Real-time magnetic anomaly target recognition method and system based on physical constraint neural network

Abstract

This invention provides a real-time magnetic anomaly target identification method and system based on a physically constrained neural network (PCNN). The method includes: acquiring spatial geomagnetic total field data; preprocessing the total field data to generate input samples and labeling them with category tags; designing a lightweight convolutional neural network (HNN) as the backbone network, introducing physical constraint terms, and constructing a PCNN model. The physical constraint terms specifically include: the magnetic field in the passive region should satisfy the Laplace equation and the magnetic anomaly signal should conform to the dipole field decay law; training and optimizing the PCNN model based on the labeled input sample data and simulated target data; receiving the magnetic signal stream in real time, inputting it into the trained PCNN model according to time windows, and completing the target identification and output. This invention addresses the technical problems of low identification accuracy, poor interpretability, and weak generalization ability in existing magnetic anomaly target identification methods under conditions of small samples and low signal-to-noise ratio.

Description

Technical Field

[0001] This invention relates to the field of magnetic detection technology in geophysical exploration, and in particular to a method and system for real-time identification of magnetic anomaly targets based on a physically constrained neural network. Background Technology

[0002] Magnetic anomaly detection is an important means of passively detecting underwater targets. Traditional methods rely on threshold detection, matched filtering, or dipole inversion, but these are sensitive to complex background noise and struggle to adapt to the diversity of target morphologies. In recent years, deep learning methods (such as CNN and LSTM) have been introduced into magnetic signal recognition, achieving some progress, but the following problems still exist:

[0003] 1. Reliance on large amounts of labeled data: In practical applications, labeled samples are scarce, resulting in poor model generalization ability;

[0004] 2. Weak interpretability: Black box models are difficult to guarantee the reliability of decision-making, which is not conducive to military applications;

[0005] 3. Ignoring physical laws: Existing models do not utilize the physical priors of magnetic field propagation (Laplace equation, magnetic dipole spatial field distribution), resulting in a high false alarm rate at low signal-to-noise ratios;

[0006] 4. Insufficient real-time performance: Complex network structures are difficult to deploy on edge devices for real-time detection and identification.

[0007] In existing technologies, CN114675335A, "A Method for Processing Fluxgate Aeromagnetic Data Based on Machine Learning," proposes using a trained machine learning model to predict three-component fluxgate data to reduce steering errors. However, this method focuses on preprocessing for aeromagnetic compensation and does not involve target recognition. CN113484919B, "A Method, System, Terminal, and Medium for Magnetic Anomaly Inversion Based on Convolutional Neural Networks," first performs forward simulation to generate a sample dataset, then constructs a VGG magnetic anomaly inversion network, trains the VGG magnetic anomaly inversion network using the sample dataset, and finally inputs the magnetic anomaly data into the trained inversion network to obtain the results. This method is still a traditional, purely data-driven approach and does not impose physical constraints on the model during training, resulting in insufficient robustness and generalization ability. CN115826083A, "A Method, System, and Terminal for Joint Gravity and Magnetic Inversion Based on Deep Learning Coupling Constraints," still trains the network by implementing a "soft" mapping between gravity and magnetic data through pure data-driven methods, which suffers from inefficiency in practice. Currently, there is no known technical solution that embeds the physical equations of magnetic fields as regularization terms into the neural network training process to achieve collaborative identification through "data-driven + physical guidance". The real-time magnetic anomaly target identification method and system based on Physically Informed Neural Network (PCNN) proposed in this invention solves the problems of traditional methods relying on human experience and the high false alarm rate of pure data-driven AI models. It is particularly suitable for real-time detection, classification and identification of magnetic anomaly signals generated by underwater moving targets by magnetic detection systems mounted on UAVs, ships or fixed platforms, and can be widely used in scenarios such as marine anti-submarine warfare, maritime security, and underwater facility monitoring. Summary of the Invention

[0008] This invention provides a real-time magnetic anomaly target identification method and system based on a physical constraint neural network, which can solve the technical problems of low identification accuracy, poor interpretability and weak generalization ability of existing magnetic anomaly target identification methods under conditions of small sample size and low signal-to-noise ratio.

[0009] According to one aspect of the present invention, a real-time magnetic anomaly target identification method based on a physically constrained neural network is provided. The method includes: Step 1, collecting space geomagnetic total field data through a magnetic exploration platform, preprocessing the space geomagnetic total field data to generate input samples, and labeling the input sample data with category tags; Step 2, designing a lightweight convolutional neural network as the backbone network, introducing physical constraint terms, and constructing a physically constrained neural network (PCNN) model. The physical constraint terms specifically include: the magnetic field in the passive region should satisfy the Laplace equation and the magnetic anomaly signal should conform to the dipole field attenuation law; Step 3, training and optimizing the parameters of the physically constrained neural network (PCNN) model based on the category-labeled input sample data from Step 1 and the simulated target data; Step 4, receiving the magnetic signal stream in real time, inputting it into the trained physically constrained neural network (PCNN) model according to time windows, and completing the target identification and output.

[0010] Further, step one specifically includes: collecting spatial geomagnetic total field data B(lon,lat,height,mag,time) through a magnetic exploration platform (such as a drone), where lon, lat, and height represent the longitude, latitude, and altitude of a spatial point, respectively, mag represents the total magnetic field strength at that point, and time is the corresponding time, with a sampling rate ≥5Hz; preprocessing the raw signal and dividing it into a sliding time window (length T = 1024 points) to generate a set of input samples X; labeling the input sample data with category labels {0,1,2}, where 0 represents no target, 1 represents underwater moving target, and 2 represents other targets.

[0011] Furthermore, in step two, the lightweight convolutional neural network includes: an input layer, a convolutional layer, a pooling layer, and a fully connected layer. The input layer receives a one-dimensional magnetic signal of length T; there are three convolutional layers with kernel sizes of 16, 32, and 64, respectively, and the activation function is LeakyReLU; the pooling layer uses max pooling to reduce dimensionality; and the fully connected layer is used to output the class probability {y} = f{X}.

[0012] Furthermore, in step two, the loss function of the Physically Constrained Neural Network (PCNN) model is L = L_data + λ * L_physics, where L_data is the data error term (such as mean square error), L_physics is the physical constraint term, and λ is the weight coefficient.

[0013] Furthermore, in step two, the magnetic field in the source-free region satisfies Laplace's equation; in the region without magnetic sources, the geomagnetic field satisfies: In discrete signals, this equation is approximated as the arithmetic statistics of second-order differences: in, Let be the arithmetic statistics of the second difference of the discrete signal, N be the number of discrete points involved in the calculation, N = T - 3, T be the length of the discrete signal, B[i] be the magnetic field value of the i-th point, B[i-1] be the magnetic field value of the (i-1)-th point, and B[i+1] be the magnetic field value of the (i+1)-th point.

[0014] Furthermore, in step two, the magnetic anomaly signal conforms to the dipole field decay law, and the magnetic field strength at a distance r from the target satisfies B∝1 / r^3. In the time series, if the target passes through at a constant speed, the magnetic signal peak should approximately satisfy the power-law decay.

[0015] Furthermore, in step four, the trained Physically Constrained Neural Network (PCNN) model is deployed on an edge computing device to receive magnetic signal streams in real time and input them into the PCNN model according to time windows; the target category probability and confidence level are output; when the confidence level > threshold (0.85), an alarm is triggered, and the target type, occurrence time and location are output.

[0016] According to another aspect of the present invention, a real-time magnetic anomaly target identification system based on a physical constraint neural network is provided, which is used to implement the real-time magnetic anomaly target identification method based on a physical constraint neural network as described above.

[0017] Furthermore, the real-time magnetic anomaly target identification system includes a magnetic signal acquisition unit, an edge computing processing unit, a communication unit, a human-computer interaction unit, and a storage unit. The magnetic signal acquisition unit is used to collect space geomagnetic total field data; the edge computing processing unit includes an embedded AI processor for running a Physically Constrained Neural Network (PCNN) model; the communication unit supports 4G / 5G or satellite links for data uploading and command reception; the human-computer interaction unit is used to display identification results and confidence levels; and the storage unit is used to store raw data and identification logs.

[0018] This invention provides a real-time identification method for magnetic anomaly targets based on a physically constrained neural network. This method uses fundamental physical laws of magnetic fields (Laplace's equation, magnetic dipole spatial field distribution) as soft constraints embedded in the neural network loss function to construct a hybrid model of "physical guidance + data driving," achieving high-precision, robust, and interpretable real-time identification of weak magnetic anomaly signals. Compared with existing technologies, this invention has the following advantages:

[0019] 1. High recognition accuracy: The physical constraint model focuses on features that conform to physical laws, and the recognition accuracy is over 90% when the signal-to-noise ratio exceeds 3dB;

[0020] 2. Strong small-sample learning ability: Only 100 sets of labeled data are needed to train an effective model, reducing the dependence on a large number of labels and making it highly practical;

[0021] 3. High interpretability: The model output can be accompanied by a "physical consistency score" to assist human interpretation, which is more interpretable than the black box model of traditional machine learning;

[0022] 4. Good anti-interference performance: The physical constraint model suppresses false anomalies that do not conform to the laws of magnetic fields, reducing the false alarm rate by more than half compared with traditional methods;

[0023] 5. Excellent real-time performance: The lightweight network has a single-frame inference time of <50ms, which meets the requirements of real-time processing. Attached Figure Description

[0024] The accompanying drawings, which form part of this specification, are provided to further illustrate embodiments of the invention and, together with the textual description, explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0025] Figure 1 A flowchart illustrating the overall process of a real-time magnetic anomaly target identification method based on a physically constrained neural network according to a specific embodiment of the present invention is shown.

[0026] Figure 2 A schematic diagram of the physical constraint neural network structure and loss function provided according to a specific embodiment of the present invention is shown;

[0027] Figure 3 A comparison diagram of the recognition performance of a conventional CNN and the PCNN of the present invention, provided by a specific embodiment of the present invention, is shown. Detailed Implementation

[0028] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0030] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0031] According to a specific embodiment of the present invention, a real-time magnetic anomaly target identification method based on a physically constrained neural network is provided. This method includes: Step 1, collecting space geomagnetic total field data through a magnetic exploration platform, preprocessing the space geomagnetic total field data to generate input samples, and labeling the input sample data with category tags; Step 2, designing a lightweight convolutional neural network as the backbone network, introducing physical constraint terms, and constructing a physically constrained neural network (PCNN) model. The physical constraint terms specifically include: the magnetic field in the passive region should satisfy the Laplace equation and the magnetic anomaly signal should conform to the dipole field attenuation law; Step 3, training and optimizing the parameters of the physically constrained neural network (PCNN) model based on the category-labeled input sample data from Step 1 and the simulated target data; Step 4, receiving the magnetic signal stream in real time, inputting it into the trained physically constrained neural network (PCNN) model according to time windows, and completing the target identification and output.

[0032] This configuration provides a real-time magnetic anomaly target identification method based on a physically constrained neural network. This method uses fundamental physical laws of magnetic fields (Laplace's equation, magnetic dipole spatial field distribution) as soft constraints embedded in the neural network loss function to construct a hybrid "physical guidance + data-driven" model, achieving high-precision, robust, and interpretable real-time identification of weak magnetic anomaly signals. Compared with existing technologies, this invention has the following advantages:

[0033] 1. High recognition accuracy: The physical constraint model focuses on features that conform to physical laws, and the recognition accuracy is over 90% when the signal-to-noise ratio exceeds 3dB;

[0034] 2. Strong small-sample learning ability: Only 100 sets of labeled data are needed to train an effective model, reducing the dependence on a large number of labels and making it highly practical;

[0035] 3. High interpretability: The model output can be accompanied by a "physical consistency score" to assist human interpretation, which is more interpretable than the black box model of traditional machine learning;

[0036] 4. Good anti-interference performance: The physical constraint model suppresses false anomalies that do not conform to the laws of magnetic fields, reducing the false alarm rate by more than half compared with traditional methods;

[0037] 5. Excellent real-time performance: The lightweight network has a single-frame inference time of <50ms, which meets the requirements of real-time processing.

[0038] Specifically, in order to achieve real-time identification of magnetic anomaly targets, it is first necessary to collect space geomagnetic total field data through a magnetic exploration platform, preprocess the space geomagnetic total field data to generate input samples, and label the input sample data with category labels.

[0039] In this invention, step one specifically includes: collecting spatial geomagnetic total field data B(lon,lat,height,mag,time) through a magnetic exploration platform (such as a drone), where lon, lat, and height represent the longitude, latitude, and altitude of a spatial point, respectively, mag represents the total magnetic field strength at that point, and time is the corresponding time, with a sampling rate ≥5Hz; preprocessing the original signal and dividing it into a sliding time window (length T = 1024 points) to generate a set of input samples X; labeling the input sample data with category labels {0,1,2}, where 0 represents no target, 1 represents underwater moving target, and 2 represents other targets.

[0040] Next, a lightweight convolutional neural network was designed as the backbone network, and physical constraints were introduced to construct a Physically Constrained Neural Network (PCNN) model. The physical constraints specifically include: the magnetic field in the passive region should satisfy the Laplace equation and the magnetic anomaly signal should conform to the dipole field decay law.

[0041] In step two, the lightweight convolutional neural network includes an input layer, a convolutional layer, a pooling layer, and a fully connected layer. The input layer receives a one-dimensional magnetic signal of length T. There are three convolutional layers with kernel sizes of 16, 32, and 64, respectively, and the activation function is LeakyReLU. The pooling layer uses max pooling to reduce dimensionality, and the fully connected layer is used to output the class probability {y} = f{X}.

[0042] The loss function of the Physically Constrained Neural Network (PCNN) model is L = L_data + λ * L_physics, where L_data is the data error term (such as mean square error), L_physics is the physical constraint term, and λ is the weight coefficient.

[0043] In step two, the magnetic field in the source-free region satisfies Laplace's equation; in the region without magnetic sources, the geomagnetic field satisfies: in, To find the gradient, let B be the Earth's magnetic field; in discrete signals, this equation is approximated as the arithmetic statistics of second-order differences: in, Let be the arithmetic statistics of the second difference of the discrete signal, N be the number of discrete points involved in the calculation, N = T - 3, T be the length of the discrete signal, B[i] be the magnetic field value of the i-th point, B[i-1] be the magnetic field value of the (i-1)-th point, and B[i+1] be the magnetic field value of the (i+1)-th point.

[0044] In step two, the magnetic anomaly signal conforms to the dipole field decay law, and the magnetic field strength at a distance r from the target satisfies B∝1 / r^3. In the time series, if the target passes through at a constant speed, the magnetic signal peak should approximately satisfy the power-law decay.

[0045] Furthermore, the Physically Constrained Neural Network (PCNN) model is trained and its parameters are optimized based on the labeled input sample data and simulation target data from step one.

[0046] Finally, the magnetic signal stream is received in real time and input into the trained Physically Constrained Neural Network (PCNN) model according to the time window to complete the target recognition and output.

[0047] In step four, the trained Physical Constraint Neural Network (PCNN) model is deployed on an edge computing device to receive magnetic signal streams in real time and input them into the PCNN model according to time windows; the target category probability and confidence level are output; when the confidence level > threshold (0.85), an alarm is triggered and the target type, occurrence time and location are output.

[0048] According to another aspect of the present invention, a real-time magnetic anomaly target identification system based on a physical constraint neural network is provided, which is used to implement the real-time magnetic anomaly target identification method based on a physical constraint neural network as described above.

[0049] This configuration provides a real-time magnetic anomaly target identification system based on a physical constraint neural network. The system constructs a hybrid model of "physical guidance + data drive" by embedding the basic physical laws of magnetic fields (Laplace equation, magnetic dipole spatial field distribution) as soft constraints into the neural network loss function, thereby achieving high-precision, robust, and interpretable real-time identification of weak magnetic anomaly signals.

[0050] Furthermore, the real-time magnetic anomaly target identification system includes a magnetic signal acquisition unit, an edge computing processing unit, a communication unit, a human-computer interaction unit, and a storage unit. The magnetic signal acquisition unit is used to collect space geomagnetic total field data; the edge computing processing unit includes an embedded AI processor for running a Physically Constrained Neural Network (PCNN) model; the communication unit supports 4G / 5G or satellite links for data uploading and command reception; the human-computer interaction unit is used to display identification results and confidence levels; and the storage unit is used to store raw data and identification logs.

[0051] To gain a further understanding of the present invention, the following description is provided in conjunction with... Figures 1 to 3 The present invention provides a detailed description of the real-time identification method and system for magnetic anomaly targets based on physical constraint neural networks.

[0052] This invention aims to address the problems of low accuracy, poor interpretability, and weak generalization ability in existing magnetic anomaly target identification methods under conditions of small sample size and low signal-to-noise ratio. It proposes a real-time magnetic anomaly target identification method and system based on a physically constrained neural network. By embedding fundamental physical laws of magnetic fields (Laplace's equation, magnetic dipole spatial field distribution) as soft constraints into the neural network loss function, a hybrid model of "physical guidance + data driving" is constructed, achieving high-precision, robust, and interpretable real-time identification of weak magnetic anomaly signals.

[0053] This invention provides a real-time identification method for magnetic anomaly targets based on a physically constrained neural network, comprising the following steps:

[0054] Step S1: Acquisition and Preprocessing of Magnetic Anomaly Data

[0055] • Collect space geomagnetic total field data B(lon,lat,height,mag,time) through a magnetic exploration platform (such as a drone), where lon, lat, and height represent the longitude, latitude, and altitude of a spatial point, respectively, mag represents the total magnetic field strength at that point, and time is the corresponding time, with a sampling rate ≥5Hz;

[0056] • The original signal is preprocessed (including filtering, diurnal variation correction and leveling), and a sliding time window (length T = 1024 points) is divided to generate a set of input samples X;

[0057] • Label the sample data: Label the input sample data with category labels {0,1,2}, where 0 represents no target, 1 represents underwater moving target, and 2 represents other targets.

[0058] Step S2: Construct a Physically Constrained Neural Network (PCNN) model

[0059] A lightweight convolutional neural network is designed as the backbone network, including: an input layer that receives a one-dimensional magnetic signal of length T; three convolutional layers with kernel sizes of 16, 32, and 64, and the activation function being LeakyReLU; a pooling layer using max pooling to reduce dimensionality; and a fully connected layer that outputs the class probability {y} = f{X}.

[0060] • Introducing a physical constraint term: Define the model's loss function L = L_data + λ * L_physics, where L_data is the data error term (e.g., mean squared error), L_physics is the physical constraint term, and λ is the weighting coefficient. The physical constraint term L_physics in this invention includes:

[0061] Physical constraint 1: The magnetic field in the source-free region satisfies Laplace's equation.

[0062] In regions without magnetic sources, the Earth's magnetic field satisfies:

[0063]

[0064] In discrete signals, it is approximately a second-order difference:

[0065]

[0066] For samples identified as targets, this value should deviate significantly from 0.

[0067] Physical constraint 2: The magnetic anomaly signal conforms to the attenuation law of the dipole field.

[0068] The magnetic field strength at a distance r from the target satisfies B∝1 / r^3. In the time series, if the target passes through at a constant velocity, the magnetic signal should approximately exhibit power-law decay near its peak. Definition:

[0069]

[0070] Where k is the fitting constant, B represents the attenuation intensity of the magnetic anomaly signal. peak:end This represents the duration from the peak to the trough of the magnetic signal.

[0071] The loss function for the physical constraint terms is:

[0072]

[0073] Where λ1 and λ2 are weighting coefficients (λ1 = 0.2, λ2 = 0.3).

[0074] Step S3: Model Training and Parameter Optimization

[0075] • Train the PCNN model using a mixed dataset, which includes real-world data without targets (background), a small amount of labeled target data, and simulated target data (to enhance diversity);

[0076] • During training, backpropagation simultaneously optimizes This enables the network to not only fit the labels but also satisfy physical laws.

[0077] Step S4: Real-time target recognition and output

[0078] • Deploy the trained model on edge computing devices (such as Jetson AGX, domestic RK3588);

[0079] • Receive magnetic signal streams in real time and input them into the PCNN model according to time windows;

[0080] • Output the probability and confidence level of the target category. When the confidence level is greater than the threshold (0.85), an alarm is triggered. Output the target type, occurrence time and location (provided by the platform positioning system).

[0081] The present invention also provides a system for implementing the above method, comprising:

[0082] • Magnetic signal acquisition unit: high-sensitivity magnetometer, GNSS / BeiDou, IMU;

[0083] • Edge computing processing unit: Embedded AI processor, running PCNN models;

[0084] • Communication unit: Supports 4G / 5G or satellite links for data uploading and command reception;

[0085] • Human-computer interaction unit: Display screen or mobile APP, showing recognition results and confidence level;

[0086] • Storage unit: Local storage of raw data and identification logs.

[0087] Compared with the prior art, the present invention has the following advantages:

[0088] 1. High recognition accuracy: The physical constraint model focuses on features that conform to physical laws, and the recognition accuracy is over 90% when the signal-to-noise ratio exceeds 3dB;

[0089] 2. Strong small-sample learning ability: Only 100 sets of labeled data are needed to train an effective model, reducing the dependence on a large number of labels and making it highly practical;

[0090] 3. High interpretability: The model output can be accompanied by a "physical consistency score" to assist human interpretation, which is more interpretable than the black box model of traditional machine learning;

[0091] 4. Good anti-interference performance: The physical constraint model suppresses false anomalies that do not conform to the laws of magnetic fields, reducing the false alarm rate by more than half compared with traditional methods;

[0092] 5. Excellent real-time performance: The lightweight network has a single-frame inference time of <50ms, which meets the requirements of real-time processing.

[0093] The following describes a specific embodiment of the present invention:

[0094] • Hardware platform: Hexacopter drone equipped with a cesium optically pumped magnetometer (sensitivity 0.01nT, sampling rate 10Hz), flight altitude 80m, flight time 30min; Edge device: NVIDIA Jetson Xavier NX, computing power 21TOPS; Positioning: Beidou III + RTK, positioning accuracy ≤1m.

[0095] Software configuration:

[0096] Operating system: Ubuntu 18.04 + ROS;

[0097] Deep learning framework: PyTorch + TensorRT optimization;

[0098] PCNN model parameters: λ1 = 0.2, λ2 = 0.3, learning rate 1e -3 The number of parameters in the PCNN network is approximately 85k.

[0099] • Data: Data was collected from a sea area in the South China Sea, totaling 10 hours, including 3 types of samples (no target, underwater moving target, and other targets), with a signal-to-noise ratio of 1-5dB; the time window for a single data point is T=1024.

[0100] • Test results:

[0101] Training: Using 80 sets of real-world labeled data + 200 sets of simulated data for training, the validation set accuracy was 95.7%; Traditional CNN (without physical constraints): accuracy 80.2%, false positive rate 28.5%; PCNN of this invention: accuracy 94.1%, false positive rate 7.3%; inference speed: 42ms / frame (Jetson NX), physical consistency score matches human judgment by >90%.

[0102] In summary, this invention provides a real-time identification method for magnetic anomaly targets based on a physically constrained neural network. This method constructs a hybrid model of "physical guidance + data drive" by embedding the fundamental physical laws of magnetic fields (Laplace's equation, spatial field distribution of magnetic dipoles) as soft constraints into the loss function of the neural network, thereby achieving high-precision, robust, and interpretable real-time identification of weak magnetic anomaly signals.

[0103] For ease of description, spatial relative terms such as "above," "on top of," "on the upper surface of," "above," etc., are used herein to describe the spatial positional relationship of a device or feature as shown in the figures to other devices or features. It should be understood that spatial relative terms are intended to encompass different orientations in use or operation beyond the orientation of the device as described in the figures. For example, if the device in the figures were inverted, a device described as "above" or "on top of" other devices or structures would subsequently be positioned as "below" or "under" other devices or structures. Thus, the exemplary term "above" can include both "above" and "below." The device may also be positioned in other different ways (rotated 90 degrees or in other orientations), and the spatial relative descriptions used herein will be interpreted accordingly.

[0104] Furthermore, it should be noted that the use of terms such as "first" and "second" to define components is merely for the purpose of distinguishing the corresponding components. Unless otherwise stated, the above terms have no special meaning and therefore should not be construed as limiting the scope of protection of this invention.

[0105] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A real-time identification method for magnetic anomaly targets based on a physically constrained neural network, characterized in that, The real-time magnetic anomaly target identification method based on physical constraint neural networks includes: Step 1: Collect space geomagnetic total field data through the magnetic exploration platform, preprocess the space geomagnetic total field data to generate input samples, and label the input sample data with category labels; Step 2: Design a lightweight convolutional neural network as the backbone network, introduce physical constraints, and construct a Physically Constrained Neural Network (PCNN) model. The physical constraints specifically include: the magnetic field in the passive region should satisfy the Laplace equation and the magnetic anomaly signal should conform to the dipole field decay law. Step 3: Based on the labeled input sample data and simulation target data from Step 1, perform Physically Constrained Neural Network (PCNN) model training and parameter optimization; Step four: Receive the magnetic signal stream in real time, input it into the trained Physically Constrained Neural Network (PCNN) model according to the time window, and complete the target recognition and output.

2. The real-time identification method for magnetic anomaly targets based on a physically constrained neural network according to claim 1, characterized in that, Step one specifically includes: The total geomagnetic field data B(lon,lat,height,mag,time) is collected by a magnetic exploration platform (such as a drone), where lon, lat, and height represent the longitude, latitude, and altitude of a point in space, respectively, mag represents the total magnetic field strength at that point, and time is the corresponding time. The sampling rate is ≥5Hz. The original signal is preprocessed and divided into a sliding time window (length T = 1024 points) to generate a set of input samples X; Label the input sample data with category labels {0,1,2}, where 0 represents no target, 1 represents underwater moving target, and 2 represents other targets.

3. The real-time identification method for magnetic anomaly targets based on a physically constrained neural network according to claim 2, characterized in that, In step two, the lightweight convolutional neural network includes an input layer, a convolutional layer, a pooling layer, and a fully connected layer. The input layer receives a one-dimensional magnetic signal of length T. The convolutional layer consists of three layers with kernel sizes of 16, 32, and 64, and the activation function is LeakyReLU. The pooling layer uses max pooling to reduce dimensionality. The fully connected layer is used to output the class probability {y} = f{X}.

4. The real-time identification method for magnetic anomaly targets based on a physically constrained neural network according to claim 3, characterized in that, In step two, the loss function of the Physically Constrained Neural Network (PCNN) model is L = L_data + λ * L_physics, where L_data is the data error term (such as mean square error), L_physics is the physical constraint term, and λ is the weight coefficient.

5. The real-time identification method for magnetic anomaly targets based on a physically constrained neural network according to claim 4, characterized in that, In step two, the magnetic field in the source-free region satisfies the Laplace equation; in the region without magnetic sources, the geomagnetic field satisfies: ▽ 2 B = 0; In discrete signals, this equation is approximately a second-order difference arithmetic statistics: in, Let be the arithmetic statistics of the second difference of the discrete signal, N be the number of discrete points involved in the calculation, N = T - 3, T be the length of the discrete signal, B[i] be the magnetic field value of the i-th point, B[i-1] be the magnetic field value of the (i-1)-th point, and B[i+1] be the magnetic field value of the (i+1)-th point.

6. The real-time identification method for magnetic anomaly targets based on a physically constrained neural network according to claim 5, characterized in that, In step two, the magnetic anomaly signal conforms to the dipole field decay law, and the magnetic field strength at a distance r from the target satisfies B∝1 / r^3. In the time series, if the target passes through at a constant speed, the magnetic signal peak should approximately satisfy the power-law decay.

7. The real-time identification method for magnetic anomaly targets based on a physically constrained neural network according to claim 6, characterized in that, In step four, the trained Physically Constrained Neural Network (PCNN) model is deployed on an edge computing device to receive magnetic signal streams in real time and input them into the PCNN model according to time windows; the target category probability and confidence level are output; when the confidence level > threshold (0.85), an alarm is triggered and the target type, occurrence time and location are output.

8. A real-time magnetic anomaly target identification system based on a physically constrained neural network, characterized in that, The real-time magnetic anomaly target identification system based on physical constraint neural network is used to implement the real-time magnetic anomaly target identification method based on physical constraint neural network as described in claims 1 to 7.

9. The real-time magnetic anomaly target identification system based on a physically constrained neural network according to claim 8, characterized in that, The real-time magnetic anomaly target identification system includes a magnetic signal acquisition unit, an edge computing processing unit, a communication unit, a human-computer interaction unit, and a storage unit. The magnetic signal acquisition unit is used to collect space geomagnetic total field data. The edge computing processing unit includes an embedded AI processor for running a Physically Constrained Neural Network (PCNN) model. The communication unit supports 4G / 5G or satellite links for data uploading and command reception. The human-computer interaction unit is used to display identification results and confidence levels. The storage unit is used to store raw data and identification logs.

Images