A high-precision magnetic target classification system based on coarse-fine two stages

Abstract

The application provides a high-precision magnetic target classification system based on coarse-precision two-stage, relates to the technical field of magnetic target classification, and comprises a magnetic field data acquisition component, a magnetic field data acquisition component configured to acquire magnetic field data of a target to be detected; a magnetic field data processing component, a magnetic field data processing component configured to sequentially perform coarse processing and fine processing according to the magnetic field data to obtain a global context enhanced feature vector corresponding to the target to be detected; the coarse processing comprises using a grey wolf algorithm and a physical information neural network to perform rapid inversion of magnetic distance parameters; the fine processing comprises using a space-time convolution neural network and a multi-head self-attention mechanism to perform dynamic focusing on key magnetic field mutations; and a target classification component, a target classification component configured to determine a target classification probability according to the global context enhanced feature vector to obtain a target classification result of the target to be detected. The application solves the problems of low reliability and practicability of the existing magnetic induction triggering technology through the above system.

Description

A high-precision magnetic target classification system based on coarse and fine levels Technical Field

[0001] This invention relates to the field of magnetic target classification technology, and in particular to a high-precision magnetic target classification system based on coarse and fine levels. Background Technology

[0002] Magnetic target signal classification and identification technology, as an important non-destructive testing method, lies in its efficient perception and identification of objects with ferromagnetic properties. Since its inception, this technology has undergone profound evolution from analog circuits to digital processing, and from single-parameter discrimination to intelligent pattern recognition. Its technical framework has developed from a simple sensor that only responded to the absolute value of magnetic field strength to a complex system integrating high-precision sensing, real-time signal processing, and intelligent decision-making algorithms. With the advent of the Internet of Things (IoT), underwater UUVs, and deep-sea exploration, higher demands are placed on this technology: it not only needs to achieve ultra-high reliability in target perception, but also must possess excellent anti-interference capabilities and extremely high identification accuracy in complex marine environments to meet the challenges of increasingly complex application scenarios.

[0003] Traditional magnetic induction triggering devices generally employ a response mechanism based on a magnetic field strength threshold. While the logic is simple, it has obvious inherent flaws. Such systems are highly susceptible to interference magnetic fields generated by other electromagnetic devices or moving ferromagnetic objects in the environment, leading to false alarms or missed alarms, thus significantly reducing the reliability and practicality of the system. Summary of the Invention

[0004] This application provides a high-precision magnetic target classification system based on coarse and fine levels to solve the problems of low reliability and practicality of existing magnetic induction triggering technology.

[0005] The system includes:

[0006] A magnetic field data acquisition component, configured to acquire magnetic field data of a target to be detected;

[0007] A magnetic field data processing component is configured to perform coarse processing and fine processing on the magnetic field data sequentially to obtain a global context-enhanced feature vector corresponding to the target to be detected. The coarse processing includes using the Grey Wolf algorithm and a physical information neural network to quickly invert the magnetic moment parameter. The fine processing includes using a spatiotemporal convolutional neural network and a multi-head self-attention mechanism to dynamically focus on key magnetic field mutations.

[0008] A target classification component is configured to determine the target classification probability based on the global context-enhanced feature vector to obtain the target classification result of the target to be detected.

[0009] Preferably, the magnetic field data processing component includes:

[0010] A coarse processing module is configured to use the Grey Wolf algorithm and the physical information neural network to quickly invert the magnetic field parameters of the magnetic field data, thereby obtaining the target positioning range and magnetic field time series data of the target to be detected.

[0011] The fine processing module is configured to use a spatiotemporal convolutional neural network and a multi-head self-attention mechanism to dynamically focus on key magnetic field mutations in the magnetic field time series data within the target positioning range, thereby obtaining the global context-enhanced feature vector.

[0012] Preferably, the physical constraints in the loss function of the physical information neural network include the Laplace equation and the magnetic dipole field formula.

[0013] Preferably, the coarse processing module includes:

[0014] The Grey Wolf Algorithm Magnetic Distance Inversion Unit has a pre-built and trained hybrid array magnetic field measurement model built into it. The Grey Wolf Algorithm Magnetic Distance Inversion Unit is configured to perform fast global inversion of the magnetic field data through the hybrid array magnetic field measurement model and the Grey Wolf algorithm to obtain the preliminary positioning range of the target location of the target to be detected.

[0015] The physical information neural network magnetic distance inversion unit has a pre-built and trained physical information neural network built into it. The physical information neural network magnetic distance inversion unit is configured to perform fine inversion on the magnetic field data of the preliminary positioning range through the physical information neural network to obtain the target positioning range and the magnetic field time series data of the target to be detected.

[0016] Preferably, the fast global inversion includes:

[0017] The magnetic field data is measured using the hybrid array magnetic field measurement model to obtain magnetic field measurement range data.

[0018] The gray wolf algorithm is used to invert the magnetic field measurement range data to obtain gray wolf inversion data;

[0019] Based on the gray wolf inversion data, the position is calculated to obtain the preliminary positioning range of the target location of the target to be detected.

[0020] Preferably, the fine processing module includes:

[0021] The spatiotemporal convolutional feature extraction unit has a pre-built and trained spatiotemporal convolutional neural network built in it. The spatiotemporal convolutional feature extraction unit is configured to perform multi-layer convolution processing on the magnetic field time series data in the target positioning range to obtain high-level feature representation.

[0022] A multi-head self-attention mechanism focusing unit is configured to perform multi-head attention mechanism calculation on the high-level feature representation to obtain the global context-enhanced feature vector.

[0023] Preferably, the multi-layer convolution processing includes:

[0024] Local spatiotemporal features in the magnetic field time series data within the target positioning range are extracted by multi-layer convolution in the spatiotemporal convolutional neural network.

[0025] The local spatiotemporal features are output in the form of the high-level features.

[0026] Preferably, the multi-head attention mechanism calculation includes:

[0027] Parallel computation is performed based on the high-level feature representation to generate multiple attention weight distribution maps;

[0028] The magnetic field time series data is dynamically focused according to the corresponding attention weight distribution map to obtain multiple key abrupt change regions; each key abrupt change region corresponds one-to-one with the attention weight distribution map.

[0029] All the key mutation regions are spliced ​​and linearly transformed to obtain the global context-enhanced feature vector.

[0030] Preferably, the target classification component is further configured as follows:

[0031] The global context enhancement feature vector is calculated using a built-in fully connected layer and a Softmax classifier to obtain the probability distribution of several target categories.

[0032] The target classification result of the target to be detected is obtained by sorting and filtering according to all the probability distributions.

[0033] Preferably, the magnetic field data acquisition component includes:

[0034] A magnetic sensing module, configured to acquire the magnetic signal of the target to be detected;

[0035] A data acquisition module, configured to convert the magnetic signal into magnetic field data;

[0036] A data transmission module, configured to transmit the magnetic field data outward.

[0037] As described above, this application provides a high-precision magnetic target classification system based on coarse and fine processing. The system includes a magnetic field data acquisition component configured to acquire magnetic field data of a target to be detected; a magnetic field data processing component configured to perform coarse and fine processing sequentially on the magnetic field data to obtain a global context-enhanced feature vector corresponding to the target; the coarse processing includes using the Grey Wolf algorithm and a physical information neural network to quickly invert magnetic moment parameters; the fine processing includes using a spatiotemporal convolutional neural network and a multi-head self-attention mechanism to dynamically focus on key magnetic field mutations; and a target classification component configured to determine the target classification probability based on the global context-enhanced feature vector to obtain the target classification result of the target to be detected. This application solves the problems of low reliability and practicality of existing magnetic induction triggering technologies through the above system. Attached Figure Description

[0038] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 is a schematic diagram of a high-precision magnetic target classification system based on coarse and fine levels according to this application;

[0040] Figure 2 is a schematic diagram of the magnetic field data processing component in a high-precision magnetic target classification system based on coarse and fine levels according to this application;

[0041] Figure 3 is a schematic diagram of the coarse processing module in a high-precision magnetic target classification system based on coarse and fine levels according to this application;

[0042] Figure 4 is a schematic diagram of the fine processing module in a high-precision magnetic target classification system based on coarse and fine levels according to this application;

[0043] Figure 5 is a schematic diagram of the target classification component in a high-precision magnetic target classification system based on coarse and fine levels according to this application. Detailed Implementation

[0044] 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 some embodiments of the present invention, and not all embodiments. 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.

[0045] It should be noted that the brief descriptions of terms in this application are only for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0046] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0047] Figure 1 is a schematic diagram of a high-precision magnetic target classification system based on coarse and fine levels according to this application.

[0048] As shown in Figure 1, this embodiment provides a high-precision magnetic target classification system based on coarse and fine levels. The system includes:

[0049] A magnetic field data acquisition component, configured to acquire magnetic field data of a target to be detected.

[0050] Specifically, in this embodiment, the magnetic field data of the target to be detected is obtained through the magnetic field data acquisition component, thereby providing a data foundation for subsequent target classification.

[0051] Figure 5 is a schematic diagram of the target classification component in a high-precision magnetic target classification system based on coarse and fine levels according to this application.

[0052] As shown in Figure 5, further, in some embodiments, the magnetic field data acquisition component includes:

[0053] A magnetic sensing module, configured to acquire the magnetic signal of the target to be detected;

[0054] A data acquisition module, configured to convert the magnetic signal into magnetic field data;

[0055] A data transmission module, configured to transmit the magnetic field data outward.

[0056] Specifically, in this embodiment, the magnetic signal of the target to be detected is captured by the magnetic sensing module and transmitted to the data acquisition module. The data acquisition module converts the magnetic signal into magnetic field data, and finally the data transmission module sends the magnetic field data outward.

[0057] The system also includes:

[0058] A magnetic field data processing component is configured to perform coarse processing and fine processing on the magnetic field data sequentially to obtain a global context-enhanced feature vector corresponding to the target to be detected. The coarse processing includes using the Grey Wolf algorithm and a physical information neural network to quickly invert the magnetic moment parameter. The fine processing includes using a spatiotemporal convolutional neural network and a multi-head self-attention mechanism to dynamically focus on key magnetic field mutations.

[0059] Specifically, in this embodiment, since the magnetic field data of the target to be detected has many features, and in order to accurately obtain the precise classification of the target to be detected, it is necessary to analyze and calculate both the local and global features in the magnetic field data. Therefore, the magnetic field data processing component performs coarse processing and fine processing on the magnetic field data in sequence to obtain all the features of the magnetic field data.

[0060] The coarse processing is based on the Gray Wolf algorithm and the physical information neural network for inversion. The Gray Wolf algorithm optimizes the magnetic distance inversion parameters to improve target recognition accuracy, and the physical information neural network performs a more refined inversion with physical laws.

[0061] The Gray Wolf Algorithm, inspired by the hunting behavior of gray wolves, is an optimization search method characterized by strong convergence performance, few parameters, and ease of implementation. It has been widely applied in fields such as workshop scheduling, parameter optimization, and image classification.

[0062] Gray wolves belong to the canid family and live in packs, occupying the top of the food chain. They strictly adhere to a social hierarchy of dominance.

[0063] The first level of the social hierarchy: The alpha wolf in the wolf pack is designated α. The alpha wolf is primarily responsible for making decisions regarding hunting, roosting, and rest schedules. Because the other wolves must obey the alpha wolf's commands, the alpha wolf is also known as the dominant wolf.

[0064] Furthermore, while the alpha wolf may not be the strongest wolf in the pack, it is certainly the best in terms of management ability.

[0065] The second tier of the social hierarchy consists of the beta wolf, which is subordinate to the alpha wolf and assists the alpha wolf in decision-making. Upon the death or old age of the alpha wolf, the beta wolf becomes the top candidate for successor. Although the beta wolf is subordinate to the alpha wolf, it can dominate wolves at other social tiers.

[0066] The third level of the social hierarchy: the δ wolf, which is subordinate to the α and β wolves, and dominates the remaining wolves. The δ wolf typically consists of cubs, sentinel wolves, hunting wolves, older wolves, and caretaker wolves.

[0067] The fourth level of the social hierarchy: the ω wolf, which is usually subordinate to wolves at other social levels. Although wolves may seem to play a minor role in the pack, without the presence of the ω wolf, internal problems such as cannibalism would arise within the pack.

[0068] The GWO optimization process includes steps such as social hierarchy stratification, tracking, surrounding, and attacking prey. The specific steps are as follows:

[0069] (1) Social Hierarchy

[0070] When designing a Gray Wolf Wolf (GWO), the first step is to construct a hierarchical model of the gray wolf social hierarchy. The fitness of each individual in the population is calculated, and the three gray wolves with the best fitness are labeled α, β, and δ, respectively, while the remaining gray wolves are labeled ω. In other words, the social hierarchy of the gray wolf pack, from highest to lowest, is: α, β, δ, ω. The optimization process of the GWO is primarily guided by the three best solutions (α, β, δ) from each generation of the population.

[0071] (2) Encircling Prey

[0072] When a gray wolf hunts, it gradually approaches and surrounds its prey. The mathematical model for this behavior is as follows:

[0073]

[0074]

[0075]

[0076] ;

[0077] In the formula: t is the current iteration number; This represents the Hadamard product operation; A and C are the co-coefficient vectors; X p Let represent the prey's position vector; X(t) represent the current gray wolf's position vector; throughout the iteration, a decreases linearly from 2 to 0; r1 and r2 are random vectors in [0, 1].

[0078] (3) Hunting

[0079] Gray wolves possess the ability to identify the location of potential prey (optimal solutions), and the search process is primarily guided by gray wolves α, β, and δ. However, the solution space characteristics of many problems are unknown, making it impossible for gray wolves to determine the precise location of prey (optimal solutions). To simulate the search behavior of gray wolves (candidate solutions), it is assumed that α, β, and δ have a strong ability to identify the location of potential prey. Therefore, in each iteration, the three best gray wolves (α, β, δ) in the current population are retained, and the positions of other search agents (including ω) are updated based on their positional information.

[0080] (4) Attacking Prey: In the process of constructing the attack prey model, according to the formula in (2), the decrease in the value of a will cause the value of A to fluctuate accordingly. In other words, A is a random vector in the interval [-a, a], where a decreases linearly during the iteration process. When A is in the interval [-1, 1], the next position of the Search Agent can be anywhere between the current gray wolf and the prey.

[0081] (5) Search for Prey: Gray wolves primarily rely on information from α, β, and δ to find prey. They begin by searching for prey location information in a dispersed manner, and then concentrate their attacks on the prey. For the establishment of the dispersed model, by making the search agent move away from the prey through |A|>1, this search method enables GWO to perform a global search. Another search coefficient in the GWO algorithm is C. As can be seen from the formula in (2), the C vector is a vector composed of random values ​​in the interval [0, 2]. This coefficient provides random weights for the prey so that it can be increased (|C|>1) or decreased (|C|<1). This helps GWO exhibit random search behavior during the optimization process to avoid the algorithm getting trapped in local optima. It is worth noting that C does not decrease linearly. C is a random value during the iteration process. This coefficient is beneficial for the algorithm to escape local optima, especially in the later stages of the iteration.

[0082] The system also includes:

[0083] A target classification component is configured to determine the target classification probability based on the global context-enhanced feature vector to obtain the target classification result of the target to be detected.

[0084] Specifically, in this embodiment, after obtaining the global context enhancement feature vector with complete magnetic field data features, the target classification component is used to determine the target classification probability, thereby obtaining the target classification result of the target to be detected.

[0085] Furthermore, in some embodiments, the target classification component is also configured to:

[0086] The global context enhancement feature vector is calculated using a built-in fully connected layer and a Softmax classifier to obtain the probability distribution of several target categories.

[0087] The target classification result of the target to be detected is obtained by sorting and filtering according to all the probability distributions.

[0088] Specifically, in this embodiment, the target classification component receives the global context-enhanced feature vector; the feature vector is calculated using a fully connected layer and a Softmax classifier to obtain the probability distribution of the target category; and the final target classification category is output based on the probability distribution, thereby achieving high-precision and real-time classification of magnetic targets in complex marine environments.

[0089] It should be noted that the final target classification result is obtained by sorting all the probability distributions and taking the probability distribution with the largest value as the target classification result.

[0090] Figure 2 is a schematic diagram of the magnetic field data processing component in a high-precision magnetic target classification system based on coarse and fine levels according to this application.

[0091] Referring to Figure 2, further, in some embodiments, the magnetic field data processing component includes:

[0092] The coarse processing module is configured to use the Grey Wolf algorithm and the physical information neural network to quickly invert the magnetic field parameters of the magnetic field data, thereby obtaining the target positioning range and magnetic field time series data of the target to be detected.

[0093] Specifically, in this embodiment, the coarse processing module in the magnetic field data processing component performs the coarse processing to obtain the target location range and magnetic field time series data of the target to be detected. In the initial stage, the Grey Wolf Algorithm (GWO) and Physical Information Neural Network (PINN) are used to quickly invert the magnetic moment parameters and narrow down the target location range.

[0094] Figure 3 is a schematic diagram of the coarse processing module in a high-precision magnetic target classification system based on coarse and fine levels according to this application.

[0095] As shown in Figure 3, further, in some embodiments, the coarse processing module includes:

[0096] The Grey Wolf Algorithm Magnetic Distance Inversion Unit has a pre-built and trained hybrid array magnetic field measurement model built into it. The Grey Wolf Algorithm Magnetic Distance Inversion Unit is configured to perform a fast global inversion of the magnetic field data through the hybrid array magnetic field measurement model and the Grey Wolf algorithm to obtain the preliminary positioning range of the target location of the target to be detected.

[0097] Specifically, in this embodiment, the Grey Wolf algorithm magnetic moment inversion unit constructs a hybrid array magnetic field measurement model; using the output data of the hybrid array magnetic field measurement model, the Grey Wolf optimization algorithm is executed to perform a fast global inversion of the target magnetic moment parameters; based on the magnetic moment parameter inversion results, the preliminary positioning range of the target location is calculated and output.

[0098] The coarse processing module also includes:

[0099] The physical information neural network magnetic distance inversion unit has a pre-built and trained physical information neural network built into it. The physical information neural network magnetic distance inversion unit is configured to perform fine inversion on the magnetic field data of the preliminary positioning range through the physical information neural network to obtain the target positioning range and the magnetic field time series data of the target to be detected.

[0100] Specifically, in this embodiment, the physical information neural network magnetic moment inversion unit constructs a physical information neural network in which the Laplace equation and magnetic dipole field formula describing the magnetic field distribution are embedded as physical constraints into the loss function of the neural network; the physical information neural network is used to perform refined inversion on the magnetic field data within the initial positioning range to further optimize the magnetic moment parameters; and the optimized magnetic moment parameters and the reduced target positioning range are output.

[0101] The magnetic field data processing component also includes:

[0102] The fine processing module is configured to use a spatiotemporal convolutional neural network and a multi-head self-attention mechanism to dynamically focus on key magnetic field mutations in the magnetic field time series data within the target positioning range, thereby obtaining the global context-enhanced feature vector.

[0103] Specifically, in this embodiment, during the fine discrimination stage, key magnetic field mutations are dynamically focused through a spatiotemporal convolutional neural network (ST-CNN) and a multi-head self-attention mechanism (MSA) to enhance classification robustness.

[0104] Figure 4 is a schematic diagram of the fine processing module in a high-precision magnetic target classification system based on coarse and fine levels according to this application.

[0105] Referring to Figure 4, further, in some embodiments, the fine processing module includes:

[0106] The spatiotemporal convolutional feature extraction unit has a pre-built and trained spatiotemporal convolutional neural network built in it. The spatiotemporal convolutional feature extraction unit is configured to perform multi-layer convolution processing on the magnetic field time series data in the target positioning range to obtain high-level feature representation.

[0107] Specifically, in this embodiment, the spatiotemporal convolutional feature extraction unit receives the magnetic field time series data for the coarse processing stage; through multi-layer convolution operations of the spatiotemporal convolutional neural network, local spatiotemporal features in the magnetic field time series data are extracted; and the extracted local spatiotemporal features are output as a high-level feature representation.

[0108] It should be noted that outputting the local spatiotemporal features as a high-level feature representation can be understood as, for example:

[0109] An object W has W1, W2, W3...Wn different types of features. However, after multi-layer convolution processing, only W can be represented, and the features W1, W2, W3...Wn are missing. Therefore, if we want to better represent the object W, we need to attach the features W1, W2, W3...Wn to the object W by using high-level feature representation, that is, outputting the local spatiotemporal features as high-level feature representation as mentioned above.

[0110] The fine processing module also includes:

[0111] A multi-head self-attention mechanism focusing unit is configured to perform multi-head attention mechanism calculation on the high-level feature representation to obtain the global context-enhanced feature vector.

[0112] Specifically, in this embodiment, the multi-head self-attention mechanism focusing unit performs parallel computation on the high-level feature representation through the multi-head self-attention mechanism to generate multiple attention weight distribution maps; based on the attention weight distribution maps, it dynamically focuses on the key abrupt change region in the magnetic field time series data; and it concatenates and linearly transforms the outputs of multiple attention heads to obtain a global context-enhanced feature vector.

[0113] This embodiment has the following advantages:

[0114] The system employs a coarse-processing stage using the Grey Wolf algorithm and a physical information neural network to rapidly invert magnetic moment parameters, effectively narrowing the target localization range and reducing false positives. The fine-processing stage utilizes a spatiotemporal convolutional neural network and a multi-head self-attention mechanism to dynamically focus on key magnetic field mutations, enhancing the robustness of feature extraction and significantly reducing false positives and false negatives caused by environmental interference, a problem common in traditional techniques. This two-stage collaborative approach—coarse-processing for rapid localization and fine-processing for refined analysis—optimizes the overall computational flow and is suitable for real-time or near-real-time applications, such as underwater detection or IoT devices. The system demonstrates strong anti-interference capabilities in complex environments (such as the ocean), improving target recognition stability in noisy environments through a combination of physical constraints and intelligent algorithms. It addresses the limited practicality of existing magnetic induction triggering technologies by automating processes, reducing the need for manual intervention, and improving the overall reliability and deployment convenience of the system.

[0115] For ease of explanation, the above description has been provided in conjunction with specific embodiments. However, the discussion in some embodiments is not intended to be exhaustive or to limit the embodiments to the specific forms disclosed above. Various modifications and variations can be obtained based on the above teachings. The selection and description of the above embodiments are for the purpose of better explaining the contents of this disclosure, thereby enabling those skilled in the art to better utilize the embodiments.

Claims

1. A high-precision magnetic target classification system based on coarse and fine levels, characterized in that, The system includes: a magnetic field data acquisition component configured to acquire magnetic field data of a target to be detected; a magnetic field data processing component configured to perform coarse and fine processing on the magnetic field data sequentially to obtain a global context-enhanced feature vector corresponding to the target to be detected; the coarse processing includes using the Grey Wolf algorithm and a physical information neural network to quickly invert the magnetic moment parameter; the fine processing includes using a spatiotemporal convolutional neural network and a multi-head self-attention mechanism to dynamically focus on key magnetic field mutations; and a target classification component configured to determine the target classification probability based on the global context-enhanced feature vector. The target classification result of the target to be detected is obtained; the magnetic field data processing component includes: a coarse processing module, which is configured to use the Grey Wolf algorithm and the physical information neural network to quickly invert the magnetic moment parameters of the magnetic field data to obtain the target positioning range and magnetic field time series data of the target to be detected; the physical constraints in the loss function of the physical information neural network include the Laplace equation and the magnetic dipole field formula; and a fine processing module, which is configured to use a spatiotemporal convolutional neural network and a multi-head self-attention mechanism to dynamically focus on key magnetic field mutations in the magnetic field time series data within the target positioning range to obtain the global context enhancement feature vector.

2. The high-precision magnetic target classification system based on coarse and fine levels according to claim 1, characterized in that, The coarse processing module includes: a Grey Wolf algorithm magnetic distance inversion unit, which has a pre-built and trained hybrid array magnetic field measurement model built into it. The Grey Wolf algorithm magnetic distance inversion unit is configured to perform a fast global inversion of the magnetic field data using the hybrid array magnetic field measurement model and the Grey Wolf algorithm to obtain the preliminary positioning range of the target location of the target to be detected; and a Physical Information Neural Network magnetic distance inversion unit, which has a pre-built and trained Physical Information Neural Network built into it. The Physical Information Neural Network magnetic distance inversion unit is configured to perform a fine inversion of the magnetic field data of the preliminary positioning range using the Physical Information Neural Network to obtain the target positioning range of the target to be detected and the magnetic field time series data.

3. The high-precision magnetic target classification system based on coarse and fine levels according to claim 2, characterized in that, The rapid global inversion includes: measuring the magnetic field data using the hybrid array magnetic field measurement model to obtain magnetic field measurement range data; inverting the magnetic field measurement range data using the Grey Wolf algorithm to obtain Grey Wolf inversion data; and calculating the position based on the Grey Wolf inversion data to obtain the preliminary positioning range of the target location of the target to be detected.

4. The high-precision magnetic target classification system based on coarse and fine levels according to claim 1, characterized in that, The fine processing module includes: a spatiotemporal convolutional feature extraction unit, which has a pre-built and trained spatiotemporal convolutional neural network built in it. The spatiotemporal convolutional feature extraction unit is configured to perform multi-layer convolution processing on the magnetic field time series data in the target positioning range to obtain a high-level feature representation; and a multi-head self-attention mechanism focusing unit, which is configured to perform multi-head attention mechanism calculation on the high-level feature representation to obtain the global context enhancement feature vector.

5. A high-precision magnetic target classification system based on coarse and fine levels according to claim 4, characterized in that, The multi-layer convolutional processing includes: extracting local spatiotemporal features from the magnetic field time series data within the target positioning range through multi-layer convolution in the spatiotemporal convolutional neural network; and outputting the local spatiotemporal features in the form of the high-level feature representation.

6. A high-precision magnetic target classification system based on coarse and fine levels according to claim 4, characterized in that, The multi-head attention mechanism calculation includes: performing parallel computation based on the high-level feature representation to generate multiple attention weight distribution maps; dynamically focusing the magnetic field time series data according to the corresponding attention weight distribution maps to obtain multiple key mutation regions; the key mutation regions correspond one-to-one with the attention weight distribution maps; and performing splicing and linear transformation processing on all the key mutation regions to obtain the global context enhancement feature vector.

7. A high-precision magnetic target classification system based on coarse and fine levels according to claim 1, characterized in that, The target classification component is further configured to: calculate the global context enhancement feature vector through a built-in fully connected layer and a Softmax classifier to obtain a probability distribution of several target categories; sort and filter according to all the probability distributions to obtain the target classification result of the target to be detected.

8. A high-precision magnetic target classification system based on coarse and fine levels according to claim 1, characterized in that, The magnetic field data acquisition component includes: a magnetic sensing module configured to acquire the magnetic signal of the target to be detected; a data acquisition module configured to convert the magnetic signal into magnetic field data; and a data transmission module configured to transmit the magnetic field data outward.

Images