A marine magnetic map matching navigation method and system based on self-supervised deep learning and multi-modal fusion

Abstract

The application provides a marine magnetic map matching navigation method and system based on self-supervised deep learning and multi-modal fusion. The method uses a deep learning network to extract geomagnetic sequence features and inertial motion features, introduces magnetic field gradient information to dynamically modulate cross-modal attention weights, and combines the IMU cumulative drift state to construct an adaptive gating fusion mechanism, so that the fusion process can dynamically adjust according to environmental changes and sensor state. Further, combined with the pre-acquired marine magnetic map, the navigation position estimation is realized through deep feature matching. Compared with the existing methods based on artificial features and single-modal matching, the application can fully utilize the geomagnetic spatial distribution information, carrier motion constraint information and sensor state information, and reduce the dependence of the model on large-scale real navigation trajectories and high-precision position labels through the self-supervised sample generation based on the magnetic map and the cross-modal contrast learning mechanism, thereby improving the robustness and positioning accuracy of the magnetic map matching in complex marine environments.

Description

Technical Field

[0001] This application belongs to the fields of geomagnetic navigation, inertial navigation, deep learning, and magnetic field measurement, and specifically relates to a marine magnetic map matching navigation method and system based on self-supervised deep learning and multimodal fusion. Background Technology

[0002] With the rapid development of marine unmanned platforms, autonomous ship navigation, and underwater robots, high-precision autonomous navigation technology has become a crucial foundation for ensuring the long-term autonomous operation of marine equipment. Currently, the Global Navigation Satellite System (GNSS) can provide high-precision positioning services, but in the open ocean, complex sea conditions, and under conditions of obstruction and interference, satellite navigation signals are easily limited, making it difficult to meet the all-weather, highly reliable autonomous navigation requirements of marine equipment.

[0003] Therefore, using geophysical field information for assisted navigation has gradually become an important research direction in the field of autonomous navigation. Among them, the geomagnetic field has the characteristics of global coverage, no need for external signal sources, and susceptibility to human interference. Navigation methods based on geomagnetic information have become a promising autonomous navigation technology.

[0004] Traditional geomagnetic matching methods typically utilize artificially designed physical features such as magnetic field strength and gradient, employing correlation matching, probability estimation, and filtering optimization to achieve location matching between real-time measurement data and pre-established magnetic maps. While these methods offer advantages such as clear physical meaning and simple implementation, they still have limitations: traditional methods often rely on artificially designed magnetic field features, which have limited expressive power. When affected by factors such as changes in the marine environment's magnetic field, sensor measurement noise, carrier attitude changes, and local magnetic anomalies, artificial features struggle to accurately describe the complex spatial distribution of the magnetic field, easily leading to a decrease in matching accuracy.

[0005] In recent years, with the development of deep learning technology, some studies have begun to utilize deep models such as Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Transformers to learn features from geomagnetic time series data. Compared with traditional manual feature extraction methods, deep learning methods can automatically extract high-dimensional features from data through multi-layer nonlinear mapping, thereby improving the adaptability of geomagnetic matching in complex environments to a certain extent.

[0006] However, existing deep learning-based geomagnetic navigation methods still have the following problems: (1) Most methods mainly model single geomagnetic data, without fully considering the correlation between geomagnetic observation information and carrier motion state, making it difficult to simultaneously utilize magnetic field spatial constraints and motion continuity constraints.

[0007] (2) Existing deep fusion methods usually calculate the correlation weights between different modes based on data-driven methods, lacking constraints from the physical characteristics of the geomagnetic field, and cannot dynamically adjust the importance of information in different regions according to the degree of spatial change of the magnetic field.

[0008] Inertial Measurement Units (IMUs) can continuously provide information on a vehicle's acceleration, angular velocity, and attitude changes, independent of the external environment, and are therefore widely used in autonomous navigation systems. However, IMU navigation requires obtaining velocity and position through integration, and its measurement errors accumulate over time, resulting in a significant cumulative drift problem.

[0009] Therefore, fusing geomagnetic information with IMU information can constrain inertial navigation errors by utilizing the spatial distribution characteristics of the geomagnetic field, thereby improving the long-term operational stability of the system. Currently, some studies have employed geomagnetic and IMU fusion methods for integrated navigation, but most methods are still based on traditional filtering frameworks such as the Extended Kalman Filter (EKF). Their fusion process typically relies on a preset motion model and manually set noise parameters. Since sensor states change over time during actual navigation—for example, IMU cumulative drift gradually increases and the magnetic field environment changes—fixed fusion parameters cannot reflect the current reliability of different sensors in real time, thus limiting the fusion effect.

[0010] In summary, the existing technology has the following main shortcomings: (1) Artificially designed magnetic field characteristics are difficult to fully express the spatial variation law of magnetic field in complex marine environments; (2) Existing deep fusion methods lack geomagnetic physical information constraints and cannot dynamically adjust cross-modal correlations using magnetic field gradients; (3) Traditional geomagnetic-IMU fusion methods are difficult to detect the cumulative drift state of IMUs and cannot dynamically adjust the fusion ratio according to the reliability of the sensors; (4) Existing magnetic map matching methods are not adaptable to noise, magnetic anomalies and error accumulation during long-term navigation.

[0011] (5) The spatial distribution of magnetic fields in different geographical regions may be similar, which may easily lead to mismatch.

[0012] (6) The training of existing depth magnetograph matching models depends on a certain scale of actual navigation data. The cost of acquiring large-scale navigation data and high-precision reference trajectories in the marine environment is high, which limits the scale of training samples and the engineering application of depth models. Summary of the Invention

[0013] To address the aforementioned issues, this application proposes a marine magnetomap matching navigation method and system based on self-supervised deep learning and multimodal fusion. By constructing multimodal deep feature representations of geomagnetic information and IMU information, the method achieves the synergistic fusion of geomagnetic spatial features and carrier motion information. Furthermore, it automatically generates self-supervised training samples using pre-acquired marine magnetomaps, reducing the dependence of deep matching model training on large-scale real navigation trajectories and high-precision location labels.

[0014] This application proposes a marine magnetograph matching navigation method based on self-supervised deep learning and multimodal fusion, including: Step S1: Obtain real-time navigation data, including geomagnetic measurement data and inertial information, and execute steps S2A and S2B; Step S2A: Input geomagnetic measurement data into the trained magnetic encoder to output high-dimensional geomagnetic features; input inertial information into the trained IMU encoder to output high-dimensional inertial features; adopt a magnetic field gradient modulation cross-attention mechanism, using magnetic field gradient information to dynamically modulate the cross-attention weights, establishing the correlation between geomagnetic and inertial information while introducing physical constraints of the spatial variation characteristics of the magnetic field; after completing the cross-attention fusion, adopt a drift-sensing gating fusion method, combining the accumulated drift state of the IMU to adaptively adjust the contribution ratio of geomagnetic and inertial features in the fusion process; subsequently, perform attention pooling on the fused sequence features to generate a query embedding that can comprehensively reflect geomagnetic spatial information and the motion state of the carrier; Step S2B: Based on the known state at the start of the current matching time window and the continuously acquired inertial measurement data within that time window, perform inertial navigation integration to continuously extrapolate the vehicle's attitude, velocity, and position state, obtaining inertial navigation position estimates for each sampling moment within the current matching time window, thus forming a coarse inertial trajectory for the vehicle; after obtaining the coarse inertial trajectory, adaptively determine the candidate trajectory search area by combining the IMU cumulative drift evaluation results; within the candidate trajectory search area, use the coarse inertial trajectory as the motion prior to discretize and construct multiple candidate trajectories that satisfy the known initial state constraint, inertial motion trend constraint, and trajectory continuity constraint. Select trajectories to form a candidate trajectory set; for each generated candidate trajectory, extract the local magnetic map region covering the candidate trajectory and its neighborhood from the pre-constructed ocean magnetic map based on the spatial location of the candidate trajectory in the ocean magnetic map, and establish a one-to-one correspondence between the candidate trajectory and the local magnetic map region; input the local magnetic map corresponding to each candidate trajectory into the trained ConvNeXt network for deep spatial feature extraction to obtain the depth feature map of the local magnetic map; aggregate the spatial features of the depth feature map through global average pooling, and obtain a map embedding in the same feature space as the query embedding through linear mapping and normalization; Step S3: Calculate the similarity between the query embedding corresponding to the real-time navigation data within the current matching time window and the map embedding corresponding to each candidate trajectory to obtain the depth feature similarity between the real-time geomagnetic-inertial fusion features and the magnetic field spatial distribution features corresponding to each candidate trajectory; select the candidate trajectory with the highest similarity as the optimal matching trajectory, and take the spatial position of the optimal matching trajectory at the end of the current matching time window as the current magnetic map matching navigation position estimation result of the vehicle.

[0015] As an improvement to the above method, the geomagnetic measurement data of the trained magnetic encoder input in step S2A is represented as follows: ; in, Represents geomagnetic measurement data; The position encoding matrix is ​​in sine and cosine form: ; ; in, pos= 0, 1, ..., N -1 represents the position index, indicating the position number of the current data point within the time window sequence. The length of the time window; i= 0, 1, ..., The dimension index represents the first element in the feature vector. i One dimension,d Indicates the number of dimensions in the feature vector; The geomagnetic feature embedding is represented as follows: ; in, This is the geomagnetic feature mapping matrix; The gradient feature mapping matrix; and These are bias parameters; For the first The magnetic field gradient at each sampling point; The displacement distance of the aircraft is estimated based on the IMU inertial integration results at adjacent sampling times. To prevent constants with a denominator of zero; A geomagnetic sequence constructed using a fixed time window; Sampling time Geomagnetic observations, , , These represent the magnetic field components in the three directions of the carrier coordinate system, with superscript indicating the direction of the magnetic field. T This represents the transpose of a vector.

[0016] As an improvement to the above method, the inertial information of the trained IMU encoder input in step S2A is represented as follows: ; in, Represents inertial information; Indicates time location encoding; Indicates inertial feature embedding, The inertial feature mapping matrix, For bias parameters, For length is N Inertial sequence, For the carrier's motion state information, , , This represents the acceleration in three directions within the carrier's coordinate system. , , Represents the angular velocity in three directions in the carrier coordinate system, indicated by the superscript. T This represents the transpose of a vector.

[0017] As an improvement to the above method, the accumulated IMU drift state in step S2A is represented as follows: ; in, Indicates the first IMU cumulative drift metric at any given time; Represents the trace of a matrix; Represents the error state vector The covariance matrix; the error state vector , Indicates positional error. Indicates speed error, Indicates attitude error. This indicates the zero bias error of the accelerometer. Indicates the zero bias error of the gyroscope, indicated by the superscript. T This represents the transpose of a vector.

[0018] As an improvement to the above method, step S2A employs a magnetic field gradient modulation cross-attention mechanism, which dynamically modulates the cross-attention weights using magnetic field gradient information, including: Using geomagnetic features as query information and inertial features as keys and values, geomagnetic cross-attention guided by inertial information is calculated; Softmax normalization is performed on the cross-modal correlation matrix row by row to obtain the attention weight of inertial information to each geomagnetic moment; the inertial value features are weighted and summed according to the attention weight to obtain the geomagnetic features modulated by inertial motion information. Using inertial features as queries and geomagnetic features as keys and values, reverse cross-attention is calculated; Softmax normalization is performed on the cross-modal correlation matrix row by row to obtain the attention weights of geomagnetic information for each inertial time step; the geomagnetic value features are weighted and summed according to the attention weights to obtain the inertial features enhanced by geomagnetic information. Multi-head cross-attention computation is employed, with different attention heads learning the correlation features between changes in magnetic field amplitude, magnetic field gradient, attitude change, angular velocity change, and acceleration change. The cross-attention output is residually connected to the original modality features to obtain preliminary enhanced features; the preliminary enhanced features are then input into a trained feedforward neural network to output bidirectional cross-enhanced features.

[0019] As an improvement to the above method, in step S2B, the candidate trajectory search region is adaptively determined by combining the IMU cumulative drift evaluation results, as shown below: ; in, This is the search area for candidate trajectories; Indicates position coordinates; The geographical area covered by the magnetic map; Indicating the first coarse inertial trajectory k One location; The length of the time window; Let be the 2-norm of the matrix; The maximum spatial offset range of the candidate trajectory corresponding to the current matching time window is represented as: ; in, The preset base search radius; This is a search range adjustment factor; This is the time weighting coefficient; This is the cumulative drift metric for the IMU corresponding to the current matching time window.

[0020] As an improvement to the above method, a self-supervised training process based on magnetic maps is also included, specifically: In a pre-built ocean magnetic map A virtual starting position is randomly selected within the coverage area, and a virtual trajectory that satisfies motion continuity constraints is generated according to the preset sailing speed, heading, and range of motion changes. ; in, Indicates the virtual trajectory sequence number. Indicates the first The virtual trajectory in the first The location coordinates at each sampling time. express x Axis coordinates express y Axis coordinates, superscript T This represents vector transpose; by changing the virtual starting position, initial heading, sailing speed, and motion state change parameters, multiple different virtual sailing trajectories can be generated from the same ocean magnetic map; For the Virtual trajectory Based on the spatial coordinates of each virtual trajectory point in the ocean magnetic map, the magnetic map is interpolated and sampled to obtain the corresponding synthetic geomagnetic observation sequence. ; Speed ​​is calculated based on the relationship between adjacent positions and time in the virtual flight path. : ; in, Indicates the time difference between adjacent positions; Further calculation of motion acceleration : ; Calculate the heading angle based on the direction of the trajectory. : ; Angular velocity is calculated based on the change in heading at adjacent moments. : ; Based on the virtual trajectory and the preset vehicle motion state model, corresponding attitude, velocity, and acceleration states are generated. Combined with the gravity model, IMU measurement model, and preset sensor error model, a synthetic IMU inertial measurement sequence consisting of three-axis acceleration and three-axis angular velocity is constructed. ; For virtual trajectory The magnetic map regions covering the virtual trajectory and its neighborhood are extracted from the pre-built ocean magnetic map as positive sample map tiles. ; Will and Constructing positive sample pairs ; Select from the magnetic map area corresponding to the virtual trajectory in other spatial locations. As negative sample map tiles, thus forming negative sample pairs. , The number of negative samples corresponding to each positive sample; Synthetic geomagnetic observation sequence and synthetic IMU inertial measurement sequence Following step S2A, geomagnetic and inertial features are extracted and fused to obtain a normalized query embedding. ; to the positive sample magnetic map area and negative sample magnetic map regions The corresponding normalized map embeddings are obtained by inputting them into the ConvNeXt network. ; Using contrastive learning loss function : ; ; in, The number of samples in a training batch. For temperature coefficient, For positive sample similarity, For the first Similarity of negative samples; By minimizing ,make at the same time ; During training, the parameters of each model in the method are jointly optimized through backpropagation based on the contrastive learning loss function value.

[0021] This application also provides a marine magnetograph matching navigation system based on self-supervised deep learning and multimodal fusion, for implementing the above method, including: The information acquisition module is used to acquire real-time navigation data, including geomagnetic measurement data and inertial information; The query embedding module is used to input geomagnetic measurement data into a trained magnetic encoder and output high-dimensional geomagnetic features; it also inputs inertial information into a trained IMU encoder and outputs high-dimensional inertial features. A magnetic field gradient modulation cross-attention mechanism is employed, dynamically modulating the cross-attention weights using magnetic field gradient information. This establishes the correlation between geomagnetic and inertial information while introducing physical constraints based on the spatial variation characteristics of the magnetic field. After cross-attention fusion, a drift-sensing gating fusion method is used, adaptively adjusting the contribution ratio of geomagnetic and inertial features in the fusion process based on the accumulated drift state of the IMU. Subsequently, attention pooling is applied to the fused sequence features to generate a query embedding that comprehensively reflects both geomagnetic spatial information and the carrier's motion state. The map embedding module is used to perform inertial navigation integration based on the known state at the start of the current matching time window and the continuously collected inertial measurement data within that time window. This continuously recursively extrapolates the vehicle's attitude, velocity, and position to obtain inertial navigation position estimates for each sampling moment within the current matching time window, forming a coarse inertial trajectory for the vehicle. After obtaining the coarse inertial trajectory, the candidate trajectory search area is adaptively determined by combining the IMU cumulative drift evaluation results. Within the candidate trajectory search area, multiple paths are discretized using the coarse inertial trajectory as the motion prior, constructing paths that satisfy the known initial state constraint, inertial motion trend constraint, and trajectory continuity constraint. Candidate trajectories are generated to form a candidate trajectory set. For each generated candidate trajectory, based on its spatial location in the ocean magnetic map, a local magnetic map region covering the candidate trajectory and its neighborhood is extracted from the pre-constructed ocean magnetic map, establishing a one-to-one correspondence between the candidate trajectory and the local magnetic map region. The local magnetic maps corresponding to each candidate trajectory are input into a trained ConvNeXt network for deep spatial feature extraction to obtain a depth feature map of the local magnetic map. The spatial features of the depth feature map are aggregated through global average pooling, and through linear mapping and normalization, a map embedding in the same feature space as the query embedding is obtained. The navigation position estimation module is used to calculate the similarity between the query embedding corresponding to the real-time navigation data within the current matching time window and the map embedding corresponding to each candidate trajectory, to obtain the depth feature similarity between the real-time geomagnetic-inertial fusion features and the magnetic field spatial distribution features corresponding to each candidate trajectory; the candidate trajectory with the highest similarity is selected as the optimal matching trajectory, and the spatial position corresponding to the optimal matching trajectory at the end of the current matching time window is taken as the current magneto-map matching navigation position estimation result of the vehicle.

[0022] Compared with existing technologies, the advantages of this application are: 1. This application adopts a geomagnetic time series feature encoding method based on deep learning to extract high-dimensional features from real-time geomagnetic observation data and introduces magnetic field gradient information to describe the spatial variation law of geomagnetic field. The original geomagnetic observation information and magnetic field gradient information are used together for geomagnetic feature representation. Therefore, it can simultaneously utilize geomagnetic field amplitude information and its spatial variation characteristics, reduce the problem of insufficient feature expression ability caused by the reliance on manually designed magnetic field features in traditional methods, and improve the feature representation ability of geomagnetic information in complex magnetic field environments.

[0023] 2. This application adopts a cross-attention mechanism based on dynamic modulation of magnetic field gradient. In the process of establishing the cross-modal correlation between geomagnetic information and inertial information, a modulation matrix constructed from magnetic field gradient information is introduced. This makes the cross-attention weight not only depend on the correlation between geomagnetic features and inertial features, but also constrained by the degree of spatial variation of the magnetic field. Therefore, it can enhance the information contribution of areas with obvious magnetic field changes, reduce the influence of weakly identifiable information in areas with gentle magnetic fields on the cross-modal fusion process, and improve the correlation expression ability between geomagnetic information and inertial motion information.

[0024] 3. This application introduces an IMU cumulative drift index in the gating fusion stage, incorporating the current inertial navigation error state into the fusion process of geomagnetic and inertial information. This allows the gating network to adaptively adjust the contribution ratio of the two modes of information based on the degree of IMU cumulative drift. When the IMU cumulative drift is small, the contribution of inertial motion information is increased to fully utilize its continuous and stable motion constraint capability in a short period of time. When the IMU cumulative drift increases, the contribution of geomagnetic spatial constraint information is increased, thereby reducing the impact of long-term inertial integration error accumulation on the magnetic map matching results and improving the stability during long-term navigation.

[0025] 4. This application utilizes the known state at the start of the current matching time window and continuously acquired IMU inertial measurement data within the time window. It continuously recursively extrapolates the vehicle's motion state through inertial navigation integration to obtain a coarse inertial trajectory within the current matching time window. Based on the cumulative IMU drift, it adaptively determines the search range of candidate trajectories around the coarse inertial trajectory. Within this search range, multiple candidate trajectories satisfying the known initial state constraint, inertial motion trend constraint, and trajectory continuity constraint are generated using the coarse inertial trajectory as the motion prior. Therefore, it can fully utilize the continuous motion information provided by the IMU to constrain the trajectory search space for subsequent magnetic map matching. When the cumulative IMU drift is small, a smaller candidate trajectory search range is used; when the cumulative IMU drift increases, the candidate trajectory search range is correspondingly expanded. This adapts to changes in the cumulative inertial integration error while reducing invalid candidate trajectories that deviate from the actual motion trend of the vehicle, narrowing the search space for subsequent magnetic map matching, and improving the rationality of candidate trajectory generation and the efficiency of magnetic map matching.

[0026] 5. For each generated candidate trajectory, this application extracts the magnetic map region covering the candidate trajectory and its neighborhood from the pre-constructed marine magnetic map, and uses the ConvNeXt network to extract the depth spatial features of the magnetic map region corresponding to each candidate trajectory. After global average pooling, feature mapping and normalization processing, the corresponding map embedding is obtained. The query embedding generated by the real-time geomagnetic-inertial fusion features within the current matching time window and the map embedding corresponding to each candidate trajectory are mapped to a unified feature space. The depth feature similarity between the real-time observation features and the magnetic field spatial distribution features corresponding to each candidate trajectory is calculated by cosine similarity, and the candidate trajectory with the highest similarity is selected as the optimal matching trajectory. The spatial position of the optimal matching trajectory at the end of the current matching time window is used as the navigation position estimation result. Therefore, it can avoid relying solely on a single location or local original magnetic field value for direct matching, and make full use of the overall correlation between the continuous geomagnetic-inertial observation features and the magnetic field spatial distribution features corresponding to the candidate trajectory within the entire matching time window. This enhances the identification ability between different candidate trajectories, reduces the probability of mismatch when there are similar local magnetic field distributions in different geographical areas, and thus improves the robustness and positioning accuracy of magnetic map matching navigation in complex marine environments.

[0027] 6. This application adopts a self-supervised training method based on marine magnetomaps. During the model training phase, a virtual trajectory that satisfies the continuity constraint of the vehicle's motion is automatically generated using a pre-constructed marine magnetomap. A synthetic geomagnetic observation sequence is generated by sampling from the magnetomap along the virtual trajectory. The corresponding synthetic IMU inertial measurement sequence is generated by combining the virtual trajectory and the pre-set vehicle motion state model. At the same time, the magnetomap region covering the virtual trajectory and its neighborhood is extracted. The spatial correspondence between the synthetic geomagnetic-IMU time series observation and the corresponding magnetomap region is established using the same virtual trajectory, thereby automatically constructing positive and negative training samples. Furthermore, cross-modal contrastive learning is used to jointly train the geomagnetic feature extraction network, the inertial feature extraction network, the cross-modal fusion network, and the magnetomap feature extraction network. This makes the geomagnetic-IMU time series features and the magnetomap spatial features corresponding to the same virtual trajectory close to each other in a unified feature space, while the features corresponding to different trajectories are separated. Therefore, it is possible to automatically construct large-scale training samples using existing marine magnetic maps, reduce the dependence of deep magnetic map matching model training on large-scale real navigation trajectories and high-precision location labels, reduce the cost of collecting and labeling actual navigation data in the marine environment, and improve the engineering feasibility of deep magnetic map matching model training. Attached Figure Description

[0028] Figure 1 The flowchart shown is for a marine magnetograph matching navigation method based on self-supervised deep learning and multimodal fusion. Figure 2 The diagram shows the Transformer model architecture; Figure 3 The diagram shows the flowchart of a multi-source information feature extraction method based on Transformer. Figure 4 The diagram shows a flowchart of a geomagnetic-inertial information fusion method based on dynamic modulation of cross-attention using magnetic field gradient. Figure 5 The following is a flowchart of the magnetic map matching method based on depth feature similarity calculation; Figure 6 The diagram shown is a ConvNeXt model architecture diagram. Detailed Implementation

[0029] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0030] This application provides a marine magnetomap matching navigation method and system based on self-supervised deep learning and multimodal fusion. It utilizes a deep learning network to extract geomagnetic sequence features and inertial motion features, and introduces magnetic field gradient information to dynamically modulate cross-modal attention weights. Simultaneously, it combines IMU cumulative drift state to construct an adaptive gating fusion mechanism, enabling the fusion process to dynamically adjust according to environmental changes and sensor states. Furthermore, it uses pre-acquired marine magnetomaps to achieve navigation position estimation through depth feature matching. In the model training phase, this application further proposes a self-supervised training method based on marine magnetomaps, achieving automatic construction of training samples and self-supervised training of the depth matching model without requiring large-scale manual annotation of real navigation trajectories. Compared to existing methods based on manual features and single-modal matching, this application can fully utilize geomagnetic spatial distribution information, carrier motion constraint information, and sensor state information. Through self-supervised sample generation based on magnetomaps and a cross-modal comparative learning mechanism, it reduces the model's dependence on large-scale real navigation trajectories and high-precision position labels, improving the robustness and positioning accuracy of magnetomap matching in complex marine environments.

[0031] Example 1 This application proposes a marine magnetomap matching navigation method based on self-supervised deep learning and multimodal fusion. During the model training phase, a virtual trajectory that satisfies the continuity constraints of vehicle motion is automatically generated using a pre-acquired marine magnetomap. A synthetic geomagnetic observation sequence is generated by sampling from the marine magnetomap along the virtual trajectory, and a corresponding synthetic IMU inertial measurement sequence is generated by combining the virtual trajectory with a preset vehicle motion state model. Simultaneously, a magnetomap region covering the virtual trajectory and its neighborhood is extracted. The spatial correspondence between the synthetic geomagnetic-IMU time-series observations and the corresponding magnetomap region is established using the virtual trajectory. Positive and negative training samples are automatically constructed. Furthermore, the geomagnetic feature extraction network, inertial feature extraction network, cross-modal fusion network, and magnetomap feature extraction network are jointly trained through cross-modal comparative learning, thereby reducing the model training's dependence on large-scale real-world navigation trajectories and high-precision position labels.

[0032] During the actual navigation phase, geomagnetic and IMU information collected within the current matching time window are input into the trained deep network. High-dimensional temporal features of the geomagnetic and inertial information are extracted separately. The cross-modal correlation between geomagnetic and inertial features is dynamically modulated using magnetic field gradients, and the fusion weights of different modal information are adaptively adjusted based on the IMU's cumulative drift state to obtain the geomagnetic-IMU fusion query features corresponding to the current matching time window. Furthermore, inertial navigation integration is performed using the known initial state and IMU motion information of the current matching time window to obtain a coarse inertial trajectory of the vehicle. Candidate trajectories are adaptively determined based on the degree of IMU cumulative drift. The system searches for a trajectory and generates multiple candidate trajectories that satisfy inertial motion constraints within the search range. For each candidate trajectory, it extracts a magnetic map region covering the candidate trajectory and its neighborhood from a pre-constructed marine magnetic map. The corresponding magnetic field spatial distribution features are obtained through a trained magnetic map feature extraction network and deep feature similarity is calculated with the geomagnetic-IMU fusion query features. The optimal matching trajectory is determined based on the similarity, and the position of the optimal matching trajectory at the end of the current matching time window is taken as the current navigation position. This improves the robustness, positioning accuracy, and long-term autonomous navigation capability of magnetic map matching navigation in complex marine environments.

[0033] Example 2 like Figure 1 As shown, this application provides a marine magnetograph matching navigation method based on self-supervised deep learning and multimodal fusion, comprising: Step S1: In the real-time navigation data processing stage, real-time navigation data of the marine vehicle is acquired, including geomagnetic measurement data and inertial information collected by the IMU. The geomagnetic data is real-time acquisition of geomagnetic field information; the IMU continuously provides motion information such as angular velocity, acceleration, and attitude changes of the marine vehicle, providing continuous motion constraints for the navigation process. Subsequently, steps S2A and S2B are executed in parallel or sequentially.

[0034] Step S2A: To fully extract the temporal features from geomagnetic and inertial information, a magnetic encoder and an IMU encoder are constructed, respectively. Both encoders use a Transformer network as their basic structure and learn global dependencies in long-term sequences through a multi-head self-attention mechanism. The magnetic encoder focuses on extracting spatial variation patterns and local magnetic anomaly features from the geomagnetic sequence, while the IMU encoder focuses on learning the trajectory and inertial variation patterns of the marine vehicle, thereby obtaining high-dimensional geomagnetic and inertial features, respectively.

[0035] To further achieve effective fusion between the two modalities, a gradient-modulated cross-attention mechanism is introduced after the dual-channel Transformer encoder. This mechanism dynamically modulates the cross-attention weights using magnetic field gradient information. While establishing the correlation between geomagnetic and inertial information, it introduces physical constraints based on the spatial variation characteristics of the magnetic field, giving higher feature attention weights to regions with significant magnetic field changes, thereby improving the effectiveness and robustness of cross-modal feature fusion.

[0036] After completing cross-attention fusion, a drift-aware gated fusion module is further introduced. This module adaptively adjusts the contribution ratio of geomagnetic and inertial features in the fusion process based on the IMU's cumulative drift state. When the IMU's cumulative drift is small, the weight of inertial information is appropriately increased; when the cumulative drift increases, the geomagnetic spatial constraint effect is automatically enhanced to reduce the impact of long-term inertial navigation integration error on the positioning result. Subsequently, attention pooling is performed on the fused sequence features to generate a query embedding that comprehensively reflects geomagnetic spatial information and the carrier's motion state, serving as a unified feature representation of the real-time navigation data.

[0037] Step S2B: To fully utilize IMU motion information to constrain the trajectory search space for subsequent magnetic map matching, inertial navigation integration is performed based on the known state at the start of the current matching time window and the continuously acquired IMU inertial measurement data within that time window. This process continuously recursively extrapolates the vehicle's attitude, velocity, and position to obtain inertial navigation position estimates for each sampling moment within the current matching time window, thereby forming a coarse inertial trajectory for the vehicle. This coarse inertial trajectory reflects the overall motion trend of the vehicle within the current matching time window, determined by IMU measurement information, and provides motion priors for generating candidate trajectories.

[0038] After obtaining the coarse inertial trajectory, the candidate trajectory search area is adaptively determined by combining the IMU cumulative drift evaluation results. When the IMU cumulative drift is small, the coarse inertial trajectory obtained by inertial integration has high reliability, thus narrowing the surrounding candidate trajectory search range. When the IMU cumulative drift increases, the uncertainty of the inertial trajectory position increases accordingly, thereby expanding the candidate trajectory search range. Within the candidate trajectory search area, multiple candidate trajectories are discretized and generated using the coarse inertial trajectory as the motion prior, satisfying the known initial state constraints, inertial motion trend constraints, and trajectory continuity constraints, forming a candidate trajectory set.

[0039] For each generated candidate trajectory, based on the spatial location of the candidate trajectory in the ocean magnetic map, a local magnetic map region covering the candidate trajectory and its neighborhood is extracted from the pre-constructed ocean magnetic map, thereby establishing a one-to-one correspondence between the candidate trajectory and the local magnetic map region.

[0040] For the local magnetic maps corresponding to each candidate trajectory, a ConvNeXt network is used for deep spatial feature extraction. The deep feature map of the local magnetic map is obtained through multi-layer convolutional coding, and the spatial features are aggregated through global average pooling. Then, through linear mapping and normalization, a map embedding in the same feature space as the query embedding is obtained.

[0041] Step S3: In the magnetograph matching stage, the query embedding corresponding to the real-time navigation data within the current matching time window is compared with the map embedding corresponding to each candidate trajectory using cosine similarity calculation. This yields the depth feature similarity between the real-time geomagnetic-inertial fusion features and the spatial distribution features of the magnetic field corresponding to each candidate trajectory. The candidate trajectory with the highest similarity is selected as the optimal matching trajectory, and the spatial position of the optimal matching trajectory at the end of the current matching time window is taken as the current magnetograph matching navigation position estimation result of the vehicle.

[0042] Example 3 This application provides a marine magnetograph matching navigation method based on self-supervised deep learning and multimodal fusion, implemented according to the method of Embodiment 1, and further comprising: 1. The multi-source information feature extraction method based on Transformer in step S2A (e.g.) Figure 3 (as shown) This application employs a multi-source information feature extraction method based on the Transformer architecture. Temporal feature models are performed on geomagnetic observation information and inertial measurement information respectively. Spatial correlation features and motion change features in long-term series are extracted through a self-attention mechanism, providing high-dimensional feature representations for subsequent geomagnetic-inertial information fusion and magnetic map matching. The Transformer architecture diagram is shown below. Figure 2 As shown.

[0043] Let the current magnetogram matching time window be... At the start of the time window, the ocean vehicle The state is a known state: (1) The synchronous acquisition length within the time window is The geomagnetic observation sequences and IMU inertial measurement sequences are both types of sequences with corresponding timestamps.

[0044] (1) Geomagnetic information sequence construction and feature embedding The offshore platform is equipped with a triaxial magnetometer to collect geomagnetic field information in real time, and the sampling time is set. The geomagnetic observations are as follows: (2) in, , , These represent the magnetic field components in the three directions of the carrier coordinate system. A geomagnetic sequence is constructed using a fixed time window. (3) in, This represents the length of the time window.

[0045] Definition of the first Magnetic field gradient at each sampling point: (4) in, The displacement distance of the aircraft is estimated based on the IMU inertial integration results at adjacent sampling times. To prevent constants with a denominator of zero.

[0046] Because the Transformer network requires the input to have a uniform dimension, the original triaxial magnetic field data and magnetic field gradients are transformed into a high-dimensional feature space through a linear mapping: (5) in, This is the geomagnetic feature mapping matrix. The gradient feature mapping matrix, go through Projected to the same feature dimension as the geomagnetic feature embedding. , For bias parameters, This is an embedding of geomagnetic features.

[0047] To preserve the temporal location information in the geomagnetic sequence, location encoding is introduced: (6) in, The position encoding matrix is ​​in sine and cosine form: (7) (8) in, pos This is the location index, representing the position number (0, 1,...,...) of the current data point within the time window sequence. N -1); i The dimension index represents the first element in the feature vector. i Each dimension (0, 1, ..., ), d This indicates the number of dimensions in the feature vector.

[0048] (2) Geomagnetic feature encoding based on Transformer The geomagnetic embedding features are input into the Transformer encoder, and the magnetic field variation patterns at different time scales are learned through a multi-head self-attention mechanism.

[0049] For the The layer Transformer encoder, whose query matrix, key matrix, and value matrix are represented as follows: (9) in, , , These are the learnable parameter matrices used to generate the Query, Key, and Value, respectively.

[0050] Calculate the self-attention weights: (10) Geomagnetic timing characteristics were obtained: (11) To enhance the model's expressive power, multi-head attention is employed: (12) in, This represents the multi-head output projection matrix; Indicates the first i One point of attention: (13) The final high-dimensional representation of geomagnetic features is obtained: (14) This feature contains information on the spatial variation patterns of the geomagnetic field and its continuous temporal variation.

[0051] (3) IMU inertial information sequence modeling The inertial measurement unit outputs information about the carrier's motion state, including triaxial angular velocity and triaxial acceleration. (15) in, , , This represents the acceleration in three directions within the carrier's coordinate system. , , This represents the angular velocity in three directions within the carrier coordinate system.

[0052] Build length is N Inertial sequence: (16) Inertial feature embedding is obtained through linear embedding: (17) in, This is the inertial feature mapping matrix; These are bias parameters; Add time location encoding: (18) Then, the attention calculation process is performed on the input IMU Transformer encoder as follows: (19) get: (20) Ultimately, the characteristics of inertial motion are obtained: (twenty one) This feature is used to describe the attitude changes, motion trends, and trajectory continuity of an offshore platform.

[0053] (4) Multi-source feature normalization and output After processing by the geomagnetic Transformer encoder and the IMU Transformer encoder, geomagnetic coded features and inertial coded features are obtained respectively: (twenty two) in, and These are the characteristic lengths of the geomagnetic sequence and the inertial sequence, respectively. and These are the characteristic dimensions of the geomagnetic sequence and the inertial sequence, respectively.

[0054] Since geomagnetic features and inertial features may have different dimensions, linear mapping is used to project them onto a dimension of . Public characteristic space: (twenty three) in, , For a trainable modality mapping matrix, d This refers to the common feature dimension, i.e., the unified target feature dimension before fusion; and These are the bias vectors for the geomagnetic sequence and the inertial sequence, respectively.

[0055] Layer normalization is performed on the two types of features after projection: (twenty four) This allows us to obtain geomagnetic features located in the same feature space. With inertial characteristics This serves as the input for subsequent cross-attention fusion and magnetic map matching.

[0056] 2. The geomagnetic-inertial information fusion method based on dynamic modulation of cross-attention by magnetic field gradient in step S2A (e.g.) Figure 4 (as shown) (1) Calculation of geomagnetic cross attention guided by inertial information Using geomagnetic features as query information and inertial features as keys and values, geomagnetic cross-attention guided by inertial information is calculated. The corresponding query matrix, key matrix, and value matrix are as follows: (25) in, , and This is a trainable parameter matrix.

[0057] The cross-modal correlation matrix is ​​as follows: (26) in, The dimension of the key vector, and the scaling factor. Used to reduce the fluctuation of dot product values ​​as the feature dimension increases. The gradient modulation matrix is ​​obtained by expanding the magnetic field gradient sequence, and its dimension is consistent with the cross-mode correlation matrix. This represents the gradient adjustment parameter.

[0058] Perform Softmax normalization on the correlation matrix row by row to obtain the attention weights of inertial information for local magnetic moments: (27) in: (28) Indicates the first The geomagnetic features and the first The degree of correlation between inertial features.

[0059] The inertia value features are weighted and summed based on attention weights: (29) The geomagnetic characteristics modulated by inertial motion information were obtained. These characteristics enable the correlation between changes in the magnetic field and changes in acceleration, angular velocity, and attitude during the corresponding time period.

[0060] (2) Calculation of inertial cross-attention guided by geomagnetic information To simultaneously utilize the spatial variations of the geomagnetic field to constrain inertial motion characteristics, inertial characteristics are used as queries, and geomagnetic characteristics are used as keys and values ​​to calculate reverse cross-attention: (30) The cross-modal correlation matrix is ​​as follows: (31) The corresponding attention weights are: (32) This leads to the inertial characteristics of enhanced geomagnetic information: (33) It is used to describe the motion state of a carrier corresponding to changes in local magnetic field, reducing the impact of error accumulation on feature representation when modeling based solely on inertial sequences.

[0061] (3) Multi-head cross-attention calculation To extract the correlation between geomagnetic and inertial information from different feature subspaces, a multi-head cross-attention mechanism is adopted.

[0062] For the The geomagnetic cross-attention guided by inertial information is represented as follows: (34) in, , For the number of heads.

[0063] The calculation results of each attention point are concatenated and transformed using the output mapping matrix: (35) in, This represents the multi-head cross-attention output, which is the geomagnetic enhancement feature resulting from splicing together the results of all heads and mapping.

[0064] Similarly, the inertial multi-head cross-attention guided by geomagnetic information is: (36) (37) Different attention heads can learn the correlation characteristics between changes in magnetic field amplitude, magnetic field gradient, attitude change, angular velocity change, and acceleration change, respectively.

[0065] (4) Residual connection and nonlinear characteristic transformation The cross-attention output is residually concatenated with the original modality features to obtain the initial enhanced features: (38) Subsequently, the initially enhanced features are input into the trained feedforward neural network. The feedforward neural network is represented as follows: (39) in, For non-linear activation functions, the following can be used: function: (40) After passing through the feedforward network and the second residual connection, the bidirectional cross-enhanced feature is obtained: (41) (5) Adaptive gating fusion based on IMU cumulative drift sensing To dynamically adjust the contributions of geomagnetic and inertial characteristics based on different navigation states and magnetic field environments, a mode fusion gating coefficient is constructed. Since IMU inertial measurements obtain velocity and position information through integration, measurement errors accumulate over time, leading to a decrease in the reliability of inertial characteristics. Therefore, an IMU cumulative drift index is introduced into the gating fusion process to describe the uncertainty of the current inertial navigation state, and the fusion ratio of geomagnetic and inertial information is dynamically adjusted based on this index.

[0066] The error state vector can be defined as: (42) in, Indicates positional error. Indicates speed error, Indicates attitude error. This indicates the zero bias error of the accelerometer. This indicates the zero bias error of the gyroscope.

[0067] The covariance matrix of the error state vector is: (43) The elements on the diagonal represent the variances of the various error quantities: (44) in, Represents the trace of a matrix; The system inertial navigation error estimate is expressed as: (45) in, Indicates the first The cumulative drift metric of the IMU at any given time. The inertial navigation error propagation model is established based on IMU measurement noise, gyroscope bias, and accelerometer bias, and the error covariance matrix is ​​propagated synchronously according to the inertial navigation state recursion process.

[0068] geomagnetic enhancement features Inertial enhancement characteristics and IMU cumulative drift characteristics Common input gating network: (46) in, Represents the parameter matrix of the gated network. Indicates the bias term. Indicates the first Adaptive retention weights for time-varying geomagnetic modes.

[0069] The final fusion feature sequence is as follows: (47) in, This indicates element-wise multiplication.

[0070] When the cumulative IMU drift is small, the gating network maintains a high contribution of inertial information; when the cumulative IMU drift increases, the weight of inertial information is reduced and the contribution of geomagnetic spatial constraint information is increased, thereby reducing the impact of long-term inertial navigation error on the matching result.

[0071] (6) Generation of fused query vectors In order to use the fused features for subsequent magnetic map candidate region matching, attention pooling is applied to the fused feature sequence.

[0072] First, calculate the aggregate weights at each time step: (48) in, Indicates the first p The fusion characteristics at each moment, , and All of these are trainable parameters.

[0073] The fused query vector is: (49) Further linear mapping of the query vector and Normalization: (50) in, For query embedding used for ocean magnetic map matching, A positive number set to prevent the denominator from being zero.

[0074] 3. The candidate trajectory generation method based on the known initial state and IMU motion constraints in step S2B; To fully utilize the constraint capability of inertial measurement information on the short-term motion state of marine vehicles and reduce the trajectory search range in the subsequent magnetograph matching process, this application obtains a coarse inertial trajectory of the vehicle within the current matching time window based on the known state at the start of the current matching time window and the continuously acquired IMU inertial measurement data within that time window through inertial navigation integral recursion. Based on this, the search range of candidate trajectories around the coarse inertial trajectory is determined according to the cumulative drift of the IMU. Using the coarse inertial trajectory as a motion prior, multiple candidate trajectories satisfying the initial state constraint and motion continuity constraint are generated within the candidate trajectory search range, forming a candidate trajectory set. This provides a spatial search basis for the subsequent depth feature extraction and matching of the magnetograph regions corresponding to the candidate trajectories.

[0075] (1) Definition of matching time window and initial state Let the current magnetogram matching time window be: (51) in, This represents the start time of the current matching time window. This represents the end time of the current matching time window.

[0076] Known aircraft in The initial state at time t, including initial position, initial velocity, and initial attitude, are represented as follows: ,in: (52) This indicates the initial position of the vehicle in the navigation coordinate system. Indicates the initial velocity. This represents the initial attitude rotation matrix from the vehicle coordinate system to the navigation coordinate system. The initial state is obtained from the system initialization results.

[0077] (2) Recursive motion state based on IMU inertial integral The IMU continuously acquires triaxial angular velocity and triaxial acceleration within the current matching time window: (53) in, .

[0078] After considering the zero bias of the gyroscope and the zero bias of the accelerometer, the IMU measurements are corrected: (54) in, To achieve zero bias in the gyroscope, This is for zero bias of the accelerometer.

[0079] First, the vehicle attitude is recursively calculated based on the corrected angular velocity: (55) in, The time interval between adjacent IMU sampling times. This represents an exponential mapping that maps a rotation vector to a rotation matrix.

[0080] The acceleration in the vehicle coordinate system is transformed to the navigation coordinate system using the current attitude, and gravity compensation is performed: (56) in, This is the gravitational acceleration vector in the navigation coordinate system. Based on this, the velocity is obtained by integrating the acceleration: (57) Further integration of the velocity yields the position: (58) By continuously recursively applying all IMU data within the current matching time window, the inertial navigation position estimate for each sampling moment within the time window is obtained, and a coarse inertial trajectory of the vehicle is constructed: (59) (3) Construction of candidate trajectory search region Establish candidate trajectory search region: (60) in, This represents the maximum spatial offset range of the candidate trajectories corresponding to the current matching time window. Indicates position coordinates, The geographical area covered by the magnetic map.

[0081] To adapt to the characteristic that the IMU integral error changes with navigation time and motion state, the candidate trajectory search range can be adaptively adjusted according to the current cumulative IMU drift.

[0082] The cumulative drift index of the IMU obtained during the aforementioned drift-sensing gating fusion process ,structure: (61) in, Time weighting coefficient: (62) in, The preset base search radius; This refers to the cumulative drift metric of the IMU corresponding to the current matching time window. This is the search range adjustment coefficient. When the IMU cumulative drift is small, the coarse inertial trajectory has high reliability, so a smaller candidate trajectory search range is used. When the IMU cumulative drift increases, the candidate trajectory is allowed to vary in a wider area to reduce the probability that the real trajectory exceeds the candidate trajectory search range due to inertial integration error.

[0083] (4) Discrete generation of candidate trajectories In the candidate trajectory search area Inside, with a rough inertial trajectory As a priori motion, it generates a discrete perturbation of trajectory position deviation, heading deviation, or motion scale deviation. Candidate trajectories: (63) Among them, the r Candidate trajectories : (64) No. The first candidate trajectory The following positions satisfy: (65) in, Let be the position disturbance, i.e., the th r The trajectory in the first k The positional offset of a given moment relative to the IMU trajectory; Because the starting point is known: (66) Right now: (67) 4. The magnetic map matching method based on depth feature similarity calculation in step S2B (e.g.) Figure 5 (as shown) (1) Construction of magnetic map regions corresponding to candidate trajectories Let the pre-constructed ocean magnetic map be represented as: (68) in, The geographical area covered by the magnetic map. For position The magnetic field properties at a location. A magnetic map consists of the total magnetic field strength, three-axis magnetic field components, and magnetic field gradient.

[0084] Based on the previously obtained set of candidate trajectories: (69) For the Candidate trajectories: (70) Based on the spatial coordinates of each trajectory point in the candidate trajectories, a pre-constructed ocean magnetic map is used. The magnetic map area covering the entire candidate trajectory is determined. To preserve the spatial distribution information of the magnetic field within a certain range around the candidate trajectory, a preset spatial extension width is set on both sides of the candidate trajectory. Thus, the first The magnetic map area corresponding to each candidate trajectory: (71) in, Indicates based on trajectory Constructed ocean magnetic map; This indicates a cropping operation, which extracts a local sub-image from a larger image based on the trajectory coordinates.

[0085] Therefore, the set of candidate trajectories corresponds to the set of magnetic map regions: (72) (2) ConvNeXt network extracts candidate magnetic map features The magnetic map area corresponding to each candidate trajectory Input ConvNeXt network (network structure as follows) Figure 6 As shown), initial features are first extracted through a convolutional embedding layer: (73) in, This indicates the kernel size and downsampling stride. The initial feature map representing the candidate magnetic map region.

[0086] For the first in the ConvNeXt network Each encoded block is used to extract local spatial features using depthwise convolution: (74) The depthwise convolution output is then subjected to layer normalization, channel expansion, non-linear activation, and channel compression in sequence: (75) The output of the current coded block is obtained through residual concatenation: (76) in, Represents depthwise convolution. Representation layer normalization, and These represent the channel expansion matrix and the channel compression matrix, respectively. This indicates a learnable scaling parameter.

[0087] go through After encoding 10 ConvNeXt blocks, the depth features of the candidate magnetic map region are obtained: (77) in, This represents the parameters of the ConvNeXt network.

[0088] (3) Generate candidate magnetic map feature embedding depth features of candidate magnetic maps Perform global average pooling: (78) in, and These represent the height and width of the depth feature map, respectively.

[0089] Map the pooled features to the fused query vector. Same feature dimensions: (79) in, and These represent the map feature mapping matrix and the bias vector, respectively.

[0090] Perform mapping on the features Normalization: (80) in, Indicates the first Normalized feature embedding of candidate magnetic map regions A positive number set to prevent the denominator from being zero.

[0091] 5. Calculate the depth feature similarity in step S3; Calculate the fusion query vector With the embedding of candidate magnetomap features Cosine similarity between them: (81) because and All have undergone Normalization, therefore simplifies to: (82) in, Indicates the fusion query features and the first The similarity of deep features among candidate magnetic map regions.

[0092] Determine the optimal candidate trajectory and output the navigation position: Select the candidate region with the highest similarity from all candidate magnetic map regions: (83) The optimal candidate trajectory is determined based on the one-to-one correspondence between candidate trajectories, local magnetic maps, and candidate map embeddings: (84) and the optimal candidate trajectory At the end of the current matching time window Corresponding position As the current magnetograph matching navigation location estimation result: (85) in, Indicates the index of the optimal candidate trajectory. This indicates the magnetograph matching position of the carrier at the current moment.

[0093] 6. Self-supervised training method based on magnetic map (1) Virtual trajectory generation based on ocean magnetic map In a pre-built ocean magnetic map A virtual starting position is randomly selected within the coverage area, and a virtual trajectory that satisfies motion continuity constraints is generated according to the preset sailing speed, heading, and range of motion changes. (86) in, Indicates the first A virtual trajectory, Indicates the first The virtual trajectory in the first The location of each sampling time. express x Axis coordinates express y Axis coordinates.

[0094] Adjacent virtual trajectory points satisfy preset constraints on maximum displacement and maximum heading changes to avoid generating trajectories that significantly deviate from the actual motion characteristics of the vessel. By changing the virtual starting position, initial heading, speed, and motion state change parameters, a large number of different virtual navigation trajectories can be automatically generated from the same ocean magnetic map.

[0095] (2) Generation of synthetic geomagnetic-IMU time series data For the Virtual trajectory Based on the spatial coordinates of each virtual trajectory point in the ocean magnetomap, the magnetomap is interpolated and sampled to obtain the corresponding synthetic geomagnetic observation sequence: (87) Subsequently, the velocity is calculated based on the relationship between adjacent positions in the virtual trajectory and time: (88) in, It represents the time difference between adjacent positions.

[0096] Further calculation of motion acceleration: (89) Calculate the heading angle based on the direction of motion: (90) And calculate the angular velocity based on the change in heading at adjacent moments: (91) Based on the virtual trajectory and the preset vehicle motion state model, corresponding attitude, velocity, and acceleration states are generated. Combined with the gravity model, IMU measurement model, and preset sensor error model, a synthetic IMU inertial measurement sequence consisting of three-axis acceleration and three-axis angular velocity is constructed. (92) (3) Automatic construction of magnetic map regions corresponding to trajectories and positive and negative samples For virtual trajectory Using the same magnetic map region construction method as before, magnetic map regions covering the virtual trajectory and its neighborhood are extracted from the pre-constructed ocean magnetic map as positive sample map tiles: (93) in, Expand the width of the trajectory neighborhood.

[0097] because and All are from the same virtual trajectory Since they are automatically generated, they have a definite spatial correspondence, and can be constructed as positive sample pairs: (94) Simultaneously, select from the magnetic map regions corresponding to virtual trajectories in other spatial locations. As negative sample map tiles, thus forming: (95) (4) Cross-modal comparative learning of geomagnetic-inertial and magnetic maps Synthetic geomagnetic sequence and synthetic IMU sequences Input the geomagnetic-inertial feature extraction and fusion network to obtain the normalized query embedding. The positive sample magnetic map area and negative sample magnetic map regions The corresponding normalized map embeddings are obtained by inputting them into the ConvNeXt magnetic map encoding network. Contrastive learning loss is used: (96) in, The number of samples in a training batch. The number of negative samples corresponding to each positive sample. For temperature coefficient, For positive sample similarity, For the first Similarity of negative samples.

[0098] By minimizing ,make at the same time .

[0099] (5) Joint optimization of network parameters During training, the model parameters are jointly optimized using the backpropagation algorithm based on the contrastive learning loss, including: geomagnetic Transformer encoder parameters; IMU Transformer encoder parameters; magnetic field gradient modulation cross-attention model parameters; drift-aware gating fusion model parameters; attention pooling and query embedding mapping parameters; ConvNeXt magnetic map encoder parameters; and map embedding mapping parameters.

[0100] After training, the model parameters are fixed, and the trained model is used for actual ocean magnetic map matching navigation. During the online navigation phase, there is no need to generate training positive and negative samples; only the real-time geomagnetic data and IMU data are processed according to the aforementioned method, and the optimal matching trajectory and the current navigation position are determined from the generated candidate trajectories.

[0101] The marine magnetograph matching navigation method based on self-supervised deep learning and multimodal fusion provided in this application has the following technological innovations: 1. A method for representing geomagnetic time series features based on magnetic field gradient enhancement This application incorporates magnetic field gradient information during the geomagnetic sequence construction process. By calculating the degree of magnetic field change between adjacent sampling locations, magnetic field gradient features are constructed, and the original geomagnetic observations and magnetic field gradients are jointly mapped to a high-dimensional feature space. By inputting the magnetic field gradient as an auxiliary feature into the geomagnetic coding network, the model can perceive the spatial variation characteristics of the magnetic field in different regions, improving its ability to represent features of areas with obvious location identification capabilities.

[0102] 2. A geomagnetic-inertial information fusion method based on dynamic modulation of cross-attention by magnetic field gradient This application proposes a magnetic field gradient dynamic modulation cross-attention mechanism. When calculating the cross-attention between geomagnetic and inertial information, a modulation matrix generated by the magnetic field gradient is introduced to dynamically adjust the cross-modal correlation matrix. This method ensures that the cross-attention weights not only depend on the correlation between geomagnetic and inertial features but are also constrained by the degree of spatial variation in the magnetic field, thus granting higher information contribution weights to regions with significant magnetic field variations.

[0103] 3. Adaptive Gated Fusion Method Based on IMU Accumulated Drift Sensing To address the issue of decreased reliability of inertial information due to the continuous accumulation of errors during long-term IMU integration, this application proposes a gated fusion method that considers the accumulated IMU drift state. By establishing an IMU accumulated drift index, the inertial navigation error state is introduced into the modal fusion process, enabling the gated network to dynamically adjust the fusion ratio of geomagnetic and inertial information based on the current reliability of the inertial information. This method can dynamically adjust the fusion weights according to navigation time and inertial error state, improving stability during long-term navigation.

[0104] 4. An Adaptive Candidate Trajectory Generation Method Based on IMU Motion Constraints and Cumulative Drift States To address the issues of a large search space and the difficulty of adapting a fixed search range to the cumulative changes in IMU inertial integration error over time during large-scale magnetic map matching, this application utilizes the known state of the vehicle at the start of the current matching time window and continuously acquired IMU inertial measurement data within the time window. By continuously recursively extrapolating the vehicle's motion state through inertial navigation integration, a coarse inertial trajectory within the current matching time window is obtained. Based on this, the search range of candidate trajectories around the coarse inertial trajectory is adaptively determined according to the IMU cumulative drift index: when the IMU cumulative drift is small, the search range is narrowed; when the IMU cumulative drift increases, the search range is correspondingly expanded. Subsequently, using the coarse inertial trajectory as a motion prior, multiple candidate trajectories satisfying the known initial state constraint, inertial motion trend constraint, and trajectory continuity constraint are discretized within the candidate trajectory search range, forming a candidate trajectory set. By introducing the IMU continuous motion constraint and cumulative drift state into the candidate trajectory generation process, the trajectory search space for subsequent magnetic map matching is narrowed while preserving the possibility of true trajectories, reducing invalid candidate trajectories, and improving the efficiency and reliability of magnetic map matching.

[0105] 5. Magnetic map matching method based on deep feature space This application, for each candidate trajectory in the candidate trajectory set, extracts the magnetic map region covering the candidate trajectory and its neighborhood based on its spatial location in a pre-constructed marine magnetic map. The ConvNeXt network is then used to extract the depth spatial features of the magnetic map region corresponding to the candidate trajectory. After global average pooling, feature mapping, and normalization, the map embedding corresponding to the candidate trajectory is obtained. Subsequently, the query embedding generated from the real-time geomagnetic-inertial fusion features within the current matching time window is mapped to a unified feature space along with the map embeddings corresponding to each candidate trajectory. The depth feature similarity between the query embedding and the spatial distribution features of the magnetic field corresponding to each candidate trajectory is calculated using cosine similarity. The candidate trajectory with the highest similarity is selected as the optimal matching trajectory. The spatial position of the optimal matching trajectory at the end of the current matching time window is used as the navigation position estimation result. By utilizing the overall correlation between the real-time observation features and the spatial distribution features of the magnetic field corresponding to the candidate trajectory within the entire matching time window, the probability of mismatches caused by local magnetic field distribution similarities is reduced, improving the positioning accuracy and robustness of magnetic map matching in complex marine environments.

[0106] 6. A self-supervised training method based on virtual trajectory generation of ocean magnetic maps To address the issue that training deep magnetic map matching models relies heavily on a large number of real navigation trajectories and high-precision location labels, and that acquiring and labeling measured data in marine environments is costly, this application proposes a self-supervised training method based on virtual trajectory generation from a marine magnetic map. Utilizing a pre-acquired marine magnetic map as the basis for training data generation, a virtual trajectory satisfying the continuity constraints of vehicle motion is automatically generated within the magnetic map coverage area. Magnetic field information is sampled from the magnetic map along the virtual trajectory to generate a synthetic geomagnetic observation sequence corresponding to the virtual trajectory. Simultaneously, a corresponding synthetic IMU inertial measurement sequence is generated by combining the virtual trajectory and a pre-defined vehicle motion state model, and the magnetic map region covering the virtual trajectory and its neighborhood is extracted, thereby automatically establishing a spatial correspondence between geomagnetic-IMU time-series observations and the magnetic map region corresponding to the trajectory. Based on this, geomagnetic-IMU time-series data generated from the same virtual trajectory and magnetic map regions are constructed as positive sample pairs, while magnetic map regions corresponding to other trajectories with different spatial locations are constructed as negative samples. Through cross-modal contrastive learning, the geomagnetic encoder, IMU encoder, cross-modal fusion network, and magnetic map encoding network are jointly optimized, enabling the model to learn a unified deep feature representation between geomagnetic-inertial time-series features and the spatial distribution features of the magnetic field corresponding to the trajectory. This reduces the model training's dependence on large-scale real navigation trajectories and high-precision position labels, improves the automatic construction capability of training data, and enhances the trainability and engineering applicability of the model in marine magnetic map matching tasks.

[0107] Example 4 This application also provides a marine magnetograph matching navigation system based on self-supervised deep learning and multimodal fusion, used in the above method, including: The information acquisition module is used to acquire real-time navigation data, including geomagnetic measurement data and inertial information; The query embedding module is used to input geomagnetic measurement data into a trained magnetic encoder and output high-dimensional geomagnetic features; it also inputs inertial information into a trained IMU encoder and outputs high-dimensional inertial features. A magnetic field gradient modulation cross-attention mechanism is employed, dynamically modulating the cross-attention weights using magnetic field gradient information. This establishes the correlation between geomagnetic and inertial information while introducing physical constraints based on the spatial variation characteristics of the magnetic field. After cross-attention fusion, a drift-sensing gating fusion method is used, adaptively adjusting the contribution ratio of geomagnetic and inertial features in the fusion process based on the accumulated drift state of the IMU. Subsequently, attention pooling is applied to the fused sequence features to generate a query embedding that comprehensively reflects both geomagnetic spatial information and the carrier's motion state. The map embedding module is used to perform inertial navigation integration based on the known state at the start of the current matching time window and the continuously collected inertial measurement data within that time window. This continuously recursively extrapolates the vehicle's attitude, velocity, and position to obtain inertial navigation position estimates for each sampling moment within the current matching time window, forming a coarse inertial trajectory for the vehicle. After obtaining the coarse inertial trajectory, the candidate trajectory search area is adaptively determined by combining the IMU cumulative drift evaluation results. Within the candidate trajectory search area, multiple paths are discretized using the coarse inertial trajectory as the motion prior, constructing paths that satisfy the constraints of the known initial state, inertial motion trend, and trajectory continuity. Candidate trajectories are generated to form a candidate trajectory set. For each generated candidate trajectory, based on its spatial location in the ocean magnetomap, a local magnetomap region covering the candidate trajectory and its neighborhood is extracted from the pre-constructed ocean magnetomap, establishing a one-to-one correspondence between the candidate trajectory and the local magnetomap region. The local magnetomaps corresponding to each candidate trajectory are input into the trained ConvNeXt network for deep spatial feature extraction to obtain the depth feature map of the local magnetomap. The spatial features of the depth feature map are aggregated through global average pooling, and through linear mapping and normalization, a map embedding in the same feature space as the query embedding is obtained. The navigation position estimation module is used to calculate the similarity between the query embedding corresponding to the real-time navigation data within the current matching time window and the map embedding corresponding to each candidate trajectory, to obtain the depth feature similarity between the real-time geomagnetic-inertial fusion features and the magnetic field spatial distribution features corresponding to each candidate trajectory; the candidate trajectory with the highest similarity is selected as the optimal matching trajectory, and the spatial position corresponding to the optimal matching trajectory at the end of the current matching time window is taken as the current magneto-map matching navigation position estimation result of the vehicle.

[0108] This application may also provide a computer device, including: at least one processor, memory, at least one network interface, and a user interface. The various components in this device are coupled together via a bus system. It is understood that the bus system is used to implement communication between these components. In addition to a data bus, the bus system also includes a power bus, a control bus, and a status signal bus.

[0109] The user interface can include a display, keyboard, or clicking device. Examples include a mouse, trackball, touchpad, or touchscreen.

[0110] It is understood that the memory in the embodiments disclosed in this application may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory may be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DRRAM). The memories described herein are intended to include, but are not limited to, these and any other suitable types of memory.

[0111] In some implementations, the memory stores elements such as executable modules or data structures, or subsets thereof, or extended sets thereof: operating systems and applications.

[0112] The operating system includes various system programs, such as the framework layer, core library layer, and driver layer, used to implement various basic business functions and handle hardware-based tasks. The application programs include various applications, such as media players and browsers, used to implement various application functions. Programs implementing the methods of the embodiments of this disclosure can be included in the application programs.

[0113] In the above embodiments, the processor can also invoke programs or instructions stored in memory, specifically programs or instructions stored in an application program, for the following purposes: Follow the steps described above.

[0114] The above methods can be applied to or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above methods can be completed by integrated logic circuits in the processor's hardware or by software instructions. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic diagrams disclosed above. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the disclosed methods can be directly implemented by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0115] It is understood that the embodiments described in this application can be implemented using hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described in this application, or combinations thereof.

[0116] For software implementation, the technology of this application can be implemented by executing the functional modules (e.g., procedures, functions, etc.) of this application. The software code can be stored in memory and executed by a processor. The memory can be implemented in the processor or outside the processor.

[0117] This application may also provide a non-volatile storage medium for storing a computer program. When the computer program is executed by a processor, it can implement the steps in the above method embodiments.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of this application do not depart from the spirit and scope of the technical solutions of this application, and should all be covered within the scope of the claims of this application.

Claims

1. A marine magnetograph matching navigation method based on self-supervised deep learning and multimodal fusion, comprising: Step S1: Obtain real-time navigation data, including geomagnetic measurement data and inertial information, and execute steps S2A and S2B; Step S2A: Input the geomagnetic measurement data into the trained magnetic encoder and output the geomagnetic high-dimensional features; input the inertial information into the trained IMU encoder and output the inertial high-dimensional features. A magnetic field gradient modulation cross-attention mechanism is adopted, which dynamically modulates the cross-attention weights using magnetic field gradient information. While establishing the correlation between geomagnetic information and inertial information, physical constraints of the spatial variation characteristics of the magnetic field are introduced. After completing the cross-attention fusion, a drift-sensing gating fusion method is adopted, which adaptively adjusts the contribution ratio of geomagnetic features and inertial features in the fusion process by combining the accumulated drift state of the IMU. Subsequently, attention pooling is performed on the fused sequence features to generate a query embedding that can comprehensively reflect geomagnetic spatial information and the motion state of the carrier. Step S2B: Based on the known state at the start of the current matching time window and the inertial measurement data continuously collected within the time window, perform inertial navigation integration, continuously recursively calculate the attitude, velocity and position state of the vehicle, obtain the inertial navigation position estimate at each sampling time within the current matching time window, and form a coarse inertial trajectory of the vehicle. After obtaining the coarse inertial trajectory, the candidate trajectory search area is adaptively determined by combining the IMU cumulative drift evaluation results. Within the candidate trajectory search area, multiple candidate trajectories are generated discretely using the coarse inertial trajectory as the motion prior, satisfying known initial state constraints, inertial motion trend constraints, and trajectory continuity constraints, forming a candidate trajectory set. For each generated candidate trajectory, based on its spatial location in the ocean magnetomap, a local magnetomap region covering the candidate trajectory and its neighborhood is extracted from the pre-constructed ocean magnetomap, establishing a one-to-one correspondence between the candidate trajectory and the local magnetomap region. The local magnetomaps corresponding to each candidate trajectory are input into the trained ConvNeXt network for deep spatial feature extraction, obtaining the depth feature map of the local magnetomap. The spatial features of the depth feature map are aggregated through global average pooling, and through linear mapping and normalization, a map embedding in the same feature space as the query embedding is obtained. Step S3: Calculate the similarity between the query embedding corresponding to the real-time navigation data within the current matching time window and the map embedding corresponding to each candidate trajectory, and obtain the deep feature similarity between the real-time geomagnetic-inertial fusion features and the magnetic field spatial distribution features corresponding to each candidate trajectory. The candidate trajectory with the highest similarity is selected as the optimal matching trajectory, and the spatial position of the optimal matching trajectory at the end of the current matching time window is taken as the current magnetograph matching navigation position estimation result of the vehicle.

2. The marine magnetograph matching navigation method based on self-supervised deep learning and multimodal fusion according to claim 1, characterized in that, The geomagnetic measurement data of the trained magnetic encoder input in step S2A is represented as follows: ； in, Represents geomagnetic measurement data; The position encoding matrix is ​​in sine and cosine form: ； ； in, pos= 0, 1, ..., N -1 represents the position index, indicating the position number of the current data point within the time window sequence. The length of the time window; i= 0, 1, ..., The dimension index represents the first element in the feature vector. i One dimension, d Indicates the number of dimensions in the feature vector; The geomagnetic feature embedding is represented as follows: ； in, This is the geomagnetic feature mapping matrix; The gradient feature mapping matrix; and These are bias parameters; For the first The magnetic field gradient at each sampling point; The displacement distance of the aircraft is estimated based on the IMU inertial integration results at adjacent sampling times. To prevent constants with a denominator of zero; A geomagnetic sequence constructed using a fixed time window; Sampling time Geomagnetic observations, , , These represent the magnetic field components in the three directions of the carrier coordinate system, with superscript indicating the direction of the magnetic field. T This represents the transpose of a vector.

3. The marine magnetograph matching navigation method based on self-supervised deep learning and multimodal fusion according to claim 1, characterized in that, The inertial information of the trained IMU encoder input in step S2A is represented as follows: ； in, Represents inertial information; Indicates time location encoding; Indicates inertial feature embedding, The inertial feature mapping matrix, For bias parameters, For length is N Inertial sequence, For the carrier's motion state information, , , This represents the acceleration in three directions within the carrier's coordinate system. , , Represents the angular velocity in three directions in the carrier coordinate system, indicated by the superscript. T This represents the transpose of a vector.

4. The marine magnetograph matching navigation method based on self-supervised deep learning and multimodal fusion according to claim 1, characterized in that, The accumulated drift state of the IMU in step S2A is represented as follows: ； in, Indicates the first IMU cumulative drift metric at any given time; Represents the trace of a matrix; Represents the error state vector The covariance matrix; the error state vector , Indicates positional error. Indicates speed error, Indicates attitude error. This indicates the zero bias error of the accelerometer. Indicates the zero bias error of the gyroscope, indicated by the superscript. T This represents the transpose of a vector.

5. The marine magnetograph matching navigation method based on self-supervised deep learning and multimodal fusion according to claim 1, characterized in that, Step S2A employs a magnetic field gradient modulation cross-attention mechanism, utilizing magnetic field gradient information to dynamically modulate the cross-attention weights, including: Using geomagnetic features as query information and inertial features as keys and values, geomagnetic cross-attention guided by inertial information is calculated; Softmax normalization is performed on the cross-modal correlation matrix row by row to obtain the attention weight of inertial information to each geomagnetic moment; the inertial value features are weighted and summed according to the attention weight to obtain the geomagnetic features modulated by inertial motion information. Using inertial features as queries and geomagnetic features as keys and values, reverse cross-attention is calculated; Softmax normalization is performed on the cross-modal correlation matrix row by row to obtain the attention weights of geomagnetic information for each inertial time step; the geomagnetic value features are weighted and summed according to the attention weights to obtain the inertial features enhanced by geomagnetic information. Multi-head cross-attention computation is employed, with different attention heads learning the correlation features between changes in magnetic field amplitude, magnetic field gradient, attitude change, angular velocity change, and acceleration change. The cross-attention output is residually connected to the original modality features to obtain preliminary enhanced features; the preliminary enhanced features are then input into a trained feedforward neural network to output bidirectional cross-enhanced features.

6. The marine magnetograph matching navigation method based on self-supervised deep learning and multimodal fusion according to claim 1, characterized in that, In step S2B, the candidate trajectory search region is adaptively determined by combining the IMU cumulative drift evaluation results, as shown below: ； in, This is the search area for candidate trajectories; Indicates position coordinates; The geographical area covered by the magnetic map; Indicating the first coarse inertial trajectory k One location; The length of the time window; Let be the 2-norm of the matrix; The maximum spatial offset range of the candidate trajectory corresponding to the current matching time window is represented as: ； in, The preset base search radius; This is a search range adjustment factor; This is the time weighting coefficient; This is the cumulative drift metric for the IMU corresponding to the current matching time window.

7. The marine magnetograph matching navigation method based on self-supervised deep learning and multimodal fusion according to claim 1, characterized in that, It also includes a self-supervised training process based on magnetic maps, specifically including: In a pre-built ocean magnetic map A virtual starting position is randomly selected within the coverage area, and a virtual trajectory that satisfies motion continuity constraints is generated according to the preset sailing speed, heading, and range of motion changes. ； in, Indicates the virtual trajectory sequence number. Indicates the first The virtual trajectory in the first The location coordinates at each sampling time. express x Axis coordinates express y Axis coordinates, superscript T This represents vector transpose; by changing the virtual starting position, initial heading, sailing speed, and motion state change parameters, multiple different virtual sailing trajectories can be generated from the same ocean magnetic map; For the Virtual trajectory Based on the spatial coordinates of each virtual trajectory point in the ocean magnetic map, the magnetic map is interpolated and sampled to obtain the corresponding synthetic geomagnetic observation sequence. ; Speed ​​is calculated based on the relationship between adjacent positions and time in the virtual flight path. : ； in, Indicates the time difference between adjacent positions; Further calculation of motion acceleration : ； Calculate the heading angle based on the direction of the trajectory. : ； Angular velocity is calculated based on the change in heading at adjacent moments. : ； Based on the virtual trajectory and the preset vehicle motion state model, corresponding attitude, velocity, and acceleration states are generated. Combined with the gravity model, IMU measurement model, and preset sensor error model, a synthetic IMU inertial measurement sequence consisting of three-axis acceleration and three-axis angular velocity is constructed. ; For virtual trajectory The magnetic map regions covering the virtual trajectory and its neighborhood are extracted from the pre-built ocean magnetic map as positive sample map tiles. ; Will and Constructing positive sample pairs ; Select from the magnetic map area corresponding to the virtual trajectory in other spatial locations. As negative sample map tiles, thus forming negative sample pairs. , The number of negative samples corresponding to each positive sample; Synthetic geomagnetic observation sequence and synthetic IMU inertial measurement sequence Following step S2A, geomagnetic and inertial features are extracted and fused to obtain a normalized query embedding. ; to the positive sample magnetic map area and negative sample magnetic map regions The corresponding normalized map embeddings are obtained by inputting them into the ConvNeXt network. ; Using contrastive learning loss function : ； ； in, The number of samples in a training batch. For temperature coefficient, For positive sample similarity, For the first Similarity of negative samples; By minimizing ,make at the same time ; During training, the parameters of each model in the method are jointly optimized through backpropagation based on the contrastive learning loss function value.

8. A marine magnetograph matching navigation system based on self-supervised deep learning and multimodal fusion, used to implement the method as described in any one of claims 1-7, characterized in that, include: The information acquisition module is used to acquire real-time navigation data, including geomagnetic measurement data and inertial information; The query embedding module is used to input geomagnetic measurement data into a trained magnetic encoder and output high-dimensional geomagnetic features. The inertial information is input into the trained IMU encoder, and the output is high-dimensional inertial features; A magnetic field gradient modulation cross-attention mechanism is adopted, which dynamically modulates the cross-attention weights using magnetic field gradient information. While establishing the correlation between geomagnetic information and inertial information, physical constraints of the spatial variation characteristics of the magnetic field are introduced. After completing the cross-attention fusion, a drift-sensing gating fusion method is adopted, which adaptively adjusts the contribution ratio of geomagnetic features and inertial features in the fusion process by combining the accumulated drift state of the IMU. Subsequently, attention pooling is performed on the fused sequence features to generate a query embedding that can comprehensively reflect geomagnetic spatial information and the motion state of the carrier. The map embedding module is used to perform inertial navigation integration based on the known state at the start of the current matching time window and the continuously collected inertial measurement data within the time window. It continuously recursively extrapolates the attitude, velocity, and position of the vehicle to obtain the inertial navigation position estimate at each sampling moment within the current matching time window, forming a coarse inertial trajectory of the vehicle. After obtaining the coarse inertial trajectory, the candidate trajectory search area is adaptively determined by combining the IMU cumulative drift evaluation results. Within the candidate trajectory search area, multiple candidate trajectories are generated discretely using the coarse inertial trajectory as the motion prior, satisfying known initial state constraints, inertial motion trend constraints, and trajectory continuity constraints, forming a candidate trajectory set. For each generated candidate trajectory, based on its spatial location in the ocean magnetomap, a local magnetomap region covering the candidate trajectory and its neighborhood is extracted from the pre-constructed ocean magnetomap, establishing a one-to-one correspondence between the candidate trajectory and the local magnetomap region. The local magnetomaps corresponding to each candidate trajectory are input into the trained ConvNeXt network for deep spatial feature extraction, obtaining the depth feature map of the local magnetomap. The spatial features of the depth feature map are aggregated through global average pooling, and through linear mapping and normalization, a map embedding in the same feature space as the query embedding is obtained. and The estimated navigation position module is used to calculate the similarity between the query embedding corresponding to the real-time navigation data within the current matching time window and the map embedding corresponding to each candidate trajectory, so as to obtain the deep feature similarity between the real-time geomagnetic-inertial fusion features and the magnetic field spatial distribution features corresponding to each candidate trajectory. The candidate trajectory with the highest similarity is selected as the optimal matching trajectory, and the spatial position of the optimal matching trajectory at the end of the current matching time window is taken as the current magnetograph matching navigation position estimation result of the vehicle.

Images