filter parameters to design a sliding window for the cycle of the electromagnetic wave emitted by the magnetoelectric antenna for the mixed signal received by the magnetic field sensor, and perform adaptive filtering on the three-axis time-series secondary field signal obtained by adaptive filtering; Then, design a sliding window according to the cycle of the electromagnetic wave emitted by the magnetoelectric antenna to segment the secondary field signal obtained by adaptive filtering; Finally, input the three-axis secondary field data into the ResNet neural network, and use the pre-trained weights to complete the binary classification detection of the presence or absence of the target signal. The principle block diagram of the whole method is as Figure 1 shown, and the specific implementation steps are as follows:

[0050] First, configure and train the adaptive cancellation system. The adaptive cancellation system consists of a transmitter, a receiver, and a data terminal. The transmitter consists of a signal generator, a voltage amplifier, and a magnetoelectric antenna. The filtering architecture is as shown in Figure 2 shown. The excitation signal generated by the signal generator is power-amplified by the high-voltage amplifier and then applied as a driving voltage to the magnetoelectric antenna. Due to the magnetoelectric coupling effect, the magnetoelectric antenna generates an electromagnetic field (primary field). Since the excitation signal generated by the signal generator has the same frequency and a fixed phase difference as the primary field signal generated by the magnetoelectric antenna, and the excitation signal can be directly measured, the excitation signal generated by the signal source can be used as a reference signal to estimate the primary field signal.

[0051] The receiver consists of a magnetic field sensor and a preamplifier. The magnetic field sensor receives the electromagnetic signals in the surrounding environment, and the preamplifier performs band-pass filtering on the signals received by the magnetic field sensor to filter out low-frequency ocean noise and high-frequency noise.

[0052] The data terminal consists of a data acquisition card and a PC adaptive filter program. The data acquisition card acquires the excitation signal of the signal generator and the pre-filtered signal of the preamplifier, and outputs these two sets of data to the PC for adaptive cancellation.

[0053] The electromagnetic field of an underwater metal target at a relatively long distance from the distance measurement system has the following complex expression:

[0054]

[0055]

[0056] H φ =0

[0057] where H r , H θ , H φ respectively represent the magnetic field intensities in the three coordinate axis directions of the spherical coordinate system of the magnetoelectric antenna, j represents the imaginary unit, r represents the distance from the underwater metal anomaly to the sensor, I0 represents the magnitude of the eddy current generated in the metal, θ represents the angle in the spherical coordinate system, k represents the complex wave number. Using the adaptive filter to process this sinusoidal magnetic field, its loss function can be expressed as:

[0058] e 2 (i)=(d(i)-W T (i)X(i)) 2

[0059] where W is the parameter of the adaptive filter, d is the actual value or expected value, X is the reference signal, and i represents the i-th iteration.

[0060] In addition, the iteration of W follows the following formula:

[0061] W(i + 1) = W(i) + μe(i)X(i)

[0062] where μ is the step size of the adaptive filter.

[0063] In the training stage, the parameters of the adaptive filter can be initially set to zero, and then the parameters of the adaptive filter are iterated according to the loss function and the step size. When the change amount of the parameter is less than the given threshold, it is considered that the parameters of the adaptive filter converge. After storing the parameters, they can be directly used in the application stage for adaptive filtering.

[0064] In the second step, the secondary field is extracted using the parameters of the adaptive filter. The output signal satisfies the following form:

[0065] s = s0 + n0 - n

[0066] where s is the output signal of the adaptive filter, s0 is the target secondary field signal, n0 is the primary field and noise collected by the magnetic field sensor in the cancellation stage, and n is the filtering result of the primary field and noise signals in the learning stage by the adaptive filter. Since the learning stage makes n infinitely approach n0, the primary field and noise signals are cancelled, and a relatively pure secondary field signal can be output by the adaptive filter.

[0067] In the third step, the signal is segmented using a sliding window method. Referring to the frequency of the primary field emission signal, a sliding window of appropriate length is selected to ensure that the sliding window length contains the entire period of the signal, and the data within the sliding window is taken as one sample; at the same time, an appropriate spacing is selected to ensure that the data is fully utilized while ensuring that the sample data volume of the training set is sufficient. The sliding window segmentation method is as Figure 3 shown.

[0068] In the fourth step, the pre-trained ResNet network is used to complete the classification of the target signal. In the present invention, the ResNet18 neural network architecture is selected, that is, the ResNet neural network containing 18 weight layers. The entire network can be divided into a weight layer and a normalization layer according to functions. Each weight layer contains a pooling layer, an activation function, and a linear fully connected layer, mainly performing convolution operations; the normalization layer contains batch normalization of data and a pooling layer, mainly performing normalization processing of data and data volume reduction to facilitate subsequent operation processing. The network framework is as Figure 4 shown.

[0069] The feature signal first enters the first convolutional layer to preprocess the data features, and the expression of the number of output channels satisfies the following formula:

[0070]

[0071] where n outis the number of output channels, n in is the number of input channels, p is the image padding number, k is the convolutional kernel size, and s is the stride.

[0072] After convolution, the signal enters the normalization layer for batch normalization and max pooling. The pooling layer does not change the number of data channels but only changes the size of the data.

[0073] The normalized signal data then passes through multiple weight layers. In ResNet18, there are four residual blocks in total. Introducing the residual mapping makes the output more sensitive to changes and has a better effect on weight optimization. Inside the residual block, the signal features pass through two convolutional layers in sequence, and an identity mapping is introduced to solve the problem of vanishing gradients and network degradation caused by the neural network. This mapping is expressed as:

[0074] F(x) = H(x) - x

[0075] where H represents the identity mapping, F represents the residual, and x is the input.

[0076] Under the condition of the same dimension of input and output, the expression of the residual block is as follows:

[0077] y = F(x, {W i}) + x

[0078] where x is the input, y is the output, F represents the network mapping relationship before summation, that is, the residual, and {W i} represents the set of weights of the network within this residual block. If the dimensions are different, it is expressed as:

[0079] y = F(x, {W i}) + W i x

[0080] After passing through the residual block, the signal is input into the average pooling layer and then through the fully connected layer for softmax classification. The expression is as follows:

[0081]

[0082] where x i is the i-th element of the input elements of the fully connected layer.

[0083] During the pre-training process, the data collected in advance is processed and classified in the same way. The cross-entropy function is selected as the loss function, the Adam optimizer is used to optimize the loss function, and the dropout method is used to prevent overfitting. After the model finally converges, it can quickly realize the classification of the input signal.

[0084] So far, the work of the airborne underwater anomaly detection method based on adaptive cancellation and ResNet neural network proposed by the present invention is completed, and a relatively high detection accuracy can be obtained.

[0085] The content not described in detail in the specification of the present invention belongs to the prior art well-known to those skilled in the art. Although the exemplary embodiments of the present invention have been described for the purpose of illustration, those skilled in the art will understand that various modifications, additions, substitutions, etc. can be made in form and detail without departing from the scope and spirit of the invention disclosed in the appended claims. All such changes shall fall within the protection scope of the appended claims of the present invention, and each part of the product and each step in the method claimed by the present invention can be combined in any combination. Therefore, the description of the embodiments disclosed in the present invention is not intended to limit the scope of the present invention, but to describe the present invention. Accordingly, the scope of the present invention is not limited by the above embodiments, but is defined by the claims or their equivalents.

Claims

1. An airborne underwater anomaly detection method based on adaptive cancellation and ResNet neural network, characterized in that It includes the following steps: (1) Use an adaptive cancellation system to train in waters without abnormal objects, iterate the parameters of the adaptive filter until convergence, minimize the mean square error between the estimated noise signal and the actual noise signal, and store the converged filter parameters; (2) On the basis of step (1), use the trained adaptive filter parameters to adaptively filter the mixed signal received by the magnetic field sensor, cancel the primary field signal, and extract a relatively pure three-axis time-series secondary field signal; (3) On the basis of step (2), design a sliding window according to the electromagnetic wave emission period of the magnetoelectric antenna, and segment the three-axis time-series secondary field signal obtained by adaptive filtering to obtain three-axis secondary field data; (4) On the basis of step (3), input the three-axis secondary field data into a ResNet neural network, and use the pre-trained weights to complete the binary classification detection of the presence or absence of the target signal; Use the pre-trained ResNet network to complete the classification of the target signal. Select the ResNet18 neural network architecture, that is, the ResNet neural network containing 18 weight layers. The entire network can be divided into a weight layer and a normalization layer according to functions. Each weight layer contains a pooling layer, an activation function, and a linear fully connected layer for convolution operations; the normalization layer contains data batch normalization and a pooling layer for data normalization processing and data volume reduction to facilitate subsequent operation processing; The feature signal first enters the first convolutional layer to preprocess the data features, and the expression of its output channel number satisfies the following formula: Among them, is the number of output channels, is the number of input channels, is the number of image padding, is the convolution kernel size, is the stride; The signal after convolution enters the normalization layer for batch normalization processing and maximum pooling processing. This pooling layer does not change the number of data channels, only changes the size of the data; The normalized signal data then passes through multiple weight layers. There are four residual blocks in ResNet18. Introducing the residual mapping makes the output more sensitive to changes and has a better effect on weight optimization; within the residual block, the signal features sequentially pass through two convolutional layers, and an identity mapping is introduced to solve the problem of gradient disappearance and network degradation caused by the neural network. This mapping is expressed as: Among them, represents the identity mapping, represents the residual, is the input; Under the condition of the same input and output dimensions, the expression of the residual block is as follows: Among them, is the input, is the output, represents the network mapping relationship before summation, i.e., the residual, represents the set of weights of the network within this residual block. If the dimensions are different, it is expressed as: The signal passes through the residual block and then enters the average pooling layer and passes through the fully connected layer softmax classification. The expression is as follows: Among them, is the th element of the input elements of the fully connected layer; During the pre-training process, use the previously collected data to perform data processing and classification in the same way. Select the cross-entropy function as the loss function, select the Adam optimizer to optimize the loss function, and use the dropout method to prevent overfitting. After the model finally converges, it can quickly realize the classification of the input signal.

2. The airborne underwater anomaly detection method based on adaptive cancellation and ResNet neural network according to claim 1, wherein: The adaptive cancellation system consists of a transmitting end, a receiving end, and a data end. The transmitting end consists of a signal generator, a voltage amplifier, and a magnetoelectric antenna. The excitation signal generated by the signal generator is power-amplified by a high-voltage amplifier and then applied as a driving voltage to the magnetoelectric antenna. Due to the magnetoelectric coupling effect, the magnetoelectric antenna generates a primary field signal. The excitation signal generated by the signal generator has the same frequency as the primary field signal generated by the magnetoelectric antenna and a fixed phase difference. The excitation signal generated by the signal generator is used as a reference signal to estimate the primary field signal. The receiving end consists of a magnetic field sensor and a preamplifier. The magnetic field sensor receives the electromagnetic signals in the surrounding environment, and the preamplifier performs band-pass filtering on the signals received by the magnetic field sensor to filter out low-frequency ocean noise and high-frequency noise. The data end consists of a data acquisition card and a PC. The data acquisition card acquires the excitation signal of the signal generator and the pre-filtered signal of the preamplifier and outputs these two sets of data to the PC for adaptive cancellation. The electromagnetic field of the underwater metal target at a relatively far distance from the distance measurement system has the same frequency as the transmitting end. An adaptive filter is used to process this electromagnetic field, and the parameters of the adaptive filter are iterated according to the loss function and the step size. When the change amount of this parameter is less than the given threshold, it is considered that the parameters of the adaptive filter converge.

3. An airborne underwater anomaly detection method based on adaptive cancellation and ResNet neural network according to claim 2, characterized in that: In step (2), the specific method of canceling the primary field signal is that the magnetic field sensor collects the same magnitude of the primary field during the training stage and the cancellation stage, and the collected noise meets the same expectation. The adaptive filter is used to remove the magnitude of the primary field during the cancellation stage, making the input approach the pure secondary field signal.

4. An airborne underwater anomaly detection method based on adaptive cancellation and ResNet neural network according to claim 3, characterized in that: The specific implementation method of step (3) is as follows: referring to the frequency of the primary field emission signal, a sliding window of appropriate length is selected to ensure that the sliding window length contains the entire period of the signal, and the data within the sliding window is used as one sample; at the same time, an appropriate spacing is selected to ensure full utilization of the data while ensuring sufficient sample data volume in the training set.

5. An airborne underwater anomaly detection method based on adaptive cancellation and ResNet neural network according to claim 4, characterized in that: The ResNet neural network architecture is a ResNet neural network containing 18 weight layers. The entire network is divided into a normalization layer and a weight layer according to functions. The normalization layer includes a data batch normalization layer and a pooling layer, which are used for data normalization processing to facilitate subsequent arithmetic processing. Each weight layer includes a pooling layer, an activation function, and a linear fully connected layer, which are used for convolution operations. The characteristic signal of the three-axis secondary field data first enters the first convolutional layer to preprocess the data characteristics. After preprocessing, the dimension and number of channels of the data characteristics change, and local feature extraction and fusion are achieved through convolution. After convolution, the signal enters the normalization layer for batch normalization processing and maximum pooling processing. This pooling layer does not change the number of data channels but only changes the size of the data. The normalized signal data then passes through multiple weight layers. There are four residual blocks in the ResNet neural grid, and the mapping of residuals is introduced. Inside the residual block, the signal features sequentially pass through two convolutional layers, and an identity mapping is introduced. Finally, after passing through the residual block, the signal is input into the average pooling layer and classified by the softmax function of the fully connected layer to realize the detection and judgment of the presence or absence of the signal.

6. The airborne underwater anomaly detection method based on adaptive cancellation and ResNet neural network according to claim 5, characterized in that: During the pre-training process, the data collected in advance is processed and classified in the same way. The cross-entropy function is selected as the loss function, and the Adam optimizer is used to optimize the loss function. The dropout method is used to prevent overfitting. After the model finally converges, the classification of the input signal is achieved.

Images