A magnetic target intelligent positioning method and system based on a hierarchical neural network architecture

Abstract

The application relates to the field of geomagnetic vector measurement, and discloses a magnetic target intelligent positioning method and system based on a layered neural network architecture. The method comprises the following steps: acquiring multi-point magnetic field data measured by a magnetic sensor array; calculating magnetic gradient tensor data according to the magnetic field data and the spatial position of the sensor; inputting the magnetic gradient tensor data into a pre-constructed layered neural network model; extracting local space-time features from the magnetic gradient tensor time series data of each measurement point through a sub-network layer; performing time series fusion and global context modeling on the feature sequences output by all sub-network branches through a global fusion network layer; and outputting the three-dimensional spatial position coordinates of the magnetic target according to the fused global features through an output layer. The implementation of the application is a deep integration of physical analysis methods and data-driven methods, can guide the mapping relationship between the magnetic gradient tensor and the position of the magnetic target of the deep learning model, and retains the strong nonlinear fitting capability, thereby enhancing the positioning accuracy of the magnetic target.

Description

Technical Field

[0001] This invention relates to the field of geomagnetic vector measurement technology, and more specifically, to a method and system for intelligent positioning of magnetic targets based on a hierarchical neural network architecture. Background Technology

[0002] Magnetic anomaly detection technology is a highly effective and covert technique. It has been applied across multiple disciplines and sectors, including geological exploration, biomedical engineering, underground engineering monitoring, and many other fields. This technology has applications in underwater submarine detection, unexploded ordnance detection, mineral exploration, and cardiovascular MRI and magnetoencephalography (MEG). The magnetic field vector generated by ferromagnetic targets affects the distribution of the surrounding geomagnetic field, leading to magnetic anomalies. By measuring the magnetic field information using magnetic sensors, and through data processing and analysis, the location information of the magnetic target can be reconstructed.

[0003] Currently, magnetic target localization methods mainly fall into two categories: analytical methods based on physical models and methods based on deep learning. Analytical methods based on physical models establish explicit analytical relationships between magnetic measurement parameters and target positions to invert the spatial coordinates of the magnetic source. For example, Nara et al. introduced a magnetic gradient tensor and a magnetic field vector into Euler's formula to solve for the position information of the magnetic target. Wilson et al. used the eigenvalues ​​and eigenvectors of the magnetic gradient tensor to achieve magnetic target localization. Davis et al. proposed a Hilbert transform data processing method based on the Euler deconvolution localization method. Wiegert et al. proposed a scalar triangulation and ranging method based on the invariant of the magnetic gradient tensor. Clark et al. studied a technique for localizing magnetic targets using the normalized magnetic source intensity method. Ding et al. employed an improved normalized source intensity method to localize magnetic targets. Sui et al. proposed using a higher-order magnetic gradient tensor to localize magnetic targets. However, this method relies on strict ideal assumptions; directly applying the linear relationships derived from these ideal assumptions to real-world nonlinear problems will lead to significant localization errors.

[0004] In recent years, deep learning methods have demonstrated significant advantages in magnetic target detection due to their powerful nonlinear fitting capabilities. Deep learning methods directly learn the end-to-end mapping from magnetic field data to target location, without requiring an explicit analytical expression of the field-position relationship. Therefore, there is an urgent need to propose a magnetic target localization scheme that combines physical models with deep learning methods to address the technical problem of existing physical localization methods being susceptible to interference from external time-varying noise. Summary of the Invention

[0005] This invention combines physical scene analysis with deep learning technology to propose a novel hierarchical neural network architecture for magnetic target localization. First, sub-network layers are used to capture local abrupt changes and morphologies in the signal waveform. Then, a global fusion network layer is used to model the overall dynamic process. This hierarchical design achieves intelligent mapping from raw data to position coordinates, providing a completely new method for high-precision magnetic positioning.

[0006] To achieve the above objectives, this invention proposes a method for intelligent localization of magnetic targets based on a hierarchical neural network architecture, comprising: Acquire multi-point magnetic field data measured by a magnetic sensor array; Based on the magnetic field data and the spatial location of the sensor, the magnetic gradient tensor data is calculated. The magnetic gradient tensor data is input into a pre-constructed hierarchical neural network model; the hierarchical neural network model includes an input layer, a sub-network layer, a global fusion network layer, and an output layer connected in sequence. The sub-network layer consists of multiple parallel one-dimensional convolutional neural network branches, used to extract local spatiotemporal features from the magnetic gradient tensor time series data of each measurement point; The global fusion network layer is composed of a long short-term memory network, which is used to perform temporal fusion and global context modeling on the feature sequences output by all sub-network branches. The output layer outputs the three-dimensional spatial coordinates of the magnetic target based on the fused global features.

[0007] Furthermore, the magnetic sensor array is a cubic array composed of eight triaxial magnetometers, with a fixed spacing between adjacent sensors, used to synchronously acquire magnetic field vector data at multiple points in space.

[0008] Furthermore, each one-dimensional convolutional neural network branch in the sub-network layer also includes a fully connected layer, which is used to perform nonlinear combination and dimensionality reduction on the extracted local features to generate a compact feature encoding vector for the measurement point.

[0009] Furthermore, the global fusion network layer utilizes the gating mechanism of LSTM to sequentially process the feature vectors of each spatial point, learning the collaborative evolution law of features at different spatial locations over time.

[0010] Furthermore, the magnetic gradient tensor data is a nine-element matrix calculated based on the spatial rate of change of the three components of the magnetic field vector in each orthogonal direction, which is used to suppress background geomagnetic field interference and enhance the target magnetic field characteristics.

[0011] On the other hand, to achieve the above objectives, this invention proposes a magnetic target intelligent positioning system based on a hierarchical neural network architecture, comprising: Magnetic sensor array, used to collect magnetic field data at multiple points in space; The data acquisition and conversion circuit is used to convert magnetic field data into digital signals and calculate the magnetic gradient tensor. An embedded or host computer processing unit loads and executes the magnetic target intelligent positioning method as described in any one of claims 1 to 5, and outputs target position information.

[0012] Furthermore, the magnetic sensor array is a cubic array consisting of eight triaxial magnetometers, with a sensor spacing of 30 cm, a sampling rate of 5 kS / s, and an analog-to-digital conversion accuracy of 24 bits.

[0013] Furthermore, the hierarchical neural network model deployed in the processing unit is a pre-trained model, and the training data includes magnetic gradient tensor data at different distances and angles and their corresponding real target positions. Compared with the prior art, the beneficial effects of the present invention are as follows: This invention designs a magnetic target localization method and system that combines physical analysis and deep learning. For model input, the magnetic gradient tensor is chosen as the input to the deep learning model. By representing the rate of change of the magnetic field vector space, the magnetic gradient tensor provides richer spatial information and suppresses uniform interference from sources such as the background geomagnetic field. The neural network model employs a hierarchical neural network architecture, utilizing sub-network layers to extract local spatiotemporal features of each measurement point, and then using a global fusion network to learn the global temporal evolution of these features. The implementation of this invention represents a deep integration of physical analysis methods and data-driven methods. This method can guide the mapping relationship between the magnetic gradient tensor and the magnetic target position in the deep learning model, while retaining its strong nonlinear fitting capability, thereby enhancing the accuracy of magnetic target localization and effectively solving the problem that magnetic target localization methods based on physical model analysis are susceptible to external time-varying noise interference. Attached Figure Description

[0014] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a schematic diagram of the magnetic sensor array in an embodiment of the present invention; Figure 2 This is a schematic diagram of the hierarchical neural network architecture in an embodiment of the present invention; Figure 3 This is a schematic diagram of the magnetic target localization error distance of the 1D CNN network in an embodiment of the present invention; Figure 4 This is a schematic diagram of the magnetic target positioning error distance of the FCL network in an embodiment of the present invention; Figure 5This is a schematic diagram of the magnetic target localization error distance of the LSTM network in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the magnetic target positioning error distance of the hierarchical neural network in an embodiment of the present invention; Figure 7 This is a schematic diagram of the magnetic target positioning process in an embodiment of the present invention; Figure 8 This is a physical diagram of the measurement system in an embodiment of the present invention; Figure 9 This is a schematic diagram of the system connection in an embodiment of the present invention. Detailed Implementation

[0015] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0016] When the distance between a magnetic target and the observation point is greater than 2.5 times the target's maximum size, the magnetic target can be equivalent to a magnetic dipole. In the development of magnetic anomaly detection technology, the magnetic dipole is an important equivalent model for magnetic sources. According to the Biot-Saffar law, a magnetic dipole in position... The magnetic field vector generated at that location It can be represented as: (1) in, This is the relative position vector from the target to the observation point. The magnetic moment vector of the magnetic dipole. is the vacuum permeability.

[0017] The magnetic gradient tensor is defined as the spatial rate of change of the three components (Bx, By, Bz) of the magnetic field vector along each orthogonal direction, consisting of nine elements. The magnetic gradient tensor matrix G can be expressed as: (2) In a source-free space, according to Maxwell's equations of magnetostatics, the curl and divergence of the magnetic field disappear, that is: (3) (4) Therefore, the magnetic gradient tensor matrix has the following properties, namely (5) (6) (7) (8) like Figure 1 As shown, the distance between any two adjacent sensors is l. An array of 8 magnetic sensors forms a cube. Let the center of the cube be O, the front be the +X face, the back be the -X face, the right be the +Y face, the left be the -Y face, the bottom be the +Z face, and the top be the -Z face.

[0018] The magnetic sensor array consists of eight triaxial magnetometers arranged in a cube, with a spacing of 30 cm between the sensors. The magnetometers used are Huashun HS-MAG03MS, with a noise density better than 10 PT RMS / √HZ@1 HZ, a linearity error less than 0.005% of full scale (FS), and an orthogonality error between the three axes less than 0.1 HZ. The analog voltage output of each magnetometer is connected to six National Instruments (NI) 9239 modules, which are housed in an NI CDAQ-9189 Ethernet chassis. Each NI 9239 module provides a 24-bit analog-to-digital converter (ADC), and the sampling rate of all channels is configured to 5 KS / s in this design. The CDAQ-9189 chassis is connected to the host computer via Ethernet. A custom data acquisition program developed using NI LabVIEW software is used on the host computer to synchronously control all acquisition channels. A physical diagram of the measurement system is shown below. Figure 5 As shown.

[0019] like Figure 2 As shown, the hierarchical neural network model of the present invention adopts a collaborative architecture of "local feature extraction-global temporal fusion", which includes an input layer, a sub-network layer, a global fusion network layer and an output layer in sequence, so as to achieve accurate mapping from the magnetic gradient tensor to the target position.

[0020] The system consists of several layers: an input layer receives temporal data of the magnetic gradient tensor; a sub-network layer contains multiple parallel one-dimensional convolutional neural network branches to extract local features from data at various spatial measurement points; a global fusion network layer performs temporal fusion and contextual modeling of the feature sequences from all sub-network branches; and an output layer outputs the target location information.

[0021] Specifically, the input layer converts the magnetic vector data into magnetic gradient tensor data of each measurement surface according to formula (2), which is then used as the model input.

[0022] The sub-network layer consists of a 1D CNN (One-Dimensional Convolutional Neural Network) and an FCL (Fully Connected Layer) network, processing temporal data from a single spatial measurement point. The 1D CNN slides along the time dimension, using its convolutional kernels to automatically and efficiently extract the local spatial-temporal pattern of the magnetic gradient tensor of the measurement point as a function of time. The fully connected network is used to nonlinearly combine and reduce the dimensionality of the extracted local high-level features, forming a compact feature encoding vector for the measurement point.

[0023] The global fusion network layer is composed of LSTM (Long Short-Term Memory) networks, which utilize their internal gating mechanisms (forget gate, input gate, output gate) to process the feature sequence sequentially. It can dynamically learn and remember the patterns of co-evolution of local features at different spatial locations over time.

[0024] The output layer receives information from the global fusion network layer and decodes the high-dimensional global features into target position coordinates. These directly correspond to the three-dimensional spatial coordinates (x, y, z) of the magnetic target in a preset coordinate system.

[0025] The implementation of the above method mainly includes the following steps: S1: Acquire spatial multi-point magnetic field data measured by the sensor array; S2: Calculate the magnetic gradient tensor data based on the magnetic field data and the spatial position of the sensor; S3: Input the magnetic gradient tensor data into a hierarchical neural network model; the hierarchical neural network model includes, in sequence, a one-dimensional convolutional neural network layer for extracting spatial features, and a long short-term memory network layer for modeling temporal dependencies; S4: Determine the position information of the magnetic target based on the output of the hierarchical neural network model.

[0026] The positioning system consists of a sensor array, data acquisition and conversion circuitry, and a power supply. The magnetic sensor array comprises eight triaxial magnetometers arranged in a cube, with a spacing of 30 cm between the sensors. The magnetometers used are Huashun HS-MAG03MS, with a noise density better than 10 PT RMS / √HZ@1 HZ, a linearity error less than 0.005% of full scale (FS), and an orthogonality error between the three axes less than 0.1 HZ. The analog voltage output of each magnetometer is connected to six National Instruments (NI) 9239 modules, which are housed in an NI CDAQ-9189 Ethernet chassis. Each NI 9239 module provides a 24-bit analog-to-digital converter (ADC), and the sampling rate of all channels is configured to 5 KS / s in this design. The CDAQ-9189 chassis is connected to the host computer via Ethernet. A custom data acquisition program developed using NI LabVIEW software is used on the host computer to synchronously control all acquisition channels. The specific experimental procedure is designed as follows: 1. Data Acquisition and Hardware System Array and Sensors: Eight Huashun HS-MAG03ms triaxial magnetometers were used to form a cube array with a side length of 30 cm.

[0027] Signal acquisition: A total of 24 analog signals are acquired simultaneously by six NI 9239 modules (integrated in the NI cDAQ-9189 chassis), with a sampling rate of 5 kS / s and an ADC resolution of 24 bits.

[0028] Control and storage: Connected to a host computer via Ethernet, synchronous acquisition and data recording are controlled by a program written in NI LabVIEW.

[0029] Measurement scheme: Within a semi-cylindrical measurement area (radius 200cm, height 57cm), place the target at preset distances (50-200cm, step size 10cm) and angles (0°, 45°, 90°, 145°, 180°) to generate a total of 80 marker points.

[0030] 2. Data Processing and Input Construction Magnetic gradient tensor calculation: Convert the magnetic vector data of the sensor into a magnetic gradient tensor.

[0031] Feature input: Input the magnetic gradient tensor into the constructed neural network.

[0032] 3. Network Design and Output Sub-network: Combining 1D CNN with FCL.

[0033] Converged network: The output of the sub-network is combined with the LSTM network. Performance evaluation: The predicted coordinates are compared with the actual coordinates, and the RMSE value is calculated as an error index.

[0034] To verify the effectiveness and superiority of the hierarchical neural network architecture proposed in this invention, a comparative experiment was designed. Under the same dataset, the same preprocessing procedure, and the same training conditions, a baseline model containing only a single network and the complete model proposed in this invention were trained and tested respectively. The dataset was randomly divided into training and test sets in an 8:2 ratio. The localization error was evaluated using the root mean square error (RMSE, unit: cm).

[0035] like Figure 6 As shown, the RMSE value of the magnetic target localization error distance using only a 1D CNN network is 0.0575 cm, and the three component errors are 0.0283 cm, 0.0413 cm, and 0.0317 cm, respectively.

[0036] like Figure 7As shown, the RMSE value of the magnetic target positioning error distance using only the FCL network is 0.1167 cm, and the three component errors are 0.0566 cm, 0.0895 cm, and 0.05 cm, respectively.

[0037] like Figure 8 As shown, the RMSE value of the magnetic target localization error distance using only the LSTM network is 0.1110 cm, and the three component errors are 0.0671 cm, 0.0755 cm, and 0.047 cm, respectively.

[0038] like Figure 9 As shown, the hierarchical neural network architecture proposed in this invention is used, which utilizes 1D CNN and FCL to form sub-networks and LSTM network for fusion network. The RMSE value of the magnetic target localization error distance is 0.0317 cm, and the three component errors are 0.0174 cm, 0.02 cm, and 0.02 cm, respectively.

[0039] In summary, the hierarchical architecture proposed in this invention achieves optimal localization accuracy on the test set. Its overall localization error is significantly lower than all single-network benchmark models. The sub-network layers and the global fusion network layer are combined in a hierarchical manner, producing a synergistic enhancement effect. This strongly demonstrates that the architectural design of this invention, which separates local feature extraction from global temporal fusion, can achieve high-precision prediction of magnetic targets.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for intelligent localization of magnetic targets based on a hierarchical neural network architecture, characterized in that, include: Acquire multi-point magnetic field data measured by a magnetic sensor array; Based on the magnetic field data and the spatial location of the sensor, the magnetic gradient tensor data is calculated. The magnetic gradient tensor data is input into a pre-constructed hierarchical neural network model; the hierarchical neural network model includes an input layer, a sub-network layer, a global fusion network layer, and an output layer connected in sequence. The sub-network layer consists of multiple parallel one-dimensional convolutional neural network branches, used to extract local spatiotemporal features from the magnetic gradient tensor time series data of each measurement point; The global fusion network layer is composed of a long short-term memory network, which is used to perform temporal fusion and global context modeling on the feature sequences output by all sub-network branches. The output layer outputs the three-dimensional spatial coordinates of the magnetic target based on the fused global features.

2. The method according to claim 1, characterized in that, The magnetic sensor array is a cubic array consisting of eight triaxial magnetometers with a fixed spacing between adjacent sensors, used to synchronously collect magnetic field vector data from multiple points in space.

3. The method according to claim 1, characterized in that, Each one-dimensional convolutional neural network branch in the sub-network layer also includes a fully connected layer, which is used to perform nonlinear combination and dimensionality reduction on the extracted local features to generate a compact feature encoding vector for the measurement point.

4. The method according to claim 1, characterized in that, The global fusion network layer uses the gating mechanism of LSTM to process the feature vectors of each spatial point sequentially, and learns the cooperative evolution law of features at different spatial locations over time.

5. The method according to claim 1, characterized in that, The magnetic gradient tensor data is a nine-element matrix calculated based on the spatial rate of change of the three components of the magnetic field vector in each orthogonal direction. It is used to suppress background geomagnetic field interference and enhance the target magnetic field characteristics.

6. A magnetic target intelligent positioning system based on a hierarchical neural network architecture, characterized in that, include: Magnetic sensor array, used to collect magnetic field data at multiple points in space; The data acquisition and conversion circuit is used to convert magnetic field data into digital signals and calculate the magnetic gradient tensor. An embedded or host computer processing unit loads and executes the magnetic target intelligent positioning method as described in any one of claims 1 to 5, and outputs target position information.

7. The system according to claim 6, characterized in that, The magnetic sensor array is a cubic array consisting of eight triaxial magnetometers with a sensor spacing of 30 cm, a sampling rate of 5 kS / s, and an analog-to-digital conversion accuracy of 24 bits.

8. The system according to claim 6, characterized in that, The hierarchical neural network model deployed in the processing unit is a pre-trained model, and the training data includes magnetic gradient tensor data at different distances and angles and their corresponding real target positions.

Images