Unexploded ordnance magnetic detection identification method and device, electronic equipment and storage medium

Abstract

Embodiments of the present application disclose a kind of unexploded bomb magnetic detection identification method, device, electronic equipment and storage medium, the method comprises: obtaining the magnetic vector field data of the detected area, magnetic total field data and magnetic vector gradient tensor data;Magnetic vector field data, magnetic total field data and magnetic vector gradient tensor data are channel superimposed, and magnetic data is obtained;Magnetic data is input into target neural network model, and the detection identification result about whether the target unexploded bomb exists in the detected area is obtained;Target neural network model includes convolutional neural network and multi-task regression head model;Convolutional neural network is used to extract the deep feature map of high-dimensional semantic information contained in magnetic data;Multi-task regression head model is used to obtain detection identification result according to deep feature map.By constructing multi-modal magnetic data input into target neural network model, unexploded bomb in the detected area can be accurately identified, and the magnetic detection identification precision under non-ideal detection condition is improved.

Description

Technical Field

[0001] This invention relates to the field of unexploded ordnance detection technology, and in particular to unexploded ordnance magnetic detection identification methods, devices, electronic equipment and storage media. Background Technology

[0002] Unexploded ordnance (UXO) refers to unexploded ordnance, such as grenades, rockets, bombs, and landmines, left behind during wartime or military exercises. Due to their explosive nature, they pose a serious threat to the lives and property of people in the surrounding area, their daily lives, and military exercises. Magnetic detection methods, with their advantages of being lightweight, non-invasive, not actively triggering targets, and highly secure, are commonly used in the field of UXO detection. However, existing magnetic detection methods are mainly based on the magnetic dipole theory model, and their overall detection accuracy decreases significantly under non-ideal detection conditions. Summary of the Invention

[0003] The purpose of this invention is to provide a method, device, electronic device, and storage medium for identifying unexploded ordnance using magnetic detection, in order to solve the problem of low overall detection accuracy under non-ideal detection conditions in the prior art.

[0004] To solve the above-mentioned technical problems, the embodiments of the present invention are implemented as follows: In a first aspect, embodiments of the present invention provide a method for magnetic detection identification of unexploded ordnance, the method comprising: Acquire magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data of the probed region; Magnetic data is obtained by overlaying the magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data into channels. The magnetic data is input into the target neural network model to obtain the detection and identification results of whether there is an unexploded ordnance in the detected area; The target neural network model includes a convolutional neural network and a multi-task regression head model; the convolutional neural network is used to extract deep feature maps containing high-dimensional semantic information from the magnetic data; the multi-task regression head model is used to obtain the detection and recognition results based on the deep feature maps.

[0005] In one implementation, the multi-task regression head model includes a fully connected regression head branch model and a fully convolutional network regression head branch model; the detection and identification results include detection classification results and detection localization results; the magnetic data is input into the target neural network model to obtain detection and identification results regarding the presence of unexploded ordnance in the detected area, including: Deep feature maps containing high-dimensional semantic information in magnetic data are extracted using convolutional neural networks; The deep feature maps are input into the fully connected regression head branch model and the fully convolutional network regression head branch model, respectively. The detection classification result is obtained through the fully connected regression head branch model; the detection localization result is obtained through the fully convolutional network regression head branch model. The detection classification result represents the category of unexploded ordnance in the detected area; the detection localization result represents the location of the unexploded ordnance in the detected area.

[0006] In one implementation, the detection and localization results are obtained through the regression head branch model of the fully convolutional network, including: The first intermediate feature is obtained by reducing the dimensionality of the deep feature map through the first convolutional layer and the first activation function. The first intermediate feature is mapped by the second convolutional layer to obtain the first response map; The detection and localization results are obtained by aggregating the first response map using a pooling layer.

[0007] In one implementation, the detection classification results are obtained through a fully connected regression head-branch model, including: Flatten the deep feature map to obtain the second intermediate feature; The second intermediate feature is used as input and processed sequentially through the first fully connected layer, the second activation function, the Dropout random deactivation layer, the second fully connected layer, and the Softmax function to obtain the detection and classification results.

[0008] In one implementation, multiple convolutional neural networks are used, each including a third convolutional layer, a third activation function, an attention mechanism, and a first pooling layer. Magnetic data is input into the target neural network model to obtain the detection and identification results of the unexploded ordnance, including: The magnetic data is input into the target neural network model, which then performs the following actions: The magnetic data is convolved by the third convolutional layer to obtain the first feature map; The negative input in the first feature map is non-linearly transformed by the third activation function to obtain the second feature map; The first attention map corresponding to the second feature map is determined by an attention mechanism, and the second feature map and the first attention map are fused to obtain the third feature map; The third feature map is pooled using the first pooling layer to obtain the deep feature map.

[0009] In one implementation, the first attention map includes a channel attention map and a spatial attention map, and the attention mechanism includes a channel attention mechanism and a spatial attention mechanism; the first attention map corresponding to the second feature map is determined through the attention mechanism, and the second feature map and the first attention map are fused to obtain the third feature map, including: The second feature map is used as input, and the channel attention map and spatial attention map are obtained through the channel attention mechanism and spatial attention mechanism, respectively. The third feature map is obtained by multiplying the second feature map, the channel attention map, and the spatial attention map element by element.

[0010] In one implementation, acquiring magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data includes: The magnetic vector field data of the probed area is obtained by using a planar square fluxgate detector array; The total magnetic field data is determined based on the magnetic field values ​​of each fluxgate in the magnetic vector field data along the first axis, second axis, and third axis. Based on the side length of the planar square fluxgate detector array and the magnetic field values ​​of each fluxgate along the first, second, and third axes in the magnetic vector field data, the magnetic vector gradient tensor data are determined using the finite difference method.

[0011] Secondly, embodiments of the present invention provide a magnetic detection device for unexploded ordnance, the device comprising: The acquisition module is used to acquire magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data of the probed area; The first processing module is used to perform channel superposition of magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data to obtain magnetic data. The identification module is used to input magnetic data into the target neural network model to obtain the detection and identification results regarding the presence of unexploded ordnance in the detected area. The target neural network model includes a convolutional neural network and a multi-task regression head model. The convolutional neural network is used to extract deep feature maps containing high-dimensional semantic information from the magnetic data. The multi-task regression head model is used to obtain the detection and identification results based on the deep feature maps.

[0012] In one implementation, the multi-task regression head model includes a fully connected regression head branch model and a fully convolutional network regression head branch model; the detection and recognition results include detection classification results and detection localization results; the recognition module is also used for: Deep feature maps containing high-dimensional semantic information in magnetic data are extracted using convolutional neural networks; The deep feature maps are input into the fully connected regression head branch model and the fully convolutional network regression head branch model, respectively. The detection classification result is obtained through the fully connected regression head branch model; the detection localization result is obtained through the fully convolutional network regression head branch model. The detection classification result represents the category of unexploded ordnance in the detected area; the detection localization result represents the location of the unexploded ordnance in the detected area.

[0013] In one implementation, the identification module is also used for: The first intermediate feature is obtained by reducing the dimensionality of the deep feature map through the first convolutional layer and the first activation function. The first intermediate feature is mapped by the second convolutional layer to obtain the first response map; The detection and localization results are obtained by aggregating the first response map using a pooling layer.

[0014] In one implementation, the identification module is also used for: Flatten the deep feature map to obtain the second intermediate feature; The second intermediate feature is used as input and processed sequentially through the first fully connected layer, the second activation function, the Dropout random deactivation layer, the second fully connected layer, and the Softmax function to obtain the detection and classification results.

[0015] In one implementation, there are multiple convolutional neural networks, each including a third convolutional layer, a third activation function, an attention mechanism, and a first pooling layer; the recognition module is also used for: The magnetic data is convolved by the third convolutional layer to obtain the first feature map; The negative input in the first feature map is non-linearly transformed by the third activation function to obtain the second feature map; The first attention map corresponding to the second feature map is determined by an attention mechanism, and the second feature map and the first attention map are fused to obtain the third feature map; The third feature map is pooled using the first pooling layer to obtain the deep feature map.

[0016] In one implementation, the first attention map includes a channel attention map and a spatial attention map, and the attention mechanism includes a channel attention mechanism and a spatial attention mechanism; the recognition module is also used for: The second feature map is used as input, and the channel attention map and spatial attention map are obtained through the channel attention mechanism and spatial attention mechanism, respectively. The third feature map is obtained by multiplying the second feature map, the channel attention map, and the spatial attention map element by element.

[0017] In one implementation, the acquisition module includes: a first acquisition module, a first determination module, and a second determination module; The first acquisition module is used to acquire magnetic vector field data of the probed area through a planar square fluxgate detector array; The first determining module is used to determine the total magnetic field data based on the magnetic field values ​​of each fluxgate in the magnetic vector field data along the first axis, the second axis, and the third axis. The second determining module is used to determine the magnetic vector gradient tensor data by means of the finite difference method based on the side length of the planar square fluxgate detector array and the magnetic field value of each fluxgate along the first axis, the second axis and the third axis in the magnetic vector field data.

[0018] Thirdly, embodiments of the present invention provide an electronic device for identifying unexploded ordnance magnetic probes, including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the unexploded ordnance magnetic probe identification method provided in the above embodiments.

[0019] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the unexploded ordnance magnetic detection method provided in the above embodiments.

[0020] The above-described at least one technical solution adopted in the embodiments of the present invention can achieve the following beneficial effects: The unexploded ordnance magnetic detection identification method provided in this invention includes acquiring magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data of the area to be detected; performing channel superposition on the magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data to obtain magnetic data; inputting the magnetic data into a target neural network model to obtain a detection and identification result regarding the presence of a target unexploded ordnance in the area to be detected; the target neural network model includes a convolutional neural network and a multi-task regression head model; the convolutional neural network is used to extract deep feature maps containing high-dimensional semantic information in the magnetic data; the multi-task regression head model is used to obtain the detection and identification result based on the deep feature maps. This method inputs multimodal magnetic data, including magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data, into a target neural network model composed of a convolutional neural network and a multi-task regression head model. The convolutional neural network extracts the high-dimensional semantic information contained in the multimodal magnetic data and generates a deep feature map, effectively capturing the characteristics of weak magnetic anomaly amplitude, strong spatial diffusion, and significant local gradients. Subsequently, the deep feature map is input into the multi-task regression head model to accurately identify unexploded ordnance in the detected area, thereby improving the magnetic detection accuracy under non-ideal detection conditions. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1A flowchart illustrating an unexploded ordnance magnetic detection identification method provided in this embodiment of the invention. Figure 1 ; Figure 2 This is a schematic diagram of the planar square fluxgate detector array for an unexploded ordnance magnetic detection method provided in an embodiment of the present invention; Figure 3 This is a two-dimensional distribution diagram of magnetic data for an unexploded ordnance magnetic detection identification method provided in an embodiment of the present invention; Figure 4 A schematic diagram of a convolutional neural network provided in an embodiment of the present invention. Figure 1 ; Figure 5 A schematic diagram of a convolutional neural network provided in an embodiment of the present invention. Figure 2 ; Figure 6 A comparison diagram of the output values ​​of different activation functions of a convolutional neural network provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of a hybrid attention mechanism provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of a multi-task regression head model provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of a fully convolutional network regression head branch model provided in an embodiment of the present invention; Figure 10 A schematic diagram illustrating the training process of a target neural network model for an unexploded ordnance magnetic detection method provided in an embodiment of the present invention; Figure 11 The geometric model of the unexploded ordnance target provided in the embodiments of the present invention; Figure 12 This is a two-dimensional distribution diagram of nine channels of projectile target data provided in an embodiment of the present invention; Figure 13 Magnetic data of a projectile target in an actual experiment provided for an embodiment of the present invention; Figure 14 Magnetic moment variation diagrams of five unexploded ordnance targets under different attitudes provided in embodiments of the present invention; Figure 15 A two-dimensional distribution diagram of the nine channels of data of the interference target provided in an embodiment of the present invention; Figure 16 The fluxgate sensor provided in this embodiment of the invention collects the local environmental magnetic field total field time-domain signal; Figure 17 The probability density function and cumulative distribution function curves of the environmental magnetic field provided in the embodiments of the present invention; Figure 18This is a schematic diagram of the hardware structure of the magnetic detection device for unexploded ordnance identification provided in an embodiment of the present invention; Figure 19 This is a schematic diagram of the hardware structure of the unexploded ordnance magnetic detection electronic device provided in an embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0024] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0025] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of the invention described herein can be implemented in an order other than that illustrated or described herein.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit this disclosure.

[0027] To better understand the embodiments of the present invention, the following descriptions are provided through some scenario examples: Example 1 like Figure 1 As shown, this embodiment of the invention provides a method for identifying unexploded ordnance using magnetic detection, the method comprising: Step S101: Obtain magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data of the probed area; Specifically, a planar square fluxgate detector array with a side length of 0.8m is used to acquire magnetic vector field data of the probed area. The center of the probed area (e.g., a 10m×10m two-dimensional grid) is aligned with the geometric center of the array, and sampling points are arranged in a grid pattern using the midpoint of the array as a reference to achieve full area coverage. To obtain high-resolution data, the grid spacing of the sampling plane is set to 0.01m, meaning that within the set 10m×10m probed area, the system can densely acquire magnetic field data with centimeter-level spatial resolution, forming a high-density grid of 1001×1001.

[0028] In one implementation, acquiring magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data includes: The magnetic vector field data of the probed area is obtained by using a planar square fluxgate detector array; The total magnetic field data is determined based on the magnetic field values ​​of each fluxgate in the magnetic vector field data along the first axis, second axis, and third axis. Based on the side length of the planar square fluxgate detector array and the magnetic field values ​​of each fluxgate along the first, second, and third axes in the magnetic vector field data, the magnetic vector gradient tensor data are determined using the finite difference method.

[0029] like Figure 2 As shown, the array contains four fluxgate sensors (numbered S1-S4). These four fluxgate sensors synchronously acquire magnetic field vector components (i.e., magnetic field values) along the first, second, and third axes, constituting the magnetic vector field data of the probed area. The first, second, and third axes correspond to... Figure 2 In Axial direction, Axial direction and Axial direction. Accordingly, For along Magnetic field vector components along the axial direction, For along Magnetic field vector components along the axial direction, For along The magnetic field vector component along the axial direction.

[0030] In this implementation, the total magnetic field data can be determined based on the magnetic field values ​​of each fluxgate in the magnetic vector field data along the first axis, the second axis, and the third axis; and the magnetic vector gradient tensor data can be determined by the finite difference method based on the side length of the planar square fluxgate detector array and the magnetic field values ​​of each fluxgate in the magnetic vector field data along the first axis, the second axis, and the third axis.

[0031] Specifically, the total magnetic field at the center of the array is calculated using the following formula 1 to obtain the total magnetic field data.

[0032]

[0033] in, For the total magnetic field, , , For the first A fluxgate edge , , The magnetic field value measured along the axial direction.

[0034] The magnetic vector gradient at the center of the array can be calculated using the finite difference method, yielding the magnetic vector gradient tensor data. Considering the passive and irrotational nature of the magnetic field, the magnetic gradient tensor contains only five independent variables, calculated according to the following formula 2:

[0035] in, Represents the magnetic field vector components Rate of change along the x-axis Represents the magnetic field vector components Rate of change along the x-axis Represents the magnetic field vector components Rate of change along the x-axis Represents the magnetic field vector components Rate of change along the z-axis, Represents the magnetic field vector components Rate of change along the y-axis m is the side length of the detection array.

[0036] Step S102: Channel superposition of magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data to obtain magnetic data; Specifically, the magnetic vector field data, total magnetic field data and magnetic vector gradient tensor data obtained from the probed area are superimposed to obtain magnetic data, totaling 9 data channels.

[0037] like Figure 3 The image shows a two-dimensional distribution diagram of the nine channels of the KAB1500 target obtained from finite element simulation. The magnetic data used for detection and identification consists of magnetic vector field, total magnetic field, and magnetic vector gradient tensor data superimposed on a 501×501 (5m×5m) two-dimensional grid, including: total magnetic field data. , axial magnetic field vector components , axial magnetic field vector components , axial magnetic field vector components Magnetic field vector components along Rate of change of axis, magnetic field vector components along Rate of change of axis, magnetic field vector components along Rate of change of axis, magnetic field vector components along Rate of change of axis, magnetic field vector components along Rate of change of the axis.

[0038] Step S103: Input the magnetic data into the target neural network model to obtain the detection and identification results of whether there is an unexploded ordnance in the detected area; The target neural network model includes a convolutional neural network and a multi-task regression head model; the convolutional neural network is used to extract deep feature maps containing high-dimensional semantic information in the magnetic data; the multi-task regression head model is used to obtain the detection and recognition results based on the deep feature maps.

[0039] In one implementation, multiple convolutional neural networks are used, each including a third convolutional layer, a third activation function, an attention mechanism, and a first pooling layer. Magnetic data is input into the target neural network model to obtain the detection and identification results of the unexploded ordnance, including: The magnetic data is input into the target neural network model, which then performs the following actions: The magnetic data is convolved by the third convolutional layer to obtain the first feature map; The negative input in the first feature map is non-linearly transformed by the third activation function to obtain the second feature map; The first attention map corresponding to the second feature map is determined by an attention mechanism, and the second feature map and the first attention map are fused to obtain the third feature map; The third feature map is pooled using the first pooling layer to obtain the deep feature map.

[0040] In this implementation, Convolutional Neural Networks (CNNs) have achieved significant results in computer vision tasks such as image classification and object recognition. Deep structures, represented by the VGG series networks, are widely used due to their regular convolutional framework and efficient feature representation capabilities. Traditional VGG networks achieve layer-by-layer abstraction of image texture, edge, and shape features by stacking numerous 3×3 convolutional and pooling layers, primarily focusing on static texture analysis of natural images. However, in magnetic anomaly detection scenarios, the input data is no longer image texture but a continuous field distribution formed by magnetic gradient tensors. Magnetic anomaly signals are characterized by weak amplitude, strong spatial diffusion, and significant local gradients, which differ significantly from the significant edge characteristics of traditional images. If the original VGG network structure is directly used for feature extraction, its highly adapted structure to natural image characteristics may not be able to fully capture the local perturbation features of magnetic anomalies in multi-scale space. Based on these differences, this embodiment of the invention makes targeted modifications and optimizations to the feature extraction backbone based on the VGG structure.

[0041] Specifically, such as Figure 4 , Figure 5 As shown, this embodiment of the invention uses five convolutional neural networks connected in series. Each convolutional neural network serves as a feature extraction module, performing convolution calculations on the input magnetic data. Each feature extraction module includes a third convolutional layer (3×3 convolution), a third activation function (ELU activation function), an attention mechanism, and a first pooling layer (2×2 max pooling). Each feature extraction module maintains the standard 3×3 convolutional structure of VGG while introducing an Exponential Linear Unit (ELU) activation function and an attention mechanism to adapt to the statistical characteristics and physical distribution patterns of anomalous magnetic data. Using the constructed 9-channel magnetic data as input, after multi-layer convolution calculations by multiple convolutional neural networks, the final output is a deep feature map containing high-dimensional semantic information.

[0042] In one implementation, the first attention map includes a channel attention map and a spatial attention map, and the attention mechanism includes a channel attention mechanism and a spatial attention mechanism; the first attention map corresponding to the second feature map is determined through the attention mechanism, and the second feature map and the first attention map are fused to obtain the third feature map, including: The second feature map is used as input, and the channel attention map and spatial attention map are obtained through the channel attention mechanism and spatial attention mechanism, respectively. The third feature map is obtained by multiplying the second feature map, the channel attention map, and the spatial attention map element by element.

[0043] Specifically, within each feature extraction module, the input magnetic data is first processed through a 3×3 third convolutional layer to extract local region features, resulting in a first feature map. The ELU activation function provides a smooth negative half-axis nonlinear mapping on top of the first feature map output by the convolution, yielding a second feature map. Compared to ReLU, the mean of its output second feature map is closer to zero, mitigating gradient shift and improving network convergence stability, which is crucial for effectively capturing weak magnetic anomalies. Subsequently, to overcome the limitations of the local receptive field in convolutional operations and enhance the model's response to perturbations at different spatial scales of magnetic anomalies, a hybrid attention mechanism is added to the module to extract the channel attention map and spatial attention map corresponding to the second feature map, serving as the first attention map. This achieves dual recalibration of the channel and spatial dimensions of the second feature map, re-obtaining the third feature map. Finally, a 2×2 max-pooling layer with a stride of 2×2 is used to downsample the third feature map obtained by fusing the first attention map, further refining deep features and enabling the network to gradually abstract from shallow local gradients to the deep overall morphology of the magnetic anomaly.

[0044] (1) Third convolutional layer (3×3 convolution operation) The third convolutional layer, as the core module of the Convolutional Neural Network (CNN), primarily undertakes the task of extracting spatial features from the input data. This layer contains N learnable convolutional kernels, where N directly determines the number of channels in the output feature map after convolution. The convolution operation utilizes the principle of local correlation in images. The convolutional kernel performs a sliding window scan on the input signal with a set stride, executing a weighted summation operation, thereby gradually abstracting low-level local features into high-level semantic features. This mechanism not only achieves local perception of the input data but also significantly reduces the number of parameters and computational complexity of the model through a "weight sharing" strategy (i.e., the same convolutional kernel shares parameters at different locations in the image). The specific mathematical expression of the convolution operation is shown in Equation 3:

[0045] in, To output the first feature map in The value of the position, express Convolution operation, For the input magnetic data, This represents the stride of the convolutional kernel as it moves across the input feature map. For fill size, The matrix represents a size of convolution kernel, For input magnetic data in the region covered by the convolution kernel The corresponding location feature value.

[0046] (2) Third activation function (ELU activation function) The third activation function is a key component in convolutional neural networks (CNNs) that introduces nonlinear characteristics. Its main function is to convert the input signals of neurons into output signals, thereby giving the network the ability to fit complex nonlinear functions. Without an activation function, no matter how deep the neural network is, it is essentially just a superposition of linear transformations and cannot handle high-dimensional nonlinear separable problems. After performing a linear weighted summation operation in the third convolutional layer, a nonlinear activation layer is usually followed to enhance the richness of feature representation and the discriminative power of the model. This invention primarily uses the classic sigmoid function and the improved exponential linear unit (ELU) function.

[0047] The Sigmoid function is one of the most commonly used activation functions in early neural networks, and its mathematical expression is shown in Equation 4:

[0048] in, This is the second feature map after the activation function. This is the first feature map output by the third convolutional layer. This is the Sigmoid function.

[0049] The sigmoid function exhibits a monotonically increasing "S"-shaped curve characteristic, capable of compressing and mapping inputs from any real number field to an open interval (0,1), thus it is often used in probabilistic outputs or gating mechanisms. However, in deep network training, the sigmoid function has significant limitations: when the absolute value of the input signal is large, the function enters the saturation region, and its derivative approaches zero. According to the chain rule, this will lead to gradient vanishing during backpropagation, hindering the update of deep parameters. Furthermore, the output of the sigmoid function is not centered at zero, which causes the weight update path to oscillate in a sawtooth pattern, reducing optimization efficiency.

[0050] To overcome the shortcomings of traditional activation functions and improve model training efficiency, this embodiment of the invention employs the Exponential Linear Unit (ELU) as the core activation function. Unlike the Corrected Linear Unit (ReLU), which directly truncates negative inputs to zero, the ELU function introduces a smooth exponential decay along the negative half-axis, as shown in Equation 5:

[0051] in, This is the second feature map after the activation function. This is the first feature map output by the third convolutional layer. This is an adjustable parameter, typically set to 1. It is the activation function for the exponential linear unit.

[0052] like Figure 6 As shown, when When ELU maintains a linear identity mapping, it effectively avoids gradient saturation in the positive interval; when In ELU, negative outputs are generated using an exponential function. This design makes the mean of activation values ​​closer to zero, thus producing a self-normalization effect similar to batch normalization. This feature not only avoids the "neuron death" problem caused by ReLU's derivative being zero in the negative region, but also significantly suppresses gradient oscillations, accelerates the convergence speed of the model during training, and enhances the network's robustness to input noise.

[0053] (3) Hybrid attention mechanism Hybrid attention mechanisms consist of channel attention and spatial attention mechanisms. For example... Figure 7 As shown, after the second feature map passes through the channel attention mechanism and the spatial attention mechanism, a spatial attention map and a channel attention map are generated respectively. After element-wise multiplication with the second feature map, a global adaptive weight calibration feature map (i.e., the third feature map) is obtained.

[0054] Specifically, the Squeeze and Excitation (SE) method is used to calculate the channel attention weights. The calculation method is shown in Formula 6:

[0055] in, This is a channel attention map. This is the second feature map after the activation function. For global average pooling, It is a multilayer perceptron. It is the Sigmoid activation function. This is a channel attention mechanism.

[0056] First, the features of each channel of the upper-layer input are compressed into a scalar by global average pooling. Then, a multilayer perceptron (MLP) is used to represent the interrelationships between channels, and non-negative weight coefficients are generated by the sigmoid activation function.

[0057] Spatial attention weights are calculated using the following formula 7. :

[0058] in, This is a spatial attention map. Average pooling / max pooling at the channel level It is the Sigmoid activation function. For spatial attention mechanisms, This is the second feature map after the activation function. This represents a 7×7 convolution operation.

[0059] First, max pooling and average pooling are performed on the upper input along the channel dimension. Then, a large-size (7×7) convolutional kernel is used to calculate the concatenated max pooling and average pooling feature maps, and non-negative weight coefficients are generated by the Sigmoid activation function.

[0060]

[0061] in, This is the third feature map. This is the second feature map. For hybrid attention functions, Spatial attention map This is a channel attention map. This is for element-wise multiplication.

[0062] The channel attention mechanism in this embodiment of the invention enables adaptive selection of features in the channel dimension, while the spatial attention mechanism focuses on local regions with significant magnetic anomaly perturbations in the feature map. The coupling of the two enables the network to simultaneously pay attention to the global magnetic field perturbation trend and local gradient changes, significantly improving the robustness and effectiveness of magnetic anomaly feature representation.

[0063] (4) First pooling layer (2×2 max pooling) The primary function of the first pooling layer is to downsample the feature map, preserving key information while creating conditions for extracting deeper features at higher levels. Commonly used pooling methods include average pooling, max pooling, overlapping pooling, Gaussian pooling, and random pooling. This embodiment uses max pooling with a pooling window size of 2×2 and a stride of 2 as an example. The feature map expression after pooling is shown in Formula 9 below:

[0064] in, This represents the value of the pooled deep feature map at position (m,n). The result of max pooling of the feature image. For max pooling, This is the third feature map. These are the feature values ​​of the third feature map at four adjacent positions within the pooling window. Let M be the value of the third feature map at position (m,n), where M and N are the feature map dimensions.

[0065] Through the above structural optimization, the improved feature extraction module not only maintains the simple structure and deep expression capability of the VGG network, but also adapts to the physical characteristics of magnetic anomaly data. This enables the backbone network to effectively extract discriminative deep features from magnetic field data containing interference signals, providing a reliable data representation basis for subsequent classification, identification and location calculation of unexploded ordnance.

[0066] In one implementation, the multi-task regression head model includes a fully connected regression head branch model and a fully convolutional network regression head branch model; the detection and identification results include detection classification results and detection localization results; the magnetic data is input into the target neural network model to obtain detection and identification results regarding the presence of unexploded ordnance in the detected area, including: Deep feature maps containing high-dimensional semantic information in magnetic data are extracted using convolutional neural networks; The deep feature maps are input into the fully connected regression head branch model and the fully convolutional network regression head branch model, respectively. The detection classification result is obtained through the fully connected regression head branch model; the detection localization result is obtained through the fully convolutional network regression head branch model. The detection classification result represents the category of unexploded ordnance in the detected area; the detection localization result represents the location of the unexploded ordnance in the detected area.

[0067] Specifically, such as Figure 8 As shown, in order to simultaneously achieve type identification and localization calculation of the detected target, this embodiment of the invention designs a multi-task regression head model architecture for classification and localization, respectively. This architecture consists of a fully connected regression head branch model and a fully convolutional network regression head branch model. The two regression modules are connected in parallel. Figure 4 The deep feature maps output by the convolutional neural network shown are used to improve the overall task performance through joint optimization.

[0068] In one implementation, the detection classification results are obtained through a fully connected regression head-branch model, including: Flatten the deep feature map to obtain the second intermediate feature; The second intermediate feature is used as input and processed sequentially through the first fully connected layer, the second activation function, the Dropout random deactivation layer, the second fully connected layer, and the Softmax function to obtain the detection and classification results.

[0069] Specifically, unexploded ordnance type identification is essentially a multi-classification problem, aiming to extract discriminative patterns of different ordnance types from magnetic anomaly features. To this end, the fully connected regression head branch model employs a multilayer perceptron (MLP) structure. First, the depth feature map of size [15, 15, 512] output by the convolutional neural network is flattened and reshaped into a one-dimensional feature vector, obtaining the second intermediate feature. This process transforms spatial distribution information into a global semantic feature representation. Subsequently, this second intermediate feature is input into a first fully connected layer containing one hidden layer. To enhance the nonlinear expressive power of the structure and prevent overfitting, a second activation function based on the ELU activation function and a Dropout random deactivation layer are configured after the first fully connected layer. The number of output nodes of the final second fully connected layer is set to the total number of unexploded ordnance categories to be classified. The output values ​​are normalized into a probability distribution using the Softmax function, thereby determining the type of unexploded ordnance in the detected area.

[0070] In one implementation, the detection and localization results are obtained through the regression head branch model of the fully convolutional network, including: The first intermediate feature is obtained by reducing the dimensionality of the deep feature map through the first convolutional layer and the first activation function. The first intermediate feature is mapped by the second convolutional layer to obtain the first response map; The detection and localization results are obtained by aggregating the first response map using a pooling layer.

[0071] Specifically, in traditional convolutional neural networks, the final step in a regression task typically involves forcibly converting the two-dimensional feature map extracted through deep convolution into a one-dimensional vector through a "flattening" operation before feeding it into a fully connected layer for coordinate mapping. However, for magnetic anomaly localization, this design has a fundamental structural flaw: The essence of a magnetic anomaly map is a spatial distribution map of a physical field, where the position of each pixel corresponds one-to-one with spatial coordinates in the real world. Spatial features such as peak positions and gradient directions in the map are core information for locating the source of magnetic anomalies. The flattening operation crudely "straightens" the two-dimensional spatial grid into a sequence, completely destroying the spatial topology between pixels. This forces subsequent fully connected layers to learn from this scrambled information and reconstruct the existing spatial relationships, which is not only extremely inefficient but also causes irreversible loss of crucial information. To overcome these limitations, such as... Figure 9As shown, this paper employs a fully convolutional regression head. Its core idea is to replace fully connected layers with 1x1 convolutional layers, thereby achieving end-to-end mapping from high-dimensional features to target coordinates while preserving spatial information. A 1x1 convolutional layer transforms the multidimensional feature vector of each pixel into a new multidimensional feature vector through linear combination and non-linear activation. This design achieves both inter-channel information fusion and dimensionality reduction, while perfectly preserving the spatial coordinates of each pixel, laying the foundation for subsequent location prediction.

[0072] The fully convolutional network regression head branch model employs a staged mapping strategy to achieve more robust feature transformation: (1) Feature dimensionality reduction and integration After receiving the high-dimensional feature map extracted by the convolutional neural network, the dimensionality of the high-dimensional feature map extracted by the convolutional neural network is reduced by the first convolutional layer (1x1 convolutional layer) and the first activation function (ELU activation function), resulting in the first intermediate feature. While reducing the number of model parameters, the highly abstract semantic features are integrated and refined into a set of more compact intermediate features that are more focused on the regression task.

[0073] (2) Coordinate dimension mapping The second convolutional layer (1x1 convolutional layer) directly maps the first intermediate features to the final coordinate dimension. To allow the network to freely regress coordinate values ​​within any real number range, the output of this layer does not pass through an activation function, resulting in a first response map of shape [C, H, W]. Each pixel value on this first response map can be understood as the network's contribution to the final coordinates based on its local receptive field. Finally, global average pooling of the pooling layer integrates the spatial dimension information of the entire map, that is, the [C, H, W] first response map is aggregated into a [C]-dimensional vector, which is the globally optimal coordinate prediction value output by the model, i.e., the position of the unexploded ordnance in the detected area. The calculation formula of the regression head branch model of the fully convolutional network is shown in Equation 10 below:

[0074] in, For the detection and positioning results, Deep feature maps output by convolutional neural networks. For a 1×1 convolution operation, For ELU activation function, For global average pooling, This is the mapping function for the fully convolutional regression head.

[0075] Table 1 shows the network structure framework of the target neural network model in this embodiment of the invention. Specifically, it includes the operation flow, channel changes, and output resolution of multiple convolutional neural networks, the multi-task regression head model including the fully connected regression head branch model and the fully convolutional network regression head branch model.

[0076] Table 1: Network Structure Framework

[0077] The unexploded ordnance magnetic detection identification method provided in this invention includes acquiring magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data of the area to be detected; performing channel superposition on the magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data to obtain magnetic data; inputting the magnetic data into a target neural network model to obtain a detection and identification result regarding the presence of a target unexploded ordnance in the area to be detected; the target neural network model includes a convolutional neural network and a multi-task regression head model; the convolutional neural network is used to extract deep feature maps containing high-dimensional semantic information in the magnetic data; the multi-task regression head model is used to obtain the detection and identification result based on the deep feature maps. This method inputs multimodal magnetic data, including magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data, into a target neural network model composed of a convolutional neural network and a multi-task regression head model. The convolutional neural network extracts the high-dimensional semantic information contained in the multimodal magnetic data and generates a deep feature map, effectively capturing the characteristics of weak magnetic anomaly amplitude, strong spatial diffusion, and significant local gradients. Subsequently, the deep feature map is input into the multi-task regression head model to accurately identify unexploded ordnance in the detected area, thereby improving the magnetic detection accuracy under non-ideal detection conditions.

[0078] Example 2 like Figure 10 As shown, this embodiment of the invention provides a method for training a target neural network model for unexploded ordnance magnetic detection identification, the method comprising: Step S201: Construct a simulation database that includes a library of unexploded ordnance forward models and a disturbance database.

[0079] The performance of neural network models largely depends on the scale and quality of the training data. In the field of unexploded ordnance magnetic anomaly detection, acquiring large-scale experimental data often requires extremely high costs, and experimental samples are often insufficient to cover complex and variable burial attitudes and environmental conditions. To address this bottleneck of data scarcity, a high-fidelity physical simulation environment was constructed and a complete training database was generated, providing necessary data support for building a neural network-based method for unexploded ordnance magnetic anomaly detection.

[0080] This invention primarily focuses on the forward modeling simulation and database construction of magnetic anomalies in unexploded ordnance (UFOs). Five typical UFO forward model libraries are constructed based on finite element method (FEM) simulation software. To minimize the discrepancy between simulation data and real-world operating environments, and to implement anti-interference and data augmentation techniques under complex backgrounds, an interference database is established to simulate common non-target magnetic anomalies in real-world operating environments. Simultaneously, based on statistical characteristic analysis of environmental magnetic field noise collected in the field, noise injection and data augmentation processing conforming to real-world distribution patterns are implemented. The specific implementation method is as follows: Step S2011: Construct a forward model library for unexploded ordnance.

[0081] (1) Constructing the geometric model and simulation conditions of unexploded ordnance To simulate the spatial distribution characteristics of UXOs on the surface and underground in actual operating environments, and to provide high-precision input data for subsequent magnetic anomaly signal simulation and neural network training, a magnetic anomaly database needs to be established for common UXOs in the test range. The UXO targets included in the database are shown in Table 2 below. This includes: 3 typical runway bomb targets, 1 howitzer target, and 1 landmine target.

[0082] Table 2: UXO Target List

[0083] Specifically, this invention is based on Maxwell's equations and uses finite element simulation software to perform three-dimensional magnetic field simulation of the UXO target under conditions without external current.

[0084] First, digital twin models of the targets listed in Table 2 were constructed as geometric models of unexploded ordnance, such as... Figure 11 As shown. Common projectile bodies use cast iron, 45 steel, and aluminum as the main materials. In this simulation, the material of the UXO target is set to cast iron, and its relative permeability is set to 300.

[0085] A coordinate system was established using COMSOL finite element simulation software, with the y-axis pointing north, the x-axis pointing east, and the z-axis perpendicular to the horizontal plane and pointing upwards. Based on Nanjing's local latitude and longitude, the simulated background magnetic field was set according to the IGRF international geomagnetic standard model. The total geomagnetic field is approximately 50,000 nT, with a magnetic elevation angle (I) of approximately 45° and a magnetic declination angle (D) of approximately -5.5°. The three-axis components of the geomagnetic field were calculated.

[0086] (2) Mesh generation and finite element solution configuration Considering the local fine features of the magnetic anomaly field of the unexploded ordnance in the near field region and the attenuation features in the far field region, the simulation environment domain is set as a cuboid space with a length of 10m, a width of 10m, and a height of 30m, and the geometric model of the unexploded ordnance is placed at the geometric center of this region.

[0087] The quality of mesh generation directly determines the convergence and accuracy of finite element numerical calculations. Given that unexploded ordnance targets typically possess irregular geometries and high magnetic permeability, conventional adaptive mesh generation often fails to achieve ideal discretization in regions with drastically changing magnetic field gradients. Therefore, globally uniform or automatically adaptive meshing methods are abandoned in favor of a more targeted non-uniform manual mesh generation method. The following principles are followed when manually generating the mesh: Principle 1: For regions in the projectile with large variations in magnetic field strength or complex geometry, use extremely fine mesh units for discretization to accurately capture the details of local weak magnetic field distortion and reduce discretization errors. Principle 2: In regions with gentle magnetic field changes, such as the background air region far from the target, the grid cell size should be appropriately increased to concentrate computational resources on key areas. This significantly reduces the total number of degrees of freedom of the model while ensuring solution accuracy, thereby improving the overall simulation efficiency. Principle 3: At the boundary between the dense and sparse areas of the mesh, construct a mesh transition layer to ensure that the mesh size has a smooth and gradual change trend.

[0088] (3) Establish a forward model library for unexploded ordnance. Taking shell 1 in Table 2 as an example, after completing the configuration of the unexploded ordnance geometric model and finite element solution environment, the magnetic anomaly distribution around the unexploded ordnance target can be calculated. In order to construct a forward modeling database, for Figure 2 The array structure shown has four sampling points centered on the array's midpoint to represent data acquisition at the current location. Data sampling across the entire observation plane is completed using a grid-like arrangement. To obtain high-resolution data, the grid spacing of the sampling plane is set to 0.01m, meaning that magnetic field data is densely acquired with centimeter-level spatial resolution within a defined 10m × 10m detection area, forming a high-density grid of 1001 × 1001. The total magnetic field data and magnetic vector gradient tensor data at the array center can be calculated using Equations 1 and 2. Finally, the independent components of the total magnetic field data, magnetic vector field data, and magnetic vector gradient tensor data generated by the UXO target are obtained, totaling data from nine physical channels. Figure 12 The figure shown is a two-dimensional distribution diagram of the nine channels of the projectile target data obtained from finite element simulation.

[0089] From the overall distribution pattern, the simulation results clearly reflect the magnetization effect of ferromagnetic targets under the excitation of the geomagnetic background field, and each physical component exhibits significant and regular spatial geometric characteristics. The total magnetic field and the vertical vector component ( The data exhibits significant positive and negative extreme value regions. These extreme value regions are aligned along the geomagnetic direction, and their amplitude decreases with increasing distance, visually reflecting the approximate location of the target. The horizontal vector component ( The magnetic gradient tensor exhibits a centrally symmetric polar distribution. Finally, compared to the vector field, due to the fourth-power decay of the magnetic gradient tensor with distance, its anomalous energy is highly concentrated near the edge contour of the unexploded ordnance target. , , A maximum and a minimum value are generated on the observation plane, and the unexploded ordnance target is between the maximum and the minimum value; Three extreme values, one minimum value, and two maximum values ​​are generated on the observation plane, with the unexploded ordnance target located at the minimum value position; It produces a centrally symmetrical four-quadrant distribution pattern on the observation plane.

[0090] Figure 13 The magnetic data for projectile 1, obtained in an actual experiment, were obtained at the same detection altitude as the simulation conditions. Except for the magnetic vector data, which is severely affected by the carrier's attitude, the spatial distribution of the remaining physical components largely matches the simulation results; in terms of numerical magnitude, the simulated data and the measured data are quite similar. These comparison results verify the correctness of the simulation modeling method and the physical realism of the generated data.

[0091] Furthermore, to simulate the uncertainty of unexploded ordnance burial depth and the deviation of UAV flight altitude caused by terrain undulations or airflow disturbances in actual operations, the simulation strategy of a single fixed altitude was abandoned. For large unexploded ordnance targets, eight planes at different heights were randomly selected as observation planes within a detection height range of 3m to 5m (corresponding to the relative vertical distance from the sensor to the target). For howitzers and anti-tank mines, eight planes at different heights were randomly selected as observation planes within a detection height range of 1m to 2m. By combining and traversing the above 800 sets of target attitudes with eight observation heights, the distribution range of the samples in vertical space was expanded, improving the representativeness and robustness of the simulation database.

[0092] Meanwhile, to ensure the representativeness and universality of the forward model library, and considering the significant uncertainty in the attitude of unexploded ordnance targets after impact in actual battlefield environments, a randomized generation strategy was adopted for their spatial attitudes while keeping the geometric dimensions and magnetic parameters of five typical unexploded ordnance constant. Batch forward simulation was employed, randomly generating roll, pitch, and yaw angles in three-dimensional space to construct 800 distinct spatial attitude combinations for each type of unexploded ordnance model. The generated comprehensive attitude data enabled the forward model library to successfully cover the entire state space of the angle between the target's magnetic moment direction and the geomagnetic background field, providing a rich data foundation for subsequent analysis of the nonlinear mapping relationship between the target's magnetic moment characteristics and spatial position.

[0093] Finally, by combining five types of unexploded ordnance, 800 sets of random attitudes, and eight observation altitudes, a forward model library of unexploded ordnance containing 32,000 samples was constructed. This large and feature-rich database not only covers the entire state space from near field to far field and from horizontal placement to large-angle tilt, but also provides sufficient and reliable sample support for feature extraction and parameter training of subsequent target neural network models. Due to the large amount of simulation data, it is difficult to display it all. To intuitively reveal the variation of the target's magnetic characteristics with spatial attitude, the magnetic moment variation diagrams of the five types of unexploded ordnance under different attitudes were statistically analyzed and plotted. The results are as follows: Figure 14 As shown.

[0094] Figure 14 The horizontal axis represents the azimuth angle, i.e., the rotation angle of the unexploded ordnance about the z-axis, and the vertical axis represents the elevation angle, i.e., the rotation angle of the unexploded ordnance about the x-axis. From the results, it is easy to see that the magnetic moments of the four ellipsoidal unexploded ordnance exhibit a high degree of consistency with attitude, with only significant differences in magnitude. Since the mine is cylindrical, its magnetic moment is almost unaffected by azimuth changes when in a horizontal orientation. Specifically, the magnitude of the magnetic moment of shell 1 varies with the target attitude as follows: The minimum value is 11.42% of the maximum value; the magnitude of the magnetic moment of projectile 2 varies with the target attitude within the following range: The minimum value is 9.75% of the maximum value; the magnitude of the projectile's magnetic moment varies with the target's attitude within the following range: The minimum value is 12.17% of the maximum value; the magnitude of the magnetic moment of a certain type of howitzer varies with the target attitude within the following range: The minimum value is 9.67% of the maximum value. The range of the magnitude of the magnetic moment of a certain type of howitzer as a function of target attitude is as follows: The minimum value is 57.14% of the maximum value. The simulation data of the five unexploded ordnance show the same numerical pattern: the maximum value is taken when the target's long axis is aligned with the direction of the geomagnetic field, and the minimum value is taken when the target's long axis is perpendicular to the direction of the geomagnetic field; the greater the deviation between the target's long axis and the direction of the geomagnetic field, the smaller the target's magnetic moment value.

[0095] Step S2012: Construct the interference database.

[0096] In actual unexploded ordnance detection scenarios, the shallow surface layer often contains interfering materials such as fragmented shrapnel and scrap metal. The magnetic anomaly signals generated by these ferromagnetic impurities under the influence of the Earth's magnetic field are easily confused with target signals, affecting the accuracy of the identification model and leading to false alarms or missed detections. To improve the model's anti-interference capability in complex environments, this invention constructs an interference database containing interference sources. Specifically, the database is constructed using a combination of finite element parameter calculation and analytical batch generation. First, a solid model of the surface interference is established using COMSOL software, and its magnetic moment after being magnetized by the Earth's magnetic field is calculated. Given that these interference materials are small in size and their distance from the detection plane satisfies the far-field approximation condition, the distribution of the interfering magnetic field is then rapidly calculated using the MATLAB platform based on magnetic dipole theory to improve data generation efficiency.

[0097] In the specific simulation configuration, 1 to 5 fragments are randomly distributed as interference sources within each sample scenario. The size of the interference objects is randomly generated within 10 cm, and their horizontal positions are randomly generated within the detected area. The burial depth is controlled between the surface and 0.2 m underground to simulate scattered metal waste in the shallow surface layer. Based on the above strategy, 800 sets of interference field data are generated, consistent with the unexploded ordnance forward modeling strategy. Each set of data traverses 8 different observation heights. The sampling resolution, grid settings, and data channel structure of this interference database are strictly consistent with the unexploded ordnance forward modeling model library, ensuring compatibility between the two in the feature space. The results are as follows: Figure 15 As shown.

[0098] from Figure 15 As can be seen, unlike the magnetic field generated by unexploded ordnance targets, the magnetic anomalies generated by the interfering targets are characterized by local spikes or chaotic multipole superposition due to their extremely shallow burial and small size. Although the peak intensity of individual strong interference sources may be close to that of small unexploded ordnance, their signals decay at a much faster rate with increasing detection distance.

[0099] Step S202: Based on the acquired real noise samples, the simulation database is enhanced to obtain the enhanced dataset.

[0100] (1) Environmental noise analysis To compensate for the distribution differences between the data in the simulation database and the real working environment, and to improve the generalization ability and anti-interference robustness of the deep neural network model, data augmentation based on noise injection is required. Real noise samples were obtained by conducting an environmental background magnetic field acquisition experiment. Given the dynamic nature of the geomagnetic field over time and its susceptibility to random interference from surrounding environmental factors, an environmental magnetic field noise acquisition experiment was conducted in Nanjing. A fluxgate magnetometer array acquisition system was placed on a plane 2 m above the ground to collect environmental magnetic field data. The sampling frequency was set to 100 Hz, the sampling time to 90 s, and four fluxgate magnetometers collected data simultaneously. The time-domain signal of the total local environmental magnetic field acquired by the four fluxgate magnetometers is shown below. Figure 16 As shown.

[0101] exist Figure 16 In the figure, the horizontal axis represents the acquisition time from 0 to 90 s, and the vertical axis represents the magnetic induction intensity in nT. Four different colored curves represent the changes in the geomagnetic field acquired by fluxgates S1-S4. Statistical analysis was performed on the time-domain signals in the figure, and the mean, standard deviation, maximum, minimum, and median of the geomagnetic field acquired by each fluxgate sensor were calculated. The statistical results are shown in Table 3.

[0102] Table 3: Statistical Distribution of Environmental Magnetic Field Time-Domain Signals (Unit: nT)

[0103] Analysis of the statistical data in Table 3 shows that the mean and median measurements of each channel sensor are roughly close, and the background field intensity recorded by fluxgates S1 to S4 remains around 50,290 nT. Regarding fluctuation characteristics, the standard deviation of the data for each channel ranges from 1.87 nT to 2.13 nT, with an overall average standard deviation of 1.99 nT. To further reveal the statistical regularity of environmental magnetic noise, corresponding probability density function (PDF) and cumulative distribution function (CDF) curves were constructed based on the above measured data, as shown below. Figure 17 As shown.

[0104] from Figure 17 As shown in the graph, the horizontal axis represents magnetic flux density (unit: nT). The graph uses a dual vertical axis coordinate system: the left vertical axis corresponds to the probability density distribution histogram of the data, and the right vertical axis corresponds to the cumulative distribution function (CDF) curve. Analysis of the curve shape reveals that the CDF curve exhibits an S-shaped growth characteristic, with its growth rate peaking near the statistical mean and gradually slowing down towards both ends. Combining the contour presented by the histogram with the inflection point characteristics of the CDF curve, it can be shown that the environmental background magnetic field fluctuations collected by each fluxgate sensor follow a Gaussian normal distribution.

[0105] (2) Database expansion and partitioning Statistical analysis of the environmental background field shows that short-term geomagnetic field fluctuations can be approximated as a random process following a Gaussian distribution. Referring to the measured data in Table 3, the average standard deviation of the geomagnetic field fluctuations collected by each fluxgate sensor is approximately 1.99 nT. Based on this, this embodiment of the invention constructs a Gaussian normal distribution model with a mean of 0 nT and a standard deviation of 2 nT to simulate environmental noise in actual field experiments. A random noise matrix conforming to this distribution is generated using Python's NumPy numerical computation library and superimposed onto the simulation database as additive noise. This process not only simulates real background interference but also effectively expands the feature diversity of the original database.

[0106] Building upon noise enhancement, random pruning expands the database in size and diversity. The original simulated magnetic gradient tensor data has a grid size of 1001×1001, containing a large redundant background region. To adapt to the input dimension of the neural network and increase the number of samples, 15 local sub-images of size 501×501 are randomly selected from each 1001×1001 original sample as the final training samples. This pruning operation not only expands the database size by 15 times but also forces the network to learn translation-invariant feature representations by simulating the distribution of targets at different relative positions in the field of view, thus significantly improving the model's adaptability to changes in target position.

[0107] Considering that the constructed forward modeling dataset covers multimodal physical quantities such as the total magnetic field, vector field, and gradient tensor, and that the original data exhibit significant distributional differences in numerical dynamic range due to the different physical definitions and dimensions of each physical channel, this embodiment of the invention implements channel-specific normalization to eliminate the dimensional influence between different channels and construct a standardized benchmark dataset with unified statistical properties. For the first... Input data for each channel Its normalized value The calculation formula is shown in Formula 11:

[0108] in, These are the original pixel values. and These represent the maximum and minimum values ​​of the data within that channel, respectively.

[0109] Through this linear mapping, the data distribution of all channels is uniformly scaled to the [0,1] interval, which effectively accelerates the optimization process of the optimization algorithm.

[0110] Step S203: According to the principle of independent and identically distributed data, the enhanced dataset is divided into training set, validation set and test set according to the set ratio by random sampling.

[0111] After the dataset is constructed, to facilitate subsequent model performance training and evaluation, this embodiment of the invention follows the principle of independent and identically distributed (IOD) data, using random sampling to divide the expanded dataset into training, validation, and test sets in a ratio of 80%, 10%, and 10%, respectively. The training set is used for subsequent gradient descent and weight updates of model parameters; the validation set is used to monitor model performance in real time during training, assist in hyperparameter optimization, and implement early stopping strategies; the test set does not participate in the training process at all, and is only used for final objective performance evaluation after the model is finalized.

[0112] Step S204: Initialize the structure and parameters of the target neural network model, define the loss function for multi-task features, select the parameter optimization algorithm, and design a training strategy to prevent overfitting.

[0113] Specifically, the training process of a neural network model is essentially a non-convex optimization problem that seeks the global optimum in a multi-dimensional parameter space. To enable the constructed model to effectively learn features from large-scale magnetic data and achieve accurate identification and localization, a complete training method was developed, including the definition of a loss function for multi-task characteristics, the selection of efficient parameter optimization algorithms, and the design of training strategies to prevent overfitting.

[0114] The target neural network model constructed in this embodiment of the invention includes two distinct task branches: unexploded ordnance type identification and spatial location prediction. Since the classification task aims to distinguish discrete category labels, while the localization task aims to regress continuous spatial coordinates, their output spaces and optimization objectives differ fundamentally. Therefore, to ensure that the network can perform accurate gradient descent and parameter updates for specific task characteristics, appropriate loss functions are designed for the two different regression heads.

[0115] The unexploded ordnance (UFO) classification task is essentially a multi-class classification problem, aiming to maximize the consistency between the model's output probability distribution and the true class labels. Therefore, the cross-entropy loss function is chosen as the optimization objective for this classification task. The cross-entropy function effectively measures the distance between probability distributions, imposing a larger penalty on misclassified samples with high confidence, thus prompting the model to converge quickly. Assume the total number of UFO categories is... For the first For each input sample, its true class's one-hot encoded vector is: The predicted probability vector output by the fully connected classification head is Then classification loss The definition is shown in Formula 12 below:

[0116] In the formula, This represents the number of samples in the current batch. Indicates the first The sample belongs to the first The actual label of the class (0 or 1). This represents the predicted probability after Softmax normalization. By minimizing this loss, the model can learn magnetic anomaly features with high discriminative power, thereby improving the identification accuracy.

[0117] For unexploded ordnance localization, the fundamental problem is a continuous numerical regression problem, aiming to minimize the deviation between the predicted coordinates and the actual physical location. Therefore, the Mean Squared Error (MSE) loss function is chosen as the optimization objective for the localization task. The MSE loss function calculates the expected square of the difference between the predicted and actual values, and possesses the characteristic of being a convex function that is differentiable everywhere, guiding the model to smoothly approximate the true coordinates in space. Let the... The true normalized coordinates of each sample are The predicted coordinates output by the fully convolutional localization head are Then the location loss Defined as:

[0118] The loss function in Formula 13 is directly related to spatial distance. The backpropagation algorithm drives the adjustment of network parameters, enabling the fully convolutional regression head to accurately resolve the spatial topological information in the magnetic gradient tensor, thereby achieving high-precision location prediction.

[0119] To efficiently solve the aforementioned loss function and update the network weights, this embodiment of the invention employs the Adam optimization algorithm. When processing magnetic anomaly detection tasks, the input data contains multimodal information such as the total magnetic field, vector field, and gradient tensor, and the gradient sparsity and numerical magnitude of different feature channels may differ significantly. The Adam algorithm combines the advantages of the momentum method and the RMSProp algorithm, and by calculating the first and second moment estimates of the gradient, it can dynamically adjust an independent adaptive learning rate for each parameter in the network.

[0120] In terms of the specific training process configuration, this embodiment of the invention adopts the mini-batch gradient descent paradigm, setting the batch size to 50. This means that the parameters are updated once per iteration using the average gradient of 50 samples, ensuring the stability of gradient estimation while fully utilizing the parallel computing power of the GPU. An epoch is defined as the time when the model completes its traversal of the entire training set. To balance convergence speed in the early stages of training with optimization accuracy in the later stages, and to prevent overfitting on a limited simulation database, a training strategy combining dynamic learning rate adjustment and early stopping is implemented.

[0121] First, a learning rate decay mechanism is employed. A relatively large initial learning rate is set during the early stages of training to guide the model quickly out of the initial flat region. As training progresses, the learning rate is automatically reduced when the validation set loss no longer decreases within a preset number of epochs. This coarse-to-fine search strategy allows the model to be finely adjusted in the parameter space with smaller steps in the later stages of training, thereby effectively approximating the global optimum.

[0122] Step S205 involves inputting a specific number of samples selected from the training set into a convolutional neural network to output a deep feature map; then inputting the deep feature map into a fully connected regression head branch model and a fully convolutional network regression head branch model respectively to output the predicted class probability and coordinates.

[0123] Specifically, samples are randomly selected from the training set in batches of 50. The input format is (batch × channel × height × width). After checking the integrity of the sample channels and the numerical range, the samples are input into a convolutional neural network for feature extraction. The forward propagation process is as follows: Feature extraction module 1: 9-channel input → 3×3 convolution (32 channels) → ELU activation → hybrid attention (channel + spatial) → 2×2 max pooling → output; Feature extraction module 2: 32-channel input → 3×3 convolution (64 channels) → ELU activation → hybrid attention → 2×2 max pooling → output; Feature extraction module 3: 64-channel input → 3×3 convolution (128 channels) → ELU activation → hybrid attention → 2×2 max pooling → output; Feature extraction module 4: 128-channel input → 3×3 convolution (256 channels) → ELU activation → hybrid attention → 2×2 max pooling → output; Feature extraction module 5: 256-channel input → 3×3 convolution (512 channels) → ELU activation → hybrid attention → 2×2 max pooling → output deep feature map (512×15×15).

[0124] Next, the fully connected regression head branch model flattens the deep feature map to obtain a 115200-dimensional vector, then passes it through a hidden layer (first fully connected layer, second activation function, and Dropout random deactivation layer) to obtain a 128-dimensional vector. Finally, it is processed by the output layer (second fully connected layer and Softmax function) to obtain a 5-dimensional probability vector. (Predicted category probability). Simultaneously, the fully convolutional network regression head branch model first reduces the dimensionality of the deep feature map through the first convolutional layer and the first activation function to obtain a 128×15×15 first intermediate feature; then, it maps the first intermediate feature through the second convolutional layer to obtain a 2×15×15 first response map; finally, it aggregates the first response map through the pooling layer to obtain a 2-dimensional vector (predicted normalized coordinates).

[0125] Step S206: Based on the real labels and real coordinates, calculate the classification loss and localization loss using the cross-entropy loss function and the mean squared error loss function, and sum them to obtain the total loss.

[0126] Step S207: The gradient is calculated automatically using PyTorch differentiation, and the Adam optimizer updates the weights.

[0127] Specifically, the automatic differentiation engine (torch.autograd) based on the PyTorch framework performs backpropagation on the total loss, calculating the gradients of all learnable parameters of the model (convolutional kernel weights, fully connected layer weights, bias terms). Simultaneously, the Adam optimizer combines the advantages of momentum and RMSprop, dynamically adjusting the independent learning rate for each parameter by maintaining estimates of the first moment (mean) and second moment (uncentered variance) of the gradient. During the training of the target neural network model, the Adam optimizer is responsible for updating the model weights (convolutional kernels, fully connected layer parameters, etc.) based on the gradient of the loss function.

[0128] Step S208: Repeat the execution and complete one training round iteration; terminate training when the preset training rounds are completed or the early stop mechanism is triggered.

[0129] Specifically, repeat steps S205 (input sample) → S206 (calculate loss) → S207 (update weights) until the number of iterations in a single epoch is reached, which completes one training epoch. Then execute: Input the validation set into the model in batches of 50, and output the validation set prediction results. Calculate the classification loss, localization loss, and total loss of the validation set using the S206 method; Record the training set loss, validation set loss, classification accuracy, and localization RMSE for the current epoch, and plot the loss curve and metric curve. If the total loss of the validation set does not decrease for 10 consecutive epochs, adjust the Adam optimizer learning rate to 0.1 times its original value (e.g., 0.001→0.0001→0.00001), decaying a total of 2 times; If the validation set loss does not decrease for 20 consecutive epochs, the model is considered to have converged, and training is terminated early (in this embodiment, this is triggered at the 180th epoch). The model weights (including the feature extraction module and multi-task regression head parameters) are saved every 10 epochs, and only the model weights with the lowest validation set loss are retained. After completing 200 epochs of training, or triggering the early stop mechanism (180th epoch), the model weights with the lowest loss on the validation set are used as the final target neural network model for unexploded ordnance magnetic detection.

[0130] In one implementation, choosing a suitable performance metric is crucial for evaluating the classification and detection accuracy of the network model. Since the model contains two parallel output branches, classification and regression, this implementation will define quantitative evaluation criteria for each task.

[0131] For the task of identifying unexploded ordnance types, a confusion matrix is ​​first introduced as a basic analysis tool to visually display the model's classification distribution and misclassification patterns across different categories. Based on this, the following indicators are used as core quantitative metrics.

[0132] Accuracy is used to measure a model’s overall ability to identify all samples;

[0133] in, , , and These represent the number of true positive, true negative, false positive, and false negative samples, respectively.

[0134] Accuracy reflects the proportion of samples that the model predicts as belonging to a certain type of target, and demonstrates the ability to resist false alarms.

[0135] Recall rate describes the proportion of a real target of a certain type that is successfully detected. In the field of unexploded ordnance detection, recall rate has extremely high safety significance because it is directly related to the risk of "missed detection".

[0136]

[0137] The F1 score, as the harmonic mean of precision and recall, comprehensively evaluates the robustness of a model under class imbalance or when both precision and recall are required. The formulas for calculating these metrics are as follows:

[0138] For the task of predicting the horizontal position of unexploded ordnance, it is essentially a regression problem, aiming to minimize the deviation between the predicted coordinates and the actual physical coordinates. This invention uses Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) to measure the regression dispersion of the model on the independent components of the X and Y axes. RMSE is more sensitive to larger errors and can reflect the fluctuation of the predicted values; while MAE better reflects the actual average level of the prediction error. Furthermore, to intuitively reflect the actual physical offset of the detection results on the two-dimensional plane, the Mean Positioning Error (MPE) is defined as the distance between the predicted point and the actual point. This metric directly represents the geometric distance of the positioning point given by the algorithm from the actual target, and is the most critical physical indicator guiding actual excavation operations. Its calculation formula is defined as:

[0139] in, This represents the total number of samples in the test set. Using the above indicator system, the proposed neural network model can be comprehensively validated from two dimensions: recognition accuracy and localization precision.

[0140] Example 3 The above describes the unexploded ordnance magnetic detection identification method provided by the embodiments of the present invention. Based on the same idea, the embodiments of the present invention also provide an unexploded ordnance magnetic detection identification device 300, such as... Figure 18 As shown.

[0141] The magnetic detection device for identifying unexploded ordnance includes: The acquisition module 301 is used to acquire magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data of the probed area; The first processing module 302 is used to perform channel superposition on magnetic vector field data, total magnetic field data and magnetic vector gradient tensor data to obtain magnetic data. The identification module 303 is used to input magnetic data into the target neural network model to obtain the detection and identification results regarding whether there is an unexploded ordnance in the detected area. The target neural network model includes a convolutional neural network and a multi-task regression head model. The convolutional neural network is used to extract deep feature maps containing high-dimensional semantic information contained in the magnetic data. The multi-task regression head model is used to obtain the detection and identification results based on the deep feature maps.

[0142] In one implementation, the multi-task regression head model includes a fully connected regression head branch model and a fully convolutional network regression head branch model; the detection and recognition results include detection classification results and detection localization results; the recognition module 303 is also used for: Deep feature maps containing high-dimensional semantic information in magnetic data are extracted using convolutional neural networks; The deep feature maps are input into the fully connected regression head branch model and the fully convolutional network regression head branch model, respectively. The detection classification result is obtained through the fully connected regression head branch model; the detection localization result is obtained through the fully convolutional network regression head branch model. The detection classification result represents the category of unexploded ordnance in the detected area; the detection localization result represents the location of the unexploded ordnance in the detected area.

[0143] In one implementation, the identification module 303 is further used for: The first intermediate feature is obtained by reducing the dimensionality of the deep feature map through the first convolutional layer and the first activation function. The first intermediate feature is mapped by the second convolutional layer to obtain the first response map; The detection and localization results are obtained by aggregating the first response map using a pooling layer.

[0144] In one implementation, the identification module 303 is further used for: Flatten the deep feature map to obtain the second intermediate feature; The second intermediate feature is used as input and processed sequentially through the first fully connected layer, the second activation function, the Dropout random deactivation layer, the second fully connected layer, and the Softmax function to obtain the detection and classification results.

[0145] In one implementation, there are multiple convolutional neural networks, each including a third convolutional layer, a third activation function, an attention mechanism, and a first pooling layer; the recognition module 303 is also used for: The magnetic data is convolved by the third convolutional layer to obtain the first feature map; The negative input in the first feature map is non-linearly transformed by the third activation function to obtain the second feature map; The first attention map corresponding to the second feature map is determined by an attention mechanism, and the second feature map and the first attention map are fused to obtain the third feature map; The third feature map is pooled using the first pooling layer to obtain the deep feature map.

[0146] In one implementation, the first attention map includes a channel attention map and a spatial attention map, and the attention mechanism includes a channel attention mechanism and a spatial attention mechanism; the recognition module 303 is also used for: The second feature map is used as input, and the channel attention map and spatial attention map are obtained through the channel attention mechanism and spatial attention mechanism, respectively. The third feature map is obtained by multiplying the second feature map, the channel attention map, and the spatial attention map element by element.

[0147] In one implementation, the acquisition module 301 includes: a first acquisition module 3011, a first determination module 3012, and a second determination module 3013; The first acquisition module 3011 is used to acquire magnetic vector field data of the probed area through a planar square fluxgate detection array; The first determining module 3012 is used to determine the total magnetic field data based on the magnetic field values ​​of each fluxgate in the magnetic vector field data along the first axis, the second axis and the third axis. The second determining module 3013 is used to determine the magnetic vector gradient tensor data by means of the finite difference method based on the side length of the planar square fluxgate detector array and the magnetic field value of each fluxgate along the first axis, the second axis and the third axis in the magnetic vector field data.

[0148] The unexploded ordnance magnetic detection identification device provided in this invention includes an acquisition module, a first processing module, and an identification model. The acquisition module acquires magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data of the probed area. The first processing module performs channel superposition of the magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data to obtain magnetic data. The identification module then inputs the magnetic data into a target neural network model, which includes a convolutional neural network and a multi-task regression head model, to obtain a detection and identification result regarding the presence of an unexploded ordnance in the probed area. This device inputs multimodal magnetic data, including magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data, into the identification module, which includes a target neural network model. It extracts high-dimensional semantic information from the multimodal magnetic data and generates a deep feature map, effectively capturing the characteristics of weak magnetic anomaly amplitude, strong spatial diffusion, and significant local gradients, thereby accurately identifying unexploded ordnance in the probed area and improving the accuracy of magnetic detection under non-ideal detection conditions.

[0149] Example 4 Figure 19 To illustrate the hardware structure of an unexploded ordnance magnetic detection electronic device according to various embodiments of the present invention, the electronic device includes a processor 401 and a memory 402 storing computer program instructions. Specifically, the processor 401 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of the present invention.

[0150] Memory 402 may include mass storage for data or instructions. For example, and not limitingly, memory 402 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 402 may include removable or non-removable (or fixed) media. Where appropriate, memory 402 may be internal or external to an electronic device. In a particular embodiment, memory 402 may be a non-volatile solid-state memory.

[0151] In one embodiment, memory 402 may be read-only memory (ROM). In one embodiment, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0152] The processor 401 reads and executes computer program instructions stored in the memory 402 to implement any of the unexploded ordnance magnetic detection methods in the above embodiments.

[0153] In one example, the electronic device may also include a communication interface 403 and a bus 410. For example, Figure 19 As shown, the processor 401, memory 402, and communication interface 403 are connected through bus 410 and complete communication with each other.

[0154] The communication interface 403 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of the present invention.

[0155] Bus 410 includes hardware, software, or both, that couples components of an electronic device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 410 may include one or more buses. While specific buses are described and illustrated in embodiments of the invention, the invention contemplates any suitable bus or interconnect.

[0156] Furthermore, in conjunction with the unexploded ordnance magnetic detection identification method in the above embodiments, this invention can be implemented using a computer-readable storage medium. This computer-readable storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the unexploded ordnance magnetic detection identification methods in the above embodiments.

[0157] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0158] The above description is merely a specific implementation of the present invention. Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0159] Secondly, those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0160] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0161] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0162] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0163] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0164] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0165] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0166] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0167] The above description is merely an 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 principle of the present invention should be included within the scope of the claims of the present invention.

[0168] The entire process of model training and application involved in this application strictly follows Article 5 of the Patent Law of the People's Republic of China (hereinafter referred to as "A5") and the relevant examination standards of the Patent Examination Guidelines (2023 Edition) to ensure that the technical solution does not violate the law, social morality and public interest, and that the data acquisition and utilization and model training process all comply with ethical compliance requirements. Specific details are as follows: I. Explanation of whether the data used in the model meets A5 requirements 1. Data Source Legality: All datasets used in training this model were obtained through legal means, covering three categories: publicly authorized data, data authorized by partners, and self-collected compliant data. Publicly authorized data comes from compliant data sources following open-source licenses such as Apache 2.0, with complete copyright attribution and authorization scope clearly marked, and no unauthorized open-source code or data reuse. Data authorized by partners has a formal data usage agreement clearly defining the scope, duration, and confidentiality obligations, with complete authorization chain proof. For self-collected data involving personal information, strict informed consent procedures have been followed, and anonymization processes (including but not limited to field masking, feature anonymization, and differential privacy technology application) have been used to remove personally identifiable information, fully complying with the requirements of the "Interim Measures for the Administration of Generative Artificial Intelligence Services," the "Personal Information Protection Law," and other relevant laws and regulations.

[0169] 2. Data Content Compliance: This dataset has undergone multiple screening and cleaning processes to remove all content that may violate social morality or harm public interests, and does not involve the illegal acquisition or use of genetic resources. For data in sensitive fields (such as medical and financial fields), an additional privacy-preserving computation module (including federated learning and secure multi-party computation technology) is used to ensure that the data is "usable but not visible," avoiding compliance risks during the original data transmission process and ensuring that the data application scenarios and uses comply with public order and good morals and industry regulatory requirements.

[0170] 3. Data Governance Compliance: Establish a complete data traceability system to automatically record the source, collection time, annotation process, cleaning rules, and permission allocation of training data, generating traceable compliance reports to ensure data is verifiable throughout its entire lifecycle. The dataset annotation process is completed by a professional human R&D team, clearly defining the proportion of human creative contributions, avoiding reliance on AI-generated data that has not undergone substantial human modification, and complying with the "human main contribution" examination requirements in AI patent applications.

[0171] II. Explanation of Model Training Process Meeting A5 Requirements 1. Compliance of training objectives and scheme: The training objectives, training scheme and final output results of this model do not violate any mandatory provisions of laws and administrative regulations, do not harm the public interest or the legitimate rights and interests of others, and do not pose any potential risks of being used for illegal activities, privacy infringement or public safety disruption. It strictly adheres to the ethical principle of "intelligent for good".

[0172] 2. Compliance Management of Training Process: A closed-loop training framework is adopted to ensure compliance and controllability of the training process. The specific process is as follows: First, training samples are obtained through compliant data sources. After the aforementioned data cleaning and desensitization, they are input into the first neural network model to generate preliminary training results. Second, an expert system is introduced to verify the preliminary results. Based on preset rules and human expert experience, the feasibility of the results is evaluated, and outputs that may pose ethical risks or compliance hazards are corrected (such as removing decision logic that violates public order and good morals, and adjusting model parameters that do not comply with safety regulations). Finally, the loss function weights are dynamically optimized based on expert system feedback to strengthen the model's learning of compliant results, avoid overfitting errors or non-compliant labels, and form a closed-loop management system of "data input - model training - expert verification - parameter optimization - result feedback" to ensure that the entire training process complies with A5 ethical review requirements.

[0173] 3. Compliance of Training Environment and Tools: Model training is implemented on a compliant training platform. All open-source frameworks and components used in the training process have obtained the corresponding licenses, and copyright statements and patent citation information are fully retained, with no infringement or reuse. The training environment is built using virtual devices (containers / virtual machines) with fixed random seeds and initial parameter configurations to ensure the reproducibility of the training process. At the same time, through access control and operation log recording, risks such as data leakage and parameter tampering during training are prevented, ensuring the security and compliance of the training process.

[0174] 4. Ethical verification of training results: After the model is trained, it will undergo an additional third-party ethical compliance assessment and algorithm filing review to verify that the model output does not violate social morality or harm public interests. For potentially sensitive scenarios (such as public services and intelligent decision-making), a special result verification mechanism will be established to ensure that the model always complies with A5 and relevant laws and regulations in practical applications.

[0175] In summary, the data and training process used in this application model strictly comply with the relevant provisions of Article 5 of the Patent Law and the Patent Examination Guidelines (2023 Edition), and there are no violations of laws, social ethics, public interests, or illegal use of genetic resources. Therefore, it fully meets the compliance requirements for patent authorization.

Claims

1. A method for magnetic detection identification of unexploded ordnance, characterized in that, The method includes: Acquire magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data of the probed region; The magnetic vector field data, the total magnetic field data, and the magnetic vector gradient tensor data are superimposed through channels to obtain magnetic data. The magnetic data is input into the target neural network model to obtain the detection and identification results regarding whether there is an unexploded ordnance in the detected area; The target neural network model includes a convolutional neural network and a multi-task regression head model; the convolutional neural network is used to extract deep feature maps containing high-dimensional semantic information contained in the magnetic data; the multi-task regression head model is used to obtain the detection and recognition results based on the deep feature maps.

2. The method according to claim 1, characterized in that, The multi-task regression head model includes a fully connected regression head branch model and a fully convolutional network regression head branch model; the detection and identification results include detection classification results and detection localization results; the step of inputting the magnetic data into the target neural network model to obtain detection and identification results regarding the presence of unexploded ordnance in the detected area includes: The convolutional neural network is used to extract deep feature maps containing high-dimensional semantic information from the magnetic data. The deep feature map is input into the fully connected regression head branch model and the fully convolutional network regression head branch model, respectively. The detection classification result is obtained through the fully connected regression head branch model; the detection localization result is obtained through the fully convolutional network regression head branch model; the detection classification result represents the category of the unexploded ordnance in the detected area; the detection localization result represents the position of the unexploded ordnance in the detected area.

3. The method according to claim 2, characterized in that, The process of obtaining the detection and localization results through the regression head branch model of the fully convolutional network includes: The deep feature map is reduced in dimensionality by using a first convolutional layer and a first activation function to obtain a first intermediate feature. The first intermediate feature is mapped by the second convolutional layer to obtain the first response map; The detection and localization results are obtained by aggregating the first response map using a pooling layer.

4. The method according to claim 2, characterized in that, The process of obtaining the detection classification result through the fully connected regression head branch model includes: The deep feature map is flattened to obtain the second intermediate feature; The second intermediate feature is used as input and processed sequentially through a first fully connected layer, a second activation function, a Dropout random deactivation layer, a second fully connected layer, and a Softmax function to obtain the detection and classification result.

5. The method according to claim 1, characterized in that, The convolutional neural networks are multiple, each including a third convolutional layer, a third activation function, an attention mechanism, and a first pooling layer; the step of inputting the magnetic data into the target neural network model to obtain the detection and identification results of the unexploded ordnance includes: The magnetic data is input into the target neural network model, and the target neural network performs the following: The magnetic data is convolved by the third convolutional layer to obtain the first feature map; The negative input in the first feature map is nonlinearly transformed by the third activation function to obtain the second feature map; The first attention map corresponding to the second feature map is determined through the attention mechanism, and the second feature map and the first attention map are fused to obtain the third feature map; The deep feature map is obtained by pooling the third feature map through the first pooling layer.

6. The method according to claim 5, characterized in that, The first attention map includes a channel attention map and a spatial attention map, and the attention mechanism includes a channel attention mechanism and a spatial attention mechanism; the step of determining the first attention map corresponding to the second feature map through the attention mechanism, and fusing the second feature map and the first attention map to obtain the third feature map includes: The second feature map is used as input, and channel attention map and spatial attention map are obtained through channel attention mechanism and spatial attention mechanism, respectively; The second feature map, the channel attention map, and the spatial attention map are multiplied element-wise to obtain the third feature map.

7. The method according to claim 1, characterized in that, The acquisition of magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data includes: The magnetic vector field data of the probed area is obtained by using a planar square fluxgate detector array; The total magnetic field data is determined based on the magnetic field values ​​of each fluxgate in the magnetic vector field data along the first axis, the second axis, and the third axis. The magnetic vector gradient tensor data is determined by the finite difference method based on the side length of the planar square fluxgate detector array and the magnetic field values ​​of each fluxgate in the magnetic vector field data along the first axis, the second axis, and the third axis.

8. A magnetic detection device for unexploded ordnance, characterized in that, The device includes: The acquisition module is used to acquire magnetic vector field data, total magnetic field data, and magnetic vector gradient tensor data of the probed area; The first processing module is used to perform channel superposition on the magnetic vector field data, the total magnetic field data, and the magnetic vector gradient tensor data to obtain magnetic data. The identification module is used to input the magnetic data into a target neural network model to obtain a detection and identification result regarding whether there is an unexploded ordnance in the detected area; the target neural network model includes a convolutional neural network and a multi-task regression head model; the convolutional neural network is used to extract a deep feature map containing high-dimensional semantic information contained in the magnetic data; the multi-task regression head model is used to obtain the detection and identification result based on the deep feature map.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the unexploded ordnance magnetic detection identification method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the unexploded ordnance magnetic detection identification method as described in any one of claims 1 to 7.

Images