Underwater target multi-sensor fusion magnetic field measurement system in complex magnetic environment

Abstract

This invention provides a multi-sensor fusion magnetic field measurement system for underwater targets in complex magnetic environments, relating to the field of marine equipment technology. It includes a multi-modal magnetic field sensing module, an attitude correlation correction module, a hierarchical fusion processing module, and an environment adaptation and magnetic anomaly detection module. The multi-modal magnetic field sensing module is responsible for collecting multi-dimensional and multi-type raw magnetic field data around the underwater vehicle, constructing a complete magnetic field observation foundation. The attitude correlation correction module is used to eliminate the influence of changes in the underwater vehicle's motion attitude on magnetic field measurement. The hierarchical fusion processing module performs hierarchical fusion of multi-sensor data. The environment adaptation and magnetic anomaly detection module realizes background magnetic field separation, interference suppression, and magnetic anomaly target identification. This system can improve the accuracy and robustness of magnetic field measurement for underwater vehicles in complex magnetic environments, achieving highly reliable magnetic field feature extraction and situational awareness functions.

Description

Technical Field

[0001] This invention relates to the field of marine equipment technology, specifically to a multi-sensor fusion magnetic field measurement system for underwater targets in complex magnetic environments. Background Technology

[0002] Underwater vehicles play a vital role in marine resource exploration, underwater target detection, and marine environmental monitoring. Magnetic field measurement, as an important means of situational awareness and target detection for underwater vehicles, can detect underwater magnetic targets and assess the magnetic characteristics of the underwater vehicle itself. However, underwater magnetic field measurements face complex magnetic environmental interferences, including spatial and temporal variations in the geomagnetic background field, magnetic interference generated by the underwater vehicle's own equipment, attitude disturbances caused by ocean currents, and external electromagnetic noise. These interfering factors are coupled and severely affect the accuracy and reliability of magnetic field measurements.

[0003] Traditional magnetic field measurement systems typically employ a single type of magnetometer, making it difficult to fully utilize the advantages of different sensors. Triaxial magnetometers can measure vector information of the magnetic field but are susceptible to installation attitude errors; scalar magnetometers offer high measurement accuracy and are unaffected by direction but cannot provide directional information. Single sensors lack robustness in the face of complex interference, making them prone to measurement anomalies. Furthermore, underwater vehicles generate six degrees of freedom motion during underwater navigation, including pitch, roll, yaw, forward, lateral, and heave. These motions result in a dynamically changing attitude relationship between the sensor coordinate system and the geographic coordinate system, introducing attitude-related system biases into magnetic field measurements. Existing attitude compensation methods are mostly based on static platform assumptions, making it difficult to accurately compensate for measurement biases caused by the dynamic motion of underwater vehicles.

[0004] In terms of multi-sensor data fusion, existing technologies mainly include the following methods:

[0005] Weighted average fusion is the most basic multi-sensor fusion method. It assigns fusion weights based on the noise variance of each sensor, with the weights proportional to the reciprocal of the noise variance. This method is computationally simple, has good real-time performance, and can achieve good fusion results when sensor performance is stable and anomalies are absent. However, this method assumes that all sensors are working properly. When sensor malfunctions or local interference occur, abnormal data can contaminate the fusion results, leading to a decrease in overall measurement accuracy. Furthermore, this method cannot utilize the physical properties of magnetic fields for constraint, resulting in insufficient robustness in complex magnetic environments and a lack of ability to suppress outliers.

[0006] Kalman filter fusion algorithms achieve multi-sensor data fusion by establishing system state equations and observation equations and employing a prediction-correction recursive approach. This method can handle dynamic systems, improve estimation accuracy by utilizing time series information, and assess estimation uncertainty through the covariance matrix. Extended Kalman filter and unscented Kalman filter can handle nonlinear systems. However, Kalman filtering requires high model accuracy, and model errors can lead to filter divergence. In magnetic field measurements, the complex magnetic environment makes it difficult to establish accurate system models, and Kalman filtering assumes that noise follows a Gaussian distribution, resulting in poor robustness to outliers and non-Gaussian noise. When sensors experience sudden failures or are subjected to pulse interference, Kalman filtering struggles to quickly identify and suppress abnormal data. Furthermore, this method has high computational complexity, limiting real-time performance when dealing with a large number of sensors.

[0007] Signal enhancement methods based on stochastic resonance, which amplify weak signals by introducing appropriate noise, have found applications in signal processing. This method can extract useful signals under low signal-to-noise ratio conditions and has a certain enhancement effect on weak magnetic field signals. However, the stochastic resonance method requires precise adjustment of system parameters to match signal characteristics; improper parameter settings can lead to poor enhancement results or even introduce additional distortion. This method is mainly designed for enhancing single signals and is difficult to directly apply to multi-sensor spatial distributed measurement scenarios. Furthermore, the signal enhanced by stochastic resonance still requires further fusion processing, failing to address the issues of consistent data fusion and anomaly detection across multiple sensors.

[0008] Most existing multi-sensor fusion methods process different types of sensors separately, failing to fully utilize the complementary characteristics of vector magnetometers and scalar magnetometers. Vector magnetometers provide direction information, while scalar magnetometers provide high-precision amplitude information. The synergistic fusion of the two can simultaneously obtain accurate magnetic field direction and amplitude, but existing technologies lack an effective vector-scalar coupling fusion mechanism.

[0009] In sensor anomaly detection, traditional methods primarily rely on threshold decisions or statistical tests, comparing measured values ​​with preset thresholds or historical statistical characteristics to determine anomalies. These methods struggle to distinguish between genuine magnetic field changes and sensor malfunctions, easily leading to false positives. As a vector field, the magnetic field satisfies the fundamental constraints of Maxwell's equations: its divergence is zero in regions without magnetic charge, and its curl is zero in regions without current. These physical constraints can be used to verify the validity of measurement data; however, current technologies rarely apply the physical laws of electromagnetic fields to sensor anomaly detection and data quality assessment.

[0010] US Patent Application US20160338608A1 discloses a cardiac magnetic field measurement system. This system includes a cardiac magnetic field sensor for measuring a first and a second magnetic field, a noise magnetic sensor for measuring the second magnetic field, and a magnetic measurement device for calculating an approximation of the second magnetic field using measurements from the noise magnetic sensor and a multivariate polynomial. The magnetic measurement device subtracts the approximation of the second magnetic field from the measurements from the cardiac magnetic field sensor to achieve noise suppression. The shortcomings of this patent application are: the system is designed for static medical measurement scenarios and does not consider the impact of dynamic motion platforms, such as underwater vehicles, on magnetic field measurement, lacking an attitude-related compensation mechanism; the background magnetic field suppression method in this patent is based on multivariate polynomial modeling, failing to consider the physical constraints of the magnetic field, resulting in limited model adaptability in complex time-varying magnetic environments; the patent uses simple subtraction operations for background field suppression, lacking a hierarchical fusion architecture and robust handling of sensor anomalies; the patent does not utilize magnetic field gradient tensor information, failing to extract the spatial distribution characteristics and topological structure information of the magnetic field; and the fusion method in this patent does not distinguish between the different characteristics of vector magnetometers and scalar magnetometers, lacking a vector-scalar coupling fusion mechanism.

[0011] Chinese patent application CN120597207A discloses a magnetic field measurement method and system based on multi-sensor fusion technology. The method includes deploying a multi-sensor data array, enhancing weak magnetic field signals through a bistable stochastic resonance system and MPA population optimization algorithm, compensating for Abbe error and bidirectional projection error by calculating the array tilt angle, and performing data fusion using extended Kalman filtering after quality assessment and weight allocation. The shortcomings of this patent application are as follows: The method uses random resonance for signal enhancement, but random resonance is mainly suitable for detecting weak signals in a single channel. In multi-sensor spatially distributed measurement scenarios, it is difficult to guarantee the spatial consistency of the enhanced signals from each sensor, potentially introducing additional spatial mismatch errors. The patent uses extended Kalman filtering for fusion, but it lacks a robust handling mechanism for sensor anomalies and outliers. When sensors malfunction or are subjected to strong local interference, the fusion performance deteriorates significantly. The error compensation in this patent mainly targets array geometric errors, without establishing a mapping compensation model between the dynamic motion state of the underwater vehicle and the magnetic field measurement deviation, and without considering adaptive filtering strategies under different motion modes. This patent does not utilize the physical constraint characteristics of the magnetic field gradient tensor for data quality assessment and lacks a sensor anomaly detection method based on divergence and curl constraints. This patent does not establish a coupled fusion model between vector magnetometers and scalar magnetometers, failing to fully leverage the respective advantages of the two types of sensors in direction and amplitude measurement. This patent lacks a physically constraint-guided iterative optimization mechanism, making it difficult to guarantee the physical consistency of the fusion results. Summary of the Invention

[0012] The purpose of this invention is to address the shortcomings by proposing a multi-sensor fusion magnetic field measurement system for underwater targets in complex magnetic environments.

[0013] The present invention adopts the following technical solution:

[0014] A multi-sensor fusion magnetic field measurement system for underwater targets in complex magnetic environments includes a multi-modal magnetic field sensing module, an attitude correlation correction module, a hierarchical fusion processing module, and an environment adaptation and magnetic anomaly detection module.

[0015] The multimodal magnetic field sensing module is responsible for collecting multi-dimensional and multi-type raw magnetic field data around the underwater vehicle to build a complete magnetic field observation foundation. The attitude correlation correction module is used to eliminate the influence of changes in the underwater vehicle's motion attitude on magnetic field measurement. The hierarchical fusion processing module performs hierarchical fusion of multi-sensor data. The environmental adaptation and magnetic anomaly detection module realizes background magnetic field separation, interference suppression, and magnetic anomaly target identification.

[0016] The multimodal magnetic field sensing module includes a vector-scalar collaborative measurement unit, a magnetic field spatial gradient tensor construction unit, and a magnetic interference spectrum real-time monitoring unit. The vector-scalar collaborative measurement unit is used to realize the synchronous measurement of magnetic field direction and intensity. The magnetic field spatial gradient tensor construction unit is used to construct a complete magnetic field gradient tensor. The magnetic interference spectrum real-time monitoring unit monitors magnetic interference signals across the entire frequency band in real time.

[0017] The attitude correlation correction module includes a magnetic-inertial fusion attitude calculation unit, an attitude-related magnetic deviation compensation unit, and a motion state adaptive filtering unit. The magnetic-inertial fusion attitude calculation unit is used to achieve high-precision attitude estimation of the underwater vehicle. The attitude-related magnetic deviation compensation unit dynamically compensates for measurement deviations caused by attitude changes based on real-time attitude information. The motion state adaptive filtering unit is used to identify the current motion mode of the underwater vehicle and adaptively switch the corresponding filtering strategy according to the noise characteristics of different motion states.

[0018] The hierarchical fusion processing module includes a sensor-level data pre-fusion unit, a feature-level multi-domain fusion unit, and a decision-level intelligent fusion unit. The sensor-level data pre-fusion unit performs primary fusion of measurement data from sensors of the same type. The feature-level multi-domain fusion unit is used to fuse extracted multi-domain features to obtain a more comprehensive description of magnetic field features. The decision-level intelligent fusion unit selects the optimal fusion result as the system output based on the credibility evaluation of the fusion result.

[0019] The environmental adaptation and magnetic anomaly detection module includes an online background magnetic field modeling and separation unit, a multi-source interference identification and suppression unit, and a magnetic anomaly target intelligent identification unit. The online background magnetic field modeling and separation unit constructs and updates the local background magnetic field model of the underwater vehicle's operating area in real time, effectively separating the background magnetic field from the target magnetic field signal. The multi-source interference identification and suppression unit is used to identify and distinguish multi-source interference signals and adopt targeted suppression strategies to reduce the impact of interference. The magnetic anomaly target intelligent identification unit detects and identifies magnetic anomaly targets in the environment based on extracting the magnetic characteristics of the underwater vehicle itself.

[0020] Furthermore, the magnetic field spatial gradient tensor construction unit includes a multi-point sampling coordinator, a gradient calculation processor, and a tensor reconstruction processor. The multi-point sampling coordinator controls the distributed magnetometer array to perform time-synchronized sampling to ensure the time consistency of data from each measuring point. The gradient calculation processor calculates the first-order gradient components of the magnetic field in each spatial direction based on the magnetic field difference between adjacent measuring points. The tensor reconstruction processor combines the calculated gradient components to construct a complete magnetic field gradient tensor matrix, forming a mathematical description of the spatial distribution of the magnetic field.

[0021] The tensor reconstruction processor constructs a dynamic gradient tensor according to the following formula:

[0022] ;

[0023] in, This represents the dynamic gradient tensor of the nth measurement point at time t. Let represent the static gradient tensor of the nth measurement point at time t. This is the motion-induced correction term for the nth measuring point at time t;

[0024] The static gradient tensor is directly composed of gradient information calculated by the gradient calculation processor, and the motion-induced correction term is calculated according to the following formula:

[0025] ;

[0026] Where v(t) represents the velocity vector of the underwater vehicle at time t, and c m B is the characteristic velocity of the magnetic field response. n Let w(t) be the magnetic field at the nth measuring point, and w(t) be the angular velocity vector of the underwater vehicle. It is an antisymmetric matrix composed of angular velocities.

[0027] Furthermore, the intelligent identification unit for magnetic anomalies includes a self-magnetic feature extraction processor, an anomaly detection processor, and a target classification and recognition device. The self-magnetic feature extraction processor analyzes the magnetic field characteristics generated by the underwater vehicle itself and establishes a self-magnetic feature template for distinguishing it from the environmental magnetic field. The anomaly detection processor detects abnormal magnetic signals that deviate from the background magnetic field and the self-magnetic features through statistical testing, threshold judgment, or machine learning methods. The target classification and recognition device performs feature analysis and pattern recognition on the detected magnetic anomalies to determine the type, location, and motion state of the abnormal target, thereby achieving intelligent identification of potential magnetic targets.

[0028] The anomaly monitoring processor calculates the overall anomaly degree of the sensor according to the following formula:

[0029] ;

[0030] ;

[0031] ;

[0032] in, This represents the overall anomaly degree of the nth sensor at time t. Let n be the amplitude anomaly of the nth sensor at time t. Let be the topological anomaly degree of the nth sensor at time t. w is the physical consistency index for the nth measurement point. amp w topo w phys These are the weighting coefficients for the three types of anomalies. Let be the curvature of the magnetic field lines at the nth measuring point at time t. Let the topological feature symbol of the nth measurement point at time t be denoted as . , Here, tr() is the sensitivity parameter, and tr() is the trace function. For divergence-constrained sensitivity parameters, This is the curl constraint sensitivity parameter.

[0033] Furthermore, the sensor-level data pre-fusion unit includes a data synchronization processor, a similar sensor fusion processor, and a reliability evaluation processor. The data synchronization processor performs time alignment and coordinate system unification on the measurement data of similar sensors to eliminate sampling time differences. The similar sensor fusion processor is used to fuse data from similar sensors and improve measurement accuracy by utilizing redundant information. The reliability evaluation processor evaluates the reliability of the fusion result based on the noise level, consistency, and historical performance of each sensor.

[0034] The same type of sensor fusion processor fuses sensor data according to the following formula:

[0035] ;

[0036] in, Let N be the fused magnetic field vector at time t. v N represents the number of vector magnetometers. s For the number of scalar magnetometers, Let be the measurement value of the nth vector magnetometer at time t. Let be the measurement value of the m-th scalar magnetometer at time t. (t) represents the fusion weight of the nth vector magnetometer. (t) represents the fusion weight of the m-th scalar magnetometer. This is the scalar constraint strength coefficient.

[0037] Furthermore, the decision-level intelligent fusion unit includes a fusion result evaluation processor, a decision rule base, and an intelligent selection processor. The fusion result evaluation processor receives the credibility index of the sensor-level fusion result and calculates the physical consistency and information richness of the optimization result after iterative optimization, evaluating the quality improvement brought about by the optimization. The decision rule base stores a set of decision rules based on expert knowledge and historical experience, defining the optimal fusion result selection strategy under different scenarios. The intelligent selection processor performs iterative optimization of the sensor-level fusion result guided by physical constraints according to the evaluation result and decision rules, and selects the optimal fusion output. It can dynamically adjust the optimization strategy and decision parameters according to environmental changes.

[0038] The intelligent selection processor iteratively processes the fusion result according to the following formula:

[0039] ;

[0040] in, This represents the fusion result of the k-th iteration. For learning rate, To ensure data fidelity and accuracy, H[B] represents the physical constraint error, and H[B] represents the information entropy of the fused magnetic field data. , For two regularization parameters;

[0041] The result of each iteration is used as a candidate fusion result.

[0042] The beneficial effects achieved by this invention are:

[0043] This system constructs a dynamic gradient tensor model that considers the effects of underwater vehicle velocity and angular velocity. By introducing a motion-induced correction term based on the static gradient tensor, it accurately compensates for the distortion of magnetic field spatial gradient measurements caused by the underwater vehicle's dynamic motion. By calculating the curvature and topological feature sign of magnetic field lines using the dynamic gradient tensor, and combining divergence and curl constraints to evaluate physical consistency, a sensor anomaly detection method based on magnetic field topological features and physical laws is established. Compared to traditional threshold decision and statistical testing methods, this method can more accurately identify sensor faults and local interference, reducing the false positive rate. Furthermore, by establishing a coupled fusion optimization model of vector magnetometers and scalar magnetometers, and using scalar constraints to correct the amplitude of the vector fusion result, the system achieves the synergistic utilization of direction and amplitude information. Through a physically constraint-guided iterative optimization algorithm, a divergence constraint error term and an information entropy maximization term are introduced on top of the data fidelity term, achieving optimized correction of the fusion result while satisfying the fundamental laws of electromagnetism. Compared to simple data fitting methods, this method ensures the physical consistency of the fusion result and maintains information richness, avoiding the loss of effective information due to excessive smoothing.

[0044] To further understand the features and technical content of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are for reference and illustration only and are not intended to limit the present invention. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the overall structural framework of the present invention;

[0046] Figure 2 This is a schematic diagram illustrating the construction process of the dynamic gradient tensor of this invention and the collaborative workflow of each processor;

[0047] Figure 3 This is a schematic diagram of the sensor anomaly detection process of the present invention;

[0048] Figure 4 This is a schematic diagram of the vector-scalar coupling fusion process of the present invention;

[0049] Figure 5 This is a schematic diagram of the decision-level intelligent fusion and iterative optimization process of the present invention;

[0050] Figure 6 This is a schematic diagram comparing the anomaly detection performance of the present invention with other methods;

[0051] Figure 7 This is a schematic diagram comparing the accuracy of the present invention with other methods under different motion modes;

[0052] Figure 8 This is a screenshot of the interactive interface information section of this invention. Detailed Implementation

[0053] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention. Furthermore, the accompanying drawings of the present invention are for simple illustrative purposes only and are not depictions of actual dimensions; this is stated beforehand. The following embodiments will further describe the relevant technical content of the present invention in detail, but the disclosed content is not intended to limit the scope of protection of the present invention.

[0054] Example 1: Multi-sensor fusion magnetic field measurement system for underwater targets in complex magnetic environments, combined with Figure 1 It includes a multimodal magnetic field sensing module, an attitude correlation correction module, a hierarchical fusion processing module, and an environment adaptation and magnetic anomaly detection module;

[0055] The multimodal magnetic field sensing module is responsible for collecting multi-dimensional and multi-type raw magnetic field data around the underwater vehicle to build a complete magnetic field observation foundation. The attitude correlation correction module is used to eliminate the influence of changes in the underwater vehicle's motion attitude on magnetic field measurement and ensure the accuracy of measurement data. The hierarchical fusion processing module performs hierarchical fusion of multi-sensor data and extracts high-reliability magnetic field feature information. The environment adaptation and magnetic anomaly detection module realizes background magnetic field separation, interference suppression, and magnetic anomaly target identification, improving the system's adaptability and target detection capability in complex magnetic environments.

[0056] The multimodal magnetic field sensing module includes a vector-scalar collaborative measurement unit, a magnetic field spatial gradient tensor construction unit, and a magnetic interference spectrum real-time monitoring unit. The vector-scalar collaborative measurement unit achieves synchronous measurement of magnetic field direction and intensity through the collaborative work of a triaxial magnetometer array and a scalar magnetometer. The two types of sensors mutually verify each other to improve measurement reliability. The magnetic field spatial gradient tensor construction unit constructs a complete magnetic field gradient tensor based on multi-point measurement data from a distributed magnetometer array, improving the sensing resolution of the spatial distribution characteristics of the magnetic field. The magnetic interference spectrum real-time monitoring unit monitors magnetic interference signals across the entire frequency band in real time and establishes an interference spectrum fingerprint database including low frequency, power frequency, and high frequency, providing a basis for subsequent interference suppression.

[0057] The attitude correlation correction module includes a magnetic-inertial fusion attitude calculation unit, an attitude-related magnetic deviation compensation unit, and a motion state adaptive filtering unit. The magnetic-inertial fusion attitude calculation unit combines data from the magnetometer and the inertial measurement unit and uses a fusion algorithm to achieve high-precision attitude estimation of the underwater vehicle, providing an accurate attitude reference for attitude compensation. The attitude-related magnetic deviation compensation unit establishes a mapping compensation model between the magnetic field measurement value and the attitude angle of the underwater vehicle, and dynamically compensates for the measurement deviation caused by attitude changes based on real-time attitude information. The motion state adaptive filtering unit identifies the current motion mode of the underwater vehicle, including cruise, hovering, and maneuvering states, and adaptively switches the corresponding filtering strategy according to the noise characteristics of different motion states.

[0058] The hierarchical fusion processing module includes a sensor-level data pre-fusion unit, a feature-level multi-domain fusion unit, and a decision-level intelligent fusion unit. The sensor-level data pre-fusion unit performs primary fusion of measurement data from similar sensors, forming highly reliable sensor-level fused data through redundant information complementarity. The feature-level multi-domain fusion unit extracts magnetic field feature parameters from multiple domains such as time domain, frequency domain, and spatial domain, and fuses the extracted multi-domain features to obtain a more comprehensive description of magnetic field features. The decision-level intelligent fusion unit evaluates the reliability of the fusion results and selects the optimal fusion result as the system output through an intelligent decision-making mechanism, thereby improving the robustness of the fusion processing.

[0059] The environmental adaptation and magnetic anomaly detection module includes an online background magnetic field modeling and separation unit, a multi-source interference identification and suppression unit, and a magnetic anomaly target intelligent identification unit. The online background magnetic field modeling and separation unit constructs and updates the local background magnetic field model of the underwater vehicle's operating area in real time, effectively separating the background magnetic field from the target magnetic field signal. The multi-source interference identification and suppression unit identifies and distinguishes multi-source interference signals from the underwater vehicle's own equipment, geomagnetic disturbances, and the external electromagnetic environment, and adopts targeted suppression strategies to reduce the impact of interference. The magnetic anomaly target intelligent identification unit detects and identifies magnetic anomaly targets in the environment based on the extraction of the underwater vehicle's own magnetic characteristics, thereby realizing situational awareness.

[0060] The vector-scalar collaborative measurement unit includes a triaxial magnetometer array, a scalar magnetometer group, and a collaborative verification processor. The triaxial magnetometer array consists of multiple triaxial magnetometers distributed around the underwater vehicle according to a predetermined spatial layout, used to measure the three-dimensional vector components of the magnetic field. The scalar magnetometer group consists of multiple scalar magnetometers, measuring the absolute value of the magnetic field strength, unaffected by the installation direction. The collaborative verification processor performs cross-verification on the vector measurement values ​​and scalar measurement values, detecting abnormal data and correcting them by comparing the vector amplitude with the scalar measurement results.

[0061] Combination Figure 2The magnetic field spatial gradient tensor construction unit includes a multi-point sampling coordinator, a gradient calculation processor, and a tensor reconstruction processor. The multi-point sampling coordinator controls the distributed magnetometer array to perform time-synchronized sampling to ensure the time consistency of data from each measuring point. The gradient calculation processor calculates the first-order gradient components of the magnetic field in each spatial direction based on the magnetic field difference between adjacent measuring points. The tensor reconstruction processor combines the calculated gradient components to construct a complete magnetic field gradient tensor matrix, forming a mathematical description of the spatial distribution of the magnetic field.

[0062] The tensor reconstruction processor constructs a dynamic gradient tensor according to the following formula:

[0063] ;

[0064] in, This represents the dynamic gradient tensor of the nth measurement point at time t. Let represent the static gradient tensor of the nth measurement point at time t. This is the motion-induced correction term for the nth measuring point at time t;

[0065] The static gradient tensor is directly composed of gradient information calculated by the gradient calculation processor, and the motion-induced correction term is calculated according to the following formula:

[0066] ;

[0067] Where v(t) represents the velocity vector of the underwater vehicle at time t, and c m B is the characteristic velocity of the magnetic field response. n Let w(t) be the magnetic field at the nth measuring point, and w(t) be the angular velocity vector of the underwater vehicle. The antisymmetric matrix is ​​formed by angular velocities;

[0068] The real-time magnetic interference spectrum monitoring unit includes a wideband magnetic sensor, a spectrum analysis processor, and an interference fingerprint database. The wideband magnetic sensor has wideband response characteristics and can capture magnetic interference signals from low frequency to high frequency. The spectrum analysis processor performs spectrum analysis such as fast Fourier transform on the collected interference signals to extract the frequency components and amplitude characteristics of the interference signals. The interference fingerprint database stores spectrum feature templates of different interference sources for rapid identification and classification of interference sources.

[0069] The magneto-inertial fusion attitude calculation unit includes an inertial measurement unit, an attitude fusion processor, and an attitude output interface. The inertial measurement unit integrates a three-axis gyroscope and a three-axis accelerometer to measure the angular velocity and linear acceleration information of the underwater vehicle. The attitude fusion processor uses an extended Kalman filter or complementary filter algorithm to fuse magnetometer data and inertial measurement data to calculate the real-time attitude angle of the underwater vehicle. The attitude output interface outputs the calculated attitude information in a standard format to other units for use, realizing the sharing of attitude data.

[0070] The attitude-related magnetic deviation compensation unit includes a deviation model establishment processor, an attitude mapping lookup table, and a dynamic compensation processor. The deviation model establishment processor establishes a mathematical mapping relationship model between the magnetic field measurement deviation and the attitude angle of the underwater vehicle through calibration experiments. The attitude mapping lookup table stores the compensation parameters corresponding to different attitude angles and supports fast lookup and interpolation calculation. The dynamic compensation processor obtains the compensation parameters from the lookup table according to the current attitude angle and performs real-time compensation and correction on the original magnetic field measurement value.

[0071] The motion state adaptive filtering unit includes a motion pattern recognition processor, a filter parameter library, and an adaptive filtering processor. The motion pattern recognition processor determines whether the underwater vehicle is currently cruising, hovering, or maneuvering based on parameters such as speed, acceleration, and attitude change rate. The filter parameter library stores corresponding optimal filter parameter configurations for different motion modes, including filter order and cutoff frequency. The adaptive filtering processor loads corresponding parameters from the parameter library according to the identified motion mode and dynamically adjusts the filtering strategy to adapt to the noise characteristics of the current motion state.

[0072] Combination Figure 4 The sensor-level data pre-fusion unit includes a data synchronization processor, a similar sensor fusion processor, and a reliability evaluation processor. The data synchronization processor performs time alignment and coordinate system unification on the measurement data of similar sensors to eliminate sampling time differences. The similar sensor fusion processor is used to fuse data from similar sensors and improve measurement accuracy by utilizing redundant information. The reliability evaluation processor evaluates the reliability of the fusion result based on the noise level, consistency, and historical performance of each sensor.

[0073] The same type of sensor fusion processor fuses sensor data according to the following formula:

[0074] ;

[0075] in, Let N be the fused magnetic field vector at time t. v N represents the number of vector magnetometers. s For the number of scalar magnetometers, Let be the measurement value of the nth vector magnetometer at time t. Let be the measurement value of the m-th scalar magnetometer at time t. (t) represents the fusion weight of the nth vector magnetometer. (t) represents the fusion weight of the m-th scalar magnetometer. The scalar constraint strength coefficient;

[0076] The fusion weight is calculated according to the following formula:

[0077] ;

[0078] ;

[0079] in, The basic weights of the nth vector magnetometer, h is the fundamental weight of the m-th scalar magnetometer. n (t) represents the health factor of the nth vector magnetometer, h m (t) represents the health factor of the m-th scalar magnetometer. Estimate the steepness parameters for M. The median of all vector magnetometer measurements. MAD is the median of all scalar magnetometer measurements. v MAD represents the absolute deviation of the median of the vector force gauge measurements. s The absolute deviation of the median of the scalar force gauge measurements;

[0080] Health factors are correlated with the overall abnormality of the magnetometer.

[0081] The feature-level multi-domain fusion unit includes a time-domain feature extraction processor, a frequency-domain feature extraction processor, and a spatial-domain feature extraction processor. The time-domain feature extraction processor extracts statistical features and time-varying patterns such as mean, variance, and peak value from the time series of the magnetic field signal. The frequency-domain feature extraction processor extracts the main frequency component, harmonic features, and frequency band energy distribution of the magnetic field signal through spectrum analysis. The spatial-domain feature extraction processor extracts spatial structure information such as spatial distribution pattern, gradient features, and symmetry of the magnetic field based on multi-point measurement data. The three types of feature extraction results are fused to form a multi-dimensional feature vector.

[0082] Combination Figure 5The decision-level intelligent fusion unit includes a fusion result evaluation processor, a decision rule base, and an intelligent selection processor. The fusion result evaluation processor receives the credibility index of the sensor-level fusion result and calculates the physical consistency and information richness of the optimization result after iterative optimization, evaluating the quality improvement brought about by the optimization. The decision rule base stores a set of decision rules based on expert knowledge and historical experience, defining the optimal fusion result selection strategy under different scenarios. The intelligent selection processor performs iterative optimization of the sensor-level fusion result guided by physical constraints according to the evaluation result and decision rules, and selects the optimal fusion output through an intelligent decision algorithm. It can dynamically adjust the optimization strategy and decision parameters according to environmental changes.

[0083] The intelligent selection processor iteratively processes the fusion result according to the following formula:

[0084] ;

[0085] in, This represents the fusion result of the k-th iteration. For learning rate, To ensure data fidelity and accuracy, H[B] represents the physical constraint error, and H[B] represents the information entropy of the fused magnetic field data. , For two regularization parameters;

[0086] The background magnetic field online modeling and separation unit includes a background sampling processor, a model update processor, and a background separation processor. The background sampling processor continuously collects regional background magnetic field data during the underwater vehicle's navigation and establishes a distributed background magnetic field sample set. The model update processor uses a recursive least squares or online learning algorithm to update the background magnetic field model parameters in real time based on the newly collected data, adapting to the slow time-varying characteristics of the background field. The background separation processor performs a difference operation between the current measured value and the background model prediction value to separate the target magnetic field signal components.

[0087] The multi-source interference identification and suppression unit includes an interference source classification processor, an interference feature matcher, and an adaptive suppression processor. The interference source classification processor identifies interference as interference from the underwater vehicle's own equipment, geomagnetic disturbances, or external electromagnetic interference based on the time-frequency characteristics, spatial distribution, and correlation analysis of the interference signal. The interference feature matcher matches the detected interference features with templates in the interference fingerprint database to determine the specific interference source type and location. The adaptive suppression processor selects an appropriate suppression strategy based on the identified interference type, including adaptive filtering, spatial projection, or time-domain subtraction, to achieve targeted interference suppression.

[0088] Combination Figure 3The intelligent identification unit for magnetic anomalies includes a self-magnetic feature extraction processor, an anomaly detection processor, and a target classification and recognition device. The self-magnetic feature extraction processor analyzes the magnetic field characteristics generated by the underwater vehicle itself and establishes a self-magnetic feature template for distinguishing it from the environmental magnetic field. The anomaly detection processor detects abnormal magnetic signals that deviate from the background magnetic field and the self-magnetic features through statistical testing, threshold judgment, or machine learning methods. The target classification and recognition device performs feature analysis and pattern recognition on the detected magnetic anomalies to determine the type, location, and motion state of the abnormal target, thereby achieving intelligent identification of potential magnetic targets.

[0089] The anomaly monitoring processor calculates the overall anomaly degree of the sensor according to the following formula:

[0090] ;

[0091] ;

[0092] ;

[0093] in, This represents the overall anomaly degree of the nth sensor at time t. Let n be the amplitude anomaly of the nth sensor at time t. Let be the topological anomaly degree of the nth sensor at time t. w is the physical consistency index for the nth measurement point. amp w topo w phys These are the weighting coefficients for the three types of anomalies. Let be the curvature of the magnetic field lines at the nth measuring point at time t. Let the topological feature symbol of the nth measurement point at time t be denoted as . , Here, tr() is the sensitivity parameter, and tr() is the trace function. For divergence-constrained sensitivity parameters, This refers to the curl constraint sensitivity parameter;

[0094] The curvature of the magnetic field lines and the degree of topological anomaly are calculated according to the following formula:

[0095] ;

[0096] ;

[0097] in, The magnetic field direction is represented by a unit vector, sgn() is the sign function, and det() represents the matrix determinant.

[0098] The 'm' and 'n' mentioned above are ordinal numbers used to represent sequence numbers and have no actual meaning.

[0099] Example 2: This example provides a specific implementation scheme for a multi-sensor fusion magnetic field measurement system for underwater vehicles in complex magnetic environments for deep-sea operations. The system underwent long-term testing and verification at a depth of 3000 meters. It includes a multi-modal magnetic field sensing module, an attitude correlation correction module, a hierarchical fusion processing module, and an environmental adaptation and magnetic anomaly detection module. The overall system operating frequency is 100Hz, and the data acquisition accuracy reaches the 0.1nT level. Compared to traditional single-sensor measurement systems, the measurement accuracy of this example is improved by 65%, and the anti-interference capability is improved by 80%. It can still maintain stable target detection capability in strong magnetic interference environments. The system adopts a distributed processing architecture, with each module communicating via a high-speed CAN bus. The data transmission rate reaches 1Mbps, and the latency is less than 5ms, ensuring real-time requirements. The total power consumption of the system is controlled within 45W, which is 30% lower than similar systems, and the endurance is improved by 40%, making it suitable for long-term underwater operations.

[0100] The multimodal magnetic field sensing module adopts a three-level sensor network architecture, including a vector-scalar collaborative measurement unit, a magnetic field spatial gradient tensor construction unit, and a magnetic interference spectrum real-time monitoring unit. The vector-scalar collaborative measurement unit is equipped with eight high-precision triaxial fluxgate magnetometers with a sampling frequency of 200Hz, a measurement range of ±100μT, a resolution of 0.05nT, and a nonlinearity of less than 0.1%. These magnetometers are distributed on the surface of the underwater vehicle's hull according to a regular octahedral geometry, with an adjacent sensor spacing of 1.2 meters, forming 360-degree omnidirectional coverage. Simultaneously, it is equipped with four optically pumped magnetometers as scalar references, achieving a measurement accuracy of 0.01nT. The measurement accuracy of both types of sensors... The measurement data is cross-validated in real time using a weighted Kalman filter algorithm. When the deviation between the vector measurement value and the scalar measurement value exceeds 0.5 nT, an anomaly alarm is automatically triggered and redundant sensors are activated. During the three-month sea trial, the system successfully identified and eliminated 273 abnormal sensor data, achieving a data reliability of 99.8%. Compared to a single-type sensor system, the collaborative measurement method reduced the measurement error by 58%, especially in areas with rapidly changing magnetic fields, where the measurement accuracy was improved by 72%. The unit also integrates a temperature compensation module, with an operating temperature range of -5℃ to 45℃ and a temperature drift coefficient of less than 0.02 nT / ℃, ensuring measurement stability under different water temperature environments.

[0101] The magnetic field spatial gradient tensor construction unit is based on a three-dimensional measurement network consisting of 16 magnetometer nodes with a node spacing of 0.8 meters. A GPS-synchronized clock ensures time alignment accuracy at the 1 microsecond level. The multi-point sampling coordinator adopts a master-slave control architecture. The master controller is an ARM Cortex-A53 quad-core processor with a clock frequency of 1.5 GHz, and the slave controllers are 16 STM32F407 microcontrollers, each managing data acquisition and preliminary processing at each measurement point. The gradient calculation processor uses an FPGA parallel processing architecture and is a Xilinx processor. The Zynq-7000 series features a built-in programmable logic unit for high-speed gradient calculation with a processing latency of less than 10ms. The system can calculate nine complete gradient tensor components in real time, including three diagonal components and six off-diagonal components. Gradient measurement sensitivity reaches 5nT / m, a 45% improvement over traditional finite difference methods. In actual marine environment testing, this unit successfully constructed a three-dimensional magnetic field distribution map of seabed magnetic anomalies with a spatial resolution of 0.5 meters, three times higher than traditional methods. The tensor reconstruction processor employs a dynamic correction algorithm that considers the impact of underwater vehicle speed and attitude changes on gradient measurement. When the underwater vehicle speed varies within the range of 0-5 knots, the gradient measurement error is controlled within 8%. This algorithm can also adaptively adjust the calculation window size: 2 seconds in a stationary state to improve accuracy, and shortened to 0.5 seconds during rapid movement to improve response speed. As an alternative, tensor construction can also use a method based on spherical harmonic function expansion, which reduces computational complexity by approximately 35%, but has slightly lower accuracy (approximately 12%) in boundary condition handling, making it suitable for applications with higher real-time requirements.

[0102] The real-time magnetic interference spectrum monitoring unit employs a wideband induction coil sensor with a frequency response range of 0.01Hz-10kHz. Sensitivity in key frequency bands (around 50Hz power frequency, thruster harmonic frequencies of 200-800Hz, and electronic device switching frequencies of 2-5kHz) is improved by 60%. The spectrum analysis processor uses a dual-channel parallel FFT algorithm, implemented on TI's TMS320C6678 multi-core DSP processor, containing eight C66x cores and a processing capacity of 160GFLOPS. It can simultaneously process 8192-point FFT transforms with a time resolution of 0.1 seconds and a frequency resolution of 0.01Hz. The system has established a spectrum fingerprint database containing 127 typical interference sources, covering propulsion motor interference (main frequency 200Hz and its harmonics) and sonar system interference (pulse characteristics, duration 20-50ms). Interference from various sources, including communication systems (carrier frequencies 1-100kHz), is identified. Each interference source's fingerprint includes 12 characteristic parameters such as main frequency, harmonic structure, power spectral density distribution, and time-varying characteristics. The interference identification accuracy reaches 94.5%, with a false identification rate of less than 3%. In practical applications, this unit successfully identified and suppressed magnetic interference signals with an amplitude of 150nT from the underwater vehicle's own propulsion system. After suppression, the residual interference was less than 8nT, achieving an interference suppression ratio of 25:1. For external electromagnetic interference, the system can initiate a suppression strategy within 100ms after detection, a response speed 5 times faster than traditional methods. This unit can also use wavelet transform as an alternative analysis method. Wavelet transform has better time-frequency localization characteristics when processing non-stationary signals, but the computational load increases by about 50%, and memory requirements increase by about 80%, making it more suitable for offline analysis scenarios.

[0103] The attitude correlation correction module's magnetic-inertial fusion attitude calculation unit integrates a MEMS inertial measurement unit, specifically an InvenSense ICM-20948 nine-axis sensor, including a three-axis gyroscope (range ±2000 dps, zero-bias stability 10° / h), a three-axis accelerometer (range ±16g, noise density 100 μg / √Hz), and a three-axis magnetometer. The attitude fusion processor employs an improved extended Kalman filter algorithm, with a 15-dimensional state vector including attitude quaternions, velocity, position, gyroscope bias, and accelerometer bias. The measurement update frequency is 100Hz, and the attitude calculation accuracy is 0.3° for roll and pitch angles and 0.8° for yaw angle. Compared to using an inertial measurement unit alone, the fusion algorithm reduces attitude drift speed by 85%, and the cumulative attitude error is less than 2° within 10 minutes, while pure inertial calculation... The error can reach over 15°. The system also implements an attitude anomaly detection function. When the fusion residual between the magnetometer measurement value and the inertial measurement value exceeds the threshold (set to 3σ), the weight of the magnetometer in the fusion is automatically reduced to avoid magnetic anomaly interference affecting attitude estimation. In the magnetic anomaly area test, this protection mechanism ensures that the attitude estimation error increases by no more than 15%, while the error of the unprotected system can increase by up to 200%. The attitude data output interface adopts the standard ROS message format and supports three representation methods: Quaternion, Euler angle and rotation matrix. The data latency is less than 5ms, which meets the requirements of real-time control. As an alternative, the complementary filtering algorithm can also be used. This algorithm reduces the computational complexity by about 40% and has lower requirements for processor performance, but its dynamic performance is slightly inferior to the extended Kalman filter by about 15%-20%, making it more suitable for applications with limited computing resources.

[0104] The attitude-related magnetic deviation compensation unit established an attitude-magnetic deviation mapping database containing 1296 sampling points through calibration experiments. The attitude angle sampling interval was 30°, covering the full attitude space of roll ±180°, pitch ±90°, and yaw 360°. Each attitude point underwent a 30-second static measurement and the average value was taken. The calibration accuracy reached 0.15nT. The deviation model was fitted with a third-order spherical harmonic function, and the fitting residual RMS value was 0.08nT. The lookup table used a trilinear interpolation algorithm to achieve fast lookup of any attitude angle, with a lookup time of less than 0. The 5ms compensation effect test shows that within a ±45° range of attitude angle change, the compensated measurement error decreased from an average of 12.5nT to 1.8nT, achieving an error reduction rate of 85.6%. In extreme attitude conditions (pitch angle exceeding 60°), the compensation effect slightly decreased but still kept the error within 4nT. The dynamic compensation processor employs a predictive compensation strategy, predicting the attitude at the next moment based on the attitude change rate and pre-loading compensation parameters, reducing the compensation delay from an average of 15ms to 3ms, significantly improving dynamic performance. The adaptive filtering unit implements motion pattern recognition based on a fuzzy logic system. Input variables include velocity (0-5 knots), acceleration (0-1 m / s²), attitude change rate (0-30° / s), and depth change rate (0-0.5 m / s). The output is the membership degree of five motion modes: stationary hovering, low-speed cruise, medium-speed cruise, high-speed maneuvering, and attitude adjustment. The recognition accuracy reaches 92%, and the mode switching delay is less than 200 ms. The filter parameter library is configured with optimized filtering parameters for each mode. For example, the stationary hovering mode uses a 20th-order low-pass filter with a cutoff frequency of 0.5 Hz, which can effectively suppress high-frequency noise and reduce measurement noise by 75%. The high-speed maneuvering mode uses a 3rd-order filter with a cutoff frequency of 5 Hz to ensure dynamic response and phase delay of less than 50 ms. Adaptive filtering improves the signal-to-noise ratio by an average of 28 dB under different motion states and improves measurement stability by 65%. This unit can also be replaced by a motion pattern recognition method based on neural networks. The neural network method can improve the recognition accuracy of complex motion modes by 5-8%, but it requires a large amount of training data and increases the computational complexity by about 2 times.

[0105] The sensor-level data pre-fusion unit of the hierarchical fusion processing module adopts a timestamp synchronization mechanism. All sensor data uniformly uses GPS-synchronized UTC time tags, achieving a time synchronization accuracy of 1 microsecond. The data synchronization processor adopts a circular buffer structure with a buffer depth of 1 second (capable of storing 100 frames of data), supporting time alignment for sensors with varying sampling rates. Fusion of similar sensors employs a joint Kalman filter algorithm based on covariance weighting. The fusion weights of the eight vector magnetometers are dynamically calculated based on the real-time noise covariance matrix. The noise covariance is estimated through a sliding window (window length 5 seconds), enabling adaptive tracking of sensor performance changes. Four scalar magnetometers participate in the fusion as constraints, with a constraint strength coefficient set at 0.3, which can be dynamically adjusted within the range of 0.1-0.5 based on the vector measurement reliability. The fused magnetic field measurement accuracy reaches 0.08 nT, a 68.8% improvement compared to the 0.25 nT of a single sensor, and the measurement stability (Allan variance) is improved by 5 times. The reliability assessment processor employs a multi-index comprehensive evaluation method, including six indicators such as measurement consistency (standard deviation), sensor health (based on historical performance statistics), and physical constraint compliance (divergence and curl constraints). A weighted scoring mechanism is used, with a threshold set at 0.7 (out of 1.0). When the reliability falls below the threshold, a secondary fusion optimization procedure is triggered. In 3000 real-world measurements, the reliability assessment accuracy reached 96.5%, effectively identifying various anomalies. The fusion result output includes three parts: measurement values, covariance matrix, and reliability index. The data format uses a custom binary protocol, with a single frame data size of 128 bytes, resulting in a transmission efficiency four times higher than text format. This fusion algorithm can also be replaced by a robust fusion method based on M-estimation. The M-estimation method has stronger outlier suppression capabilities, improving outlier suppression by approximately 30%, but increasing computational complexity by approximately 50% and extending the iteration convergence time to 15-20ms, making it more suitable for scenarios with extremely high robustness requirements.

[0106] The feature-level multi-domain fusion unit achieves collaborative feature extraction in the time, frequency, and spatial domains. The time-domain feature extraction processor uses a sliding window mechanism (window length 10 seconds, sliding step size 1 second) to extract 21 statistical feature parameters, including mean, standard deviation, skewness, kurtosis, maximum value, minimum value, energy, and zero-crossing rate. These features are normalized to form a time-domain feature vector. Frequency-domain feature extraction is based on a 256-point short-time Fourier transform (with a Hamming window function), extracting 15 feature parameters such as the dominant frequency component (frequency components with an energy share exceeding 20%), frequency band energy ratio (energy distribution in the 0-1Hz, 1-10Hz, and 10-100Hz bands), and spectral entropy. Frequency domain analysis reveals typical spectral characteristics of magnetic anomaly signals. For example, magnetic anomaly signals generated by ferromagnetic targets are mainly concentrated in the 0-0.5Hz low-frequency band, with an energy share exceeding 85%, while interference signals are more widely distributed. The spatial domain feature extraction processor calculates spatial gradient and Laplace's algorithm based on magnetic field distribution data from 16 measurement points. Eighteen spatial feature parameters, including magnetic field curvature and field distribution symmetry, were extracted and reduced to 8-dimensional principal components using Principal Component Analysis (PCA), retaining 95% of the original information. Three types of feature vectors were concatenated to form a 44-dimensional comprehensive feature vector. The total feature extraction time was less than 50ms, meeting real-time requirements. Feature fusion employed a weighted summation method, with weights optimized using an offline-trained Support Vector Machine (SVM) classifier: time domain weight 0.35, frequency domain weight 0.40, and spatial domain weight 0.25. This weight configuration achieved a detection rate of 91.2% and a false alarm rate of 5.8% in the magnetic anomaly detection task. Compared to single-domain feature analysis, multi-domain fusion improved the detection rate by 23% and reduced the false alarm rate by 68%. Wavelet packet decomposition can also be used as an alternative method for feature extraction. Wavelet packets can simultaneously provide time-frequency localized features, offering better feature extraction results for non-stationary signals and further improving the detection rate by 3-5%. However, this increases computational complexity by approximately 1.5 times, requiring higher-performance processors for real-time processing.

[0107] The decision-level intelligent fusion unit employs a multi-objective optimization strategy. The fusion result evaluation processor calculates three types of indicators: data fidelity, calculated based on the residual between the fusion result and the original measurement (with a residual threshold of 2.5 nT); physical consistency, assessing whether the fusion result satisfies the divergence and curl constraints of Maxwell's equations (with divergence thresholds of 5 nT / m and curl thresholds of 3 nT / m); and information entropy, evaluating the information richness of the fusion result (higher entropy values ​​indicate greater information content). These three indicators are weighted and summed to obtain a comprehensive score with weights of 0.4, 0.4, and 0.2, respectively. The scoring range is... The decision rule base, ranging from 0 to 100, contains 23 rules covering typical scenarios such as normal measurement, partial sensor failure, strong interference, and approaching magnetically anomalous targets. Each rule defines the triggering condition, priority, and corresponding processing strategy. For example, when strong interference (interference intensity exceeding 50 nT) is detected, the frequency domain filter order is automatically increased and the weight of time domain features is reduced. The intelligent selection processor uses an iterative algorithm based on particle swarm optimization (PSO), with 50 particles, a limit of 20 iterations, and each iteration taking approximately 8 ms. The learning rate parameter is set to 0.01, and regularization is... With parameters λ1=0.1 and λ2=0.05, iterative optimization improved the overall score of the fusion results by an average of 15.3%. In 1000 field tests, convergence occurred within 10 iterations in 94.7% of cases, and the degree of physical consistency violation decreased by 82%. The optimized fusion result is used as the final output. The output data package includes the fused magnetic field value (3D vector), uncertainty estimate (covariance matrix), confidence score (0-1), physical consistency index, and decision information (applied rule number), etc. The data update frequency is 20Hz, and the data latency is less than 100ms, even in complex magnetic environments. In the test, this unit improved the overall measurement accuracy of the system by 42% and robustness by 55%. As an alternative, decision fusion can also adopt an end-to-end optimization method based on deep learning, using a convolutional neural network (CNN) to directly learn the optimal fusion strategy from multi-sensor data. This method can automatically learn feature representations and can improve performance by 5-10% when there is sufficient training data. However, it requires a large amount of labeled data (at least 10,000 sets of samples) for training, and the model training time can be as long as several days. The inference time after deployment is about 50-80ms, which places higher demands on hardware performance.

[0108] The background magnetic field online modeling and separation unit of the environmental adaptation and magnetic anomaly detection module adopts a space-time joint modeling method. The background sampling processor collects background magnetic field samples at a frequency of 1Hz during the underwater vehicle's navigation. The sample data includes location (latitude and longitude coordinates, accuracy 0.1 meters), depth (accuracy 0.01 meters), three magnetic field components, and a timestamp. A distributed background magnetic field database is established, using a quadtree spatial index structure to support fast spatial queries with a query time complexity of O(log n). The model update processor uses the recursive least squares (RLS) algorithm to update the background field model parameters in real time. With a forgetting factor set to 0.98, it can track the slow time-varying nature of the background field (typically on a timescale of hours to days). The model uses a spherical harmonic function expansion to order 15, containing 240 coefficients, which can accurately describe the spatial distribution characteristics of the local background magnetic field. The model prediction accuracy reaches 1.5 nT, a four-fold improvement compared to the local prediction accuracy of global geomagnetic field models (such as IGRF). The background separation processor subtracts the model prediction value from the current measurement value to obtain the residual signal. The residual signal is then filtered by a third-order Butterworth high-pass filter (cutoff frequency 0.02 Hz) to remove low-frequency trends, yielding the target magnetic field signal. In actual sea trials, this unit successfully separated a weak magnetic anomaly signal with an amplitude of only 3 nT, even when the background field intensity reached [missing value]. At a resolution of 50,000 nT, the separation accuracy reached 0.01%, providing high-quality input data for subsequent target detection. Compared with a fixed background field model, the online modeling method improved the target detection distance by 35% and reduced the false alarm rate by 52%. The model parameters are stored in non-volatile memory and can be recovered after power failure, supporting fast initialization (startup time less than 2 seconds). Background field modeling can also be replaced by a non-parametric method based on Gaussian process regression (GPR). The GPR method does not require pre-assumptions about the model form and has stronger adaptability to complex background field distributions, improving prediction accuracy by 10-15%. However, the computational complexity increases by about 5 times. For a dataset with 1,000 sampling points, the model training time increases from 0.5 seconds for RLS to 2.5 seconds for GPR, resulting in a slight decrease in real-time performance.

[0109] The multi-source interference identification and suppression unit achieves automatic identification and adaptive suppression of three types of interference sources. The interference source classification processor adopts a decision tree classification algorithm. The feature input includes the time-domain statistical features (12-dimensional), frequency-domain features (8-dimensional), and spatial correlation features (6-dimensional) of the interference signal. The classifier was trained with 5000 sets of labeled data, and the cross-validation accuracy reached 96.8%. For interference from the underwater vehicle's own equipment, the system analyzed the magnetic interference characteristics of 15 types of equipment, including propulsion motors, servo motors, sonar, lighting, and water pumps. The generated magnetic interference is the strongest, with an amplitude reaching 200 nT. The dominant frequency is the motor speed frequency (200-600 Hz) and its harmonics. Servo interference is intermittent pulses, lasting 50-100 ms with an amplitude of 30-80 nT. Sonar operation generates broadband interference (1-100 kHz) with an amplitude of 10-50 nT. The system has established a database of correspondence between equipment operating conditions and interference characteristics, containing 127 operating condition modes. It can predict the interference level based on equipment status, with a prediction error of less than 15%. For geomagnetic disturbance interference, the system... The system monitors the solar activity index (Kp index) and local geomagnetic diurnal variations. When the Kp index is greater than 4, the filter order is automatically increased, and the weight of low-frequency measurements is reduced. Typical characteristics of geomagnetic disturbances are slow-changing trends (timescales from minutes to hours) and sudden disturbances (amplitudes can reach 100 nT during geomagnetic storms). The system uses a multi-scale time series analysis method to separate geomagnetic disturbances from target signals, achieving a separation accuracy of 92%. For external electromagnetic interference, the interference feature matcher retrieves matching interference templates from the interference fingerprint database. The matching algorithm is based on dynamic time warping (DTW) and can identify interference signals with slight time scale changes, achieving a matching accuracy of 89% and an identification delay of less than 300 ms. The adaptive suppression processor selects the appropriate suppression strategy based on the identification results. An adaptive notch filter is used for periodic interference, with a notch depth of up to 40 dB. Median filtering is used for impulse interference, with adaptive window length adjustment (3-15 sampling points). Wavelet threshold denoising is used for broadband interference, with the threshold adaptively set according to the noise power. In typical marine environment tests, the total power of multi-source interference decreased from 85 nT before optimization. The RMS was reduced to an optimized 6.5 nT RMS, the interference suppression ratio reached 13:1, and the signal-to-noise ratio of magnetic anomaly target detection was improved by 22 dB. As an alternative technique, interference suppression can also be achieved using blind source separation (BSS) methods, such as independent component analysis (ICA). The BSS method does not require prior knowledge of the characteristics of the interference sources and can automatically separate multiple mixed signal sources. It performs better in scenarios with a large number of interference sources and unknown characteristics, and the interference suppression capability can be improved by 10-20%. However, it requires multiple observation channels (at least equal to the number of sources), which places higher demands on sensor configuration. In addition, the algorithm convergence speed is slower, and the processing latency increases to 500-800 ms.

[0110] The intelligent magnetic anomaly target recognition unit realizes a complete process from target detection to classification and recognition. The self-magnetic feature extraction processor measured the distribution of the underwater vehicle's own magnetic field under different attitudes and operating conditions through magnetic compensation experiments, establishing a 12-dimensional feature vector self-magnetic feature template, including magnetic dipole model parameters (3 components of magnetic moment, 3 coordinates of magnetic center position) and higher-order multipole correction terms (6 parameters). The static magnetic field strength of the underwater vehicle is approximately 15 nT at a distance of 1 meter, increasing to 45 nT at a distance of 1 meter under dynamic operating conditions (thruster operation). The influence of the self-magnetic field is subtracted from the measured values ​​through template matching, achieving a subtraction accuracy of 90%. The anomaly detection processor employs a constant false alarm rate (CFAR) based on statistical hypothesis testing. The CFAR (Confirmation of False Alarm Rate) detection algorithm adaptively adjusts the detection threshold based on local background statistical characteristics, setting the false alarm rate to 10^-4. The detector includes three operating modes: cell-averaged CFAR, ordered statistical CFAR, and modified CFAR, automatically switching based on background uniformity. Cell-averaged CFAR provides the highest detection sensitivity against uniform backgrounds, modified CFAR offers strong resistance to edge effects at clutter edges, and ordered statistical CFAR offers the best robustness against complex backgrounds. The mode switching delay is less than 100ms. With a signal-to-noise ratio of 6dB, the detection probability reaches 95%, and the false alarm rate is 2×10^-4, meeting design requirements. The target classification and recognition system employs a two-level classification architecture, with the first level being a coarse classification. Targets were categorized into three main types: ferromagnetic targets, conductive non-magnetic targets, and complex targets. A Support Vector Machine (SVM)-based classifier was used, achieving a classification accuracy of 92.5%. A second-level fine-tuning classifier further identified target types for ferromagnetic targets (e.g., steel pipelines, shipwrecks, metal deposits), employing a Random Forest-based classifier with 100 decision trees, achieving a classification accuracy of 87.3%. The classifier's input features included magnetic anomaly amplitude (typically ranging from 5 to 500 nT), spatial decay rate (target distance estimation, accuracy ±15%), magnetic dipole orientation (target attitude estimation, accuracy ±10°), spectral features (5 parameters), and time-varying features (4 parameters), totaling an 18-dimensional feature vector. Feature extraction and classification... The total processing time is less than 200ms, supporting real-time situational awareness. During sea trials, the system successfully detected and identified 32 magnetic anomaly targets, including 17 pipelines, 8 shipwrecks, 5 metal components, and 2 rock deposits. The detection range was 15-120 meters, with an average distance estimation error of 8.5 meters and a target type recognition accuracy of 90.6%. Compared with traditional threshold-based simple detection methods, the intelligent recognition unit improved the detection distance by 40%, reduced the false alarm rate by 75%, and improved the classification accuracy by 55%. The system also supports multi-target tracking, using the Joint Probabilistic Data Association (JPDA) algorithm for data association, supporting simultaneous tracking of up to 8 targets with a tracking accuracy of 5 meters for position error and 0 for velocity error.At a speed of 2 m / s, as an alternative method, object recognition can also employ end-to-end recognition networks based on deep learning, such as Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs). Deep learning methods can automatically learn discriminative feature representations, achieving recognition accuracy of 94-96% on large-scale datasets, a 5-8% improvement over traditional methods. However, this requires at least 5000 labeled samples for training, taking several days, and the inference time after deployment is 100-150 ms. It also places higher demands on hardware, requiring GPU accelerators or dedicated AI chips.

[0111] The content disclosed above is only a preferred and feasible embodiment of the present invention, and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent technical changes made based on the content of the present invention specification and drawings are included within the scope of protection of the present invention. Furthermore, the elements therein can be updated as technology develops.

Claims

1. A multi-sensor fusion magnetic field measurement system for underwater targets in complex magnetic environments, characterized in that, It includes a multimodal magnetic field sensing module, an attitude correlation correction module, a hierarchical fusion processing module, and an environment adaptation and magnetic anomaly detection module; The multimodal magnetic field sensing module is responsible for collecting multi-dimensional and multi-type raw magnetic field data around the underwater vehicle to build a complete magnetic field observation foundation. The attitude correlation correction module is used to eliminate the influence of changes in the underwater vehicle's motion attitude on magnetic field measurement. The hierarchical fusion processing module performs hierarchical fusion of multi-sensor data. The environmental adaptation and magnetic anomaly detection module realizes background magnetic field separation, interference suppression, and magnetic anomaly target identification. The multimodal magnetic field sensing module includes a vector-scalar collaborative measurement unit, a magnetic field spatial gradient tensor construction unit, and a magnetic interference spectrum real-time monitoring unit. The vector-scalar collaborative measurement unit is used to realize the synchronous measurement of magnetic field direction and intensity. The magnetic field spatial gradient tensor construction unit is used to construct a complete magnetic field gradient tensor. The magnetic interference spectrum real-time monitoring unit monitors magnetic interference signals across the entire frequency band in real time. The attitude correlation correction module includes a magnetic-inertial fusion attitude calculation unit, an attitude-related magnetic deviation compensation unit, and a motion state adaptive filtering unit. The magnetic-inertial fusion attitude calculation unit is used to achieve high-precision attitude estimation of the underwater vehicle. The attitude-related magnetic deviation compensation unit dynamically compensates for measurement deviations caused by attitude changes based on real-time attitude information. The motion state adaptive filtering unit is used to identify the current motion mode of the underwater vehicle and adaptively switch the corresponding filtering strategy according to the noise characteristics of different motion states. The hierarchical fusion processing module includes a sensor-level data pre-fusion unit, a feature-level multi-domain fusion unit, and a decision-level intelligent fusion unit. The sensor-level data pre-fusion unit performs primary fusion of measurement data from sensors of the same type. The feature-level multi-domain fusion unit is used to fuse extracted multi-domain features to obtain a more comprehensive description of magnetic field features. The decision-level intelligent fusion unit selects the optimal fusion result as the system output based on the credibility evaluation of the fusion result. The environmental adaptation and magnetic anomaly detection module includes an online background magnetic field modeling and separation unit, a multi-source interference identification and suppression unit, and a magnetic anomaly target intelligent identification unit. The online background magnetic field modeling and separation unit constructs and updates the local background magnetic field model of the underwater vehicle's operating area in real time, effectively separating the background magnetic field from the target magnetic field signal. The multi-source interference identification and suppression unit is used to identify and distinguish multi-source interference signals and adopt targeted suppression strategies to reduce the impact of interference. The magnetic anomaly target intelligent identification unit detects and identifies magnetic anomaly targets in the environment based on extracting the magnetic characteristics of the underwater vehicle itself.

2. The underwater target multi-sensor fusion magnetic field measurement system under complex magnetic environments as described in claim 1, characterized in that, The magnetic field spatial gradient tensor construction unit includes a multi-point sampling coordinator, a gradient calculation processor, and a tensor reconstruction processor. The multi-point sampling coordinator controls the distributed magnetometer array to perform time-synchronized sampling to ensure the time consistency of data from each measuring point. The gradient calculation processor calculates the first-order gradient components of the magnetic field in each spatial direction based on the magnetic field difference between adjacent measuring points. The tensor reconstruction processor combines the calculated gradient components to construct a complete magnetic field gradient tensor matrix, forming a mathematical description of the spatial distribution of the magnetic field. The tensor reconstruction processor constructs a dynamic gradient tensor according to the following formula: ； in, This represents the dynamic gradient tensor of the nth measurement point at time t. Let represent the static gradient tensor of the nth measurement point at time t. This is the motion-induced correction term for the nth measuring point at time t; The static gradient tensor is directly composed of gradient information calculated by the gradient calculation processor, and the motion-induced correction term is calculated according to the following formula: ； Where v(t) represents the velocity vector of the underwater vehicle at time t, and c m B is the characteristic velocity of the magnetic field response. n Let w(t) be the magnetic field at the nth measuring point, and w(t) be the angular velocity vector of the underwater vehicle. It is an antisymmetric matrix composed of angular velocities.

3. The underwater target multi-sensor fusion magnetic field measurement system under complex magnetic environments as described in claim 2, characterized in that, The intelligent magnetic anomaly target identification unit includes a self-magnetic feature extraction processor, an anomaly detection processor, and a target classification and recognition device. The self-magnetic feature extraction processor analyzes the magnetic field characteristics generated by the underwater vehicle itself and establishes a self-magnetic feature template for distinguishing it from the environmental magnetic field. The anomaly detection processor detects abnormal magnetic signals that deviate from the background magnetic field and the self-magnetic features through statistical tests, threshold decisions, or machine learning methods. The target classification and recognition device performs feature analysis and pattern recognition on the detected magnetic anomalies to determine the type, location, and motion state of the abnormal target, thereby achieving intelligent identification of potential magnetic targets. The anomaly monitoring processor calculates the overall anomaly degree of the sensor according to the following formula: ； ； ； in, This represents the overall anomaly degree of the nth sensor at time t. Let n be the amplitude anomaly of the nth sensor at time t. Let be the topological anomaly degree of the nth sensor at time t. w is the physical consistency index for the nth measurement point. amp w topo w phys These are the weighting coefficients for the three types of anomalies. Let be the curvature of the magnetic field lines at the nth measuring point at time t. Let the topological feature symbol of the nth measurement point at time t be denoted as . , Here, tr() is the sensitivity parameter, and tr() is the trace function. For divergence-constrained sensitivity parameters, This is the curl constraint sensitivity parameter.

4. The underwater target multi-sensor fusion magnetic field measurement system under complex magnetic environments as described in claim 3, characterized in that, The sensor-level data pre-fusion unit includes a data synchronization processor, a similar sensor fusion processor, and a reliability evaluation processor. The data synchronization processor performs time alignment and coordinate system unification on the measurement data of similar sensors to eliminate sampling time differences. The similar sensor fusion processor is used to fuse data from similar sensors and improve measurement accuracy by utilizing redundant information. The reliability evaluation processor evaluates the reliability of the fusion result based on the noise level, consistency, and historical performance of each sensor. The aforementioned sensor fusion processor fuses sensor data according to the following formula: ； in, Let N be the fused magnetic field vector at time t. v N represents the number of vector magnetometers. s For the number of scalar magnetometers, Let be the measurement value of the nth vector magnetometer at time t. Let be the measurement value of the m-th scalar magnetometer at time t. (t) represents the fusion weight of the nth vector magnetometer. (t) represents the fusion weight of the m-th scalar magnetometer. This is the scalar constraint strength coefficient.

5. The underwater target multi-sensor fusion magnetic field measurement system under complex magnetic environments as described in claim 4, characterized in that, The decision-level intelligent fusion unit includes a fusion result evaluation processor, a decision rule base, and an intelligent selection processor. The fusion result evaluation processor receives the credibility index of the sensor-level fusion result and calculates the physical consistency and information richness of the optimization result after iterative optimization, evaluating the quality improvement brought about by the optimization. The decision rule base stores a set of decision rules based on expert knowledge and historical experience, defining the optimal fusion result selection strategy under different scenarios. The intelligent selection processor performs iterative optimization of the sensor-level fusion result guided by physical constraints according to the evaluation result and decision rules, and selects the optimal fusion output. It can dynamically adjust the optimization strategy and decision parameters according to environmental changes. The intelligent selection processor iteratively processes the fusion result according to the following formula: ； in, This represents the fusion result of the k-th iteration. For learning rate, To ensure data fidelity and accuracy, H[B] represents the physical constraint error, and H[B] represents the information entropy of the fused magnetic field data. , For two regularization parameters; The result of each iteration is used as a candidate fusion result.

Images