A gravity gradient matching positioning method based on multi-dimensional matchnet

Abstract

The application provides a gravity gradient matching positioning method based on a multi-dimensional MatchNet, uses the feature extraction and measurement capability of the MatchNet, combines five independent components of the gravity gradient, and builds a gravity gradient image matching network GGIMN, the matching network GGIMN fully utilizes the advantages of deep learning in image matching, has more accurate and efficient representation and measurement capability than traditional mathematical statistics features, gray features and other image features, not only improves the matching positioning accuracy, but also solves the problems of complex calculation and high requirement for image feature change in the previous matching positioning method.

Description

Technical Field

[0001] This invention belongs to the field of integrated navigation technology, and particularly relates to a gravity gradient matching localization method based on multidimensional MatchNet. Background Technology

[0002] Gravity gradient-assisted navigation is an important research direction in underwater gravity-assisted navigation. It is a combined navigation system consisting of gravity gradient information and an inertial navigation system (INS). Gravity gradient information helps correct accumulated errors in the INS, achieving high-precision navigation and positioning of underwater objects. The matching method is the core technology of gravity gradient-assisted navigation systems, and its performance directly affects the accuracy and efficiency of underwater navigation and positioning.

[0003] Deep learning has achieved groundbreaking results in tasks such as image feature extraction and measurement due to its ability to learn feature representations and patterns from raw image data by constructing deep neural network models. Recent research on gravity gradient-assisted navigation matching methods proposes using image matching to achieve gravity gradient matching. This involves fitting five independent real-time gravity gradient maps with measured gravity gradient component information and estimating the optimal matching position using the principle of image feature matching against a pre-constructed gravity gradient reference map. However, feature matching of gravity gradient reference maps with uneven feature distribution struggles to guarantee high-precision matching across the entire range. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a gravity gradient matching localization method based on multidimensional MatchNet. This method leverages the self-learning capability of deep learning for image features to achieve global matching for gravity gradient-assisted navigation, thereby improving the accuracy of the matching method. By combining single-point matching with sequence matching, the advantages of both methods are complemented, making it more intelligent and accurate than existing matching methods.

[0005] A gravity gradient matching localization method based on multidimensional MatchNet includes the following steps:

[0006] S1: Obtain the current position of the underwater vehicle from the INS and the current gravity gradient data of the underwater vehicle from the gravity gradiometer. S2: Obtain the real-time gravity gradient map and the baseline gravity gradient map corresponding to the five independent gravity gradient components in the gravity gradient data respectively; S3: Input the current indicated position, the confidence interval corresponding to the indicated position, the five real-time gravity gradient maps corresponding to the five independent gravity gradient components, and the five gravity gradient baseline maps into the trained GGIMN network to obtain the predicted position of the underwater vehicle at the current time. S4: Determine whether the curvature of the INS indication trajectory formed by the current indication position and the previous indication position in the INS output is greater than a set threshold. If yes, then use an adaptive adjoint smoothing method to smooth the prediction trajectory formed by the current prediction position and the previous prediction position. If no, then use the least squares method to smooth the prediction trajectory. S5: Use the position of the smoothed predicted trajectory at the current moment as the final estimated position of the underwater vehicle.

[0007] Furthermore, the method for obtaining the GGIMN network is as follows: S31: INS locates the underwater vehicle in real time as it traverses the underwater environment along a set path, and obtains the real-time indicated position of the underwater vehicle; at the same time, the real-time actual position of the underwater vehicle is a known quantity; S32: The gravity gradiometer acquires gravity gradient data of the underwater vehicle in real time as the vehicle traverses the underwater environment along a set path. At the same time, it extracts the real-time gravity gradient map corresponding to the five independent gravity gradient components in the gravity gradient data at each moment, as well as the gravity gradient reference map corresponding to the same gradient component at different moments. S33: Perform a sample acquisition operation on each gravity gradient component at each time moment as the current gravity gradient component to obtain the positive and negative samples corresponding to each gravity gradient component at each time moment. The sample acquisition operation is as follows: S33a: Using the position of the current indicated position on the current gravity gradient reference map corresponding to the current gravity gradient component as the center and the confidence interval corresponding to the current indicated position as the radius, a search space is defined on the current gravity gradient reference map. S33b: The current real-time gravity gradient map traverses the search space according to a set step size until the center position of the current real-time gravity gradient map coincides with the actual position of the underwater body on the current gravity gradient reference map. The reference map that overlaps with the current gravity gradient reference map and the current real-time gravity gradient map together constitute a positive sample. At the same time, a patch reference map of the same size as the current real-time gravity gradient map is randomly selected in the search space to form a negative sample together with the current real-time gravity gradient map. S34: The initial GGIMN network is trained using five gravity gradient reference maps corresponding to five independent gravity gradient components, five independent gravity gradient components at each time step, the indicated position at each time step, the confidence interval corresponding to the indicated position at each time step, and an equal number of positive and negative samples as inputs to the GGIMN network. The labels corresponding to each positive and negative sample are used as the theoretical outputs of the GGIMN network. The label corresponding to the positive sample is 1, indicating that the true position of the underwater object has been obtained, and the label of the negative sample is 0, indicating that the true position of the underwater object has not been obtained. S35: Determine whether the cross-entropy loss function used by the GGIMN network is less than the set threshold. If yes, the current GGIMN network is the GGIMN network obtained by the final training. If no, reset the path for the underwater vehicle to traverse the underwater environment, and then repeat steps S31 to S35 until the cross-entropy loss function is less than the set threshold.

[0008] Furthermore, the GGIMN network employs the cross-entropy loss function. as follows:

[0009] in, For the water-based body in the first i The actual location at any given moment The water download body output by the GGIMN network in the first i Predicted location at each moment, This represents the total number of time points corresponding to the sample data used when training the GGIMN network.

[0010] Furthermore, the method for obtaining the curvature of the INS indication trajectory in step S4 is as follows: Record the current position as ( x 3, y 3) The indicated positions at the previous two moments are respectively denoted as ( x 1, y 1) ( x 2, y 2); Obtaining INS tracking data x Axis coordinates and y Axis coordinates The expression is as follows:

[0011] in, For trajectory parameters, a 1. a 2. a 3. b 1. b 2. b All three are undetermined coefficients; Obtain the vector length between three time points and as follows:

[0012] INS-indicated trajectory x Axis coordinates and y Axis coordinates The expression is converted into a matrix solution of the linear equation system as follows:

[0013] Indicate location ( x 1, y 1) ( x 2, y 2), ( x 3, y 3) Substitute the matrix into the system of linear equations to obtain... a 1. a 2. a 3. b 1. b 2. b The possible values ​​of 3; according to a 1. a 2. a 3. b 1. b 2. b 3. The curvature of the INS indicator track is calculated as follows:

[0014] in, The curvature of the INS indicator trajectory.

[0015] Furthermore, the method for smoothing the predicted trajectory using the adaptive adjoint smoothing method in step S4 is as follows: Get the indication position of the previous L-1 time steps. and ;in, and These are the longitude and latitude at the current moment, respectively. These are the longitudes of the L-1 time points preceding the current time. These are the dimensions of the L-1 time steps preceding the current time step; The difference in indicated position between adjacent time points is obtained as follows:

[0016] in, ~ These represent the longitude differences between each adjacent time point in the previous L time points, including the current time point. ~ These are the latitude differences between each adjacent time point in the previous L time points, including the current time point; The predicted trajectory is smoothed based on the indicated position difference as follows:

[0017] in, ~ These are the longitude values ​​of the smoothed predicted trajectory for the previous L time points, including the current time. ~ These are the latitude values ​​of the smoothed predicted trajectory for the previous L time steps, including the current time step. and The predicted trajectories before smoothing are respectively the first... Longitude and latitude values ​​at any given time; Will ~ and ~ The resulting trajectory serves as the smoothed predicted trajectory.

[0018] Furthermore, the method for smoothing the predicted trajectory using the least squares method in step S4 is as follows: The second-order polynomial representing the predicted trajectory is constructed as follows:

[0019] in, , and All are undetermined coefficients; The longitude values ​​of the predicted trajectory before smoothing, including the current time, for the previous L time points. ~ The latitude values ​​of the predicted trajectory before smoothing, including the current time step, for the previous L time steps. ~ Solve for the undetermined coefficients , and ; Will , and The known second-order polynomial is used as the smoothed predicted trajectory.

[0020] Furthermore, the gravity gradiometer is a rotating accelerometer gravity gradiometer composed of three mutually perpendicular gravity gradiometers, and the five independent gravity gradient components are respectively... , , , , ,in, For located x Gravity gradient data measured by a gravity gradiometer on the axis x Components on the axis, For located x Gravity gradient data measured by a gravity gradiometer on the axisy Components on the axis, For located x Gravity gradient data measured by a gravity gradiometer on the axis z Components on the axis, For located y Gravity gradient data measured by a gravity gradiometer on the axis y Components on the axis, For located y Gravity gradient data measured by a gravity gradiometer on the axis z Components on the axis.

[0021] Beneficial effects: 1. This invention provides a gravity gradient matching localization method based on multidimensional MatchNet. By utilizing MatchNet's feature extraction and measurement capabilities and combining the five independent components of the gravity gradient, a gravity gradient image matching network GGIMN is constructed. This matching network GGIMN fully leverages the advantages of deep learning in image matching, possessing more accurate and efficient representation and measurement capabilities than traditional mathematical statistical features, grayscale features, and other image features. It not only improves the accuracy of matching localization but also solves the problems of computational complexity and high requirements for image feature changes in previous matching localization methods.

[0022] 2. This invention provides a gravity gradient matching localization method based on multidimensional MatchNet, and also proposes an adaptive adjoint smoothing method for the predicted trajectory output by the GGIMN network. Specifically, this invention proposes an adaptive adjoint smoothing method that combines single-point matching and sequence matching to further smooth the matching results, thereby improving the matching and localization accuracy and realizing high-precision gravity gradient-assisted navigation and localization of underwater vehicles.

[0023] 3. This invention provides a gravity gradient matching localization method based on multidimensional MatchNet. It constructs an equal number of positive and negative samples from a probabilistic perspective to ensure balanced training samples, and then trains the gravity gradient image matching network to improve the matching localization accuracy. Attached Figure Description

[0024] Figure 1 A flowchart of a gravity gradient matching localization method based on multidimensional MatchNet provided by the present invention; Figure 2 A schematic diagram of the gravity gradient components of the rotating accelerometer gravity gradiometer provided by the present invention; Figure 3 The GGIMN framework for gravity gradient image matching network provided by this invention; Figure 4 The diagram illustrates the principle of the adaptive adjoint smoothing method provided by this invention. Detailed Implementation

[0025] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0026] This invention proposes a gravity gradient image matching network (GGIMN), constructing a multi-dimensional MatchNet deep learning gravity gradient matching model. The network includes image preprocessing, feature extraction, and feature measurement, and the highest similarity matching position is obtained through the trained network. Furthermore, this invention proposes an adaptive adjoint smoothing method, employing a combination of single-point matching and sequence matching to further improve the localization accuracy of GGIMN. This invention fully utilizes the self-learning capability of the MatchNet deep learning network for image matching, combining multi-dimensional gravity gradient features to improve the instability of accuracy in gravity gradient-based image matching methods.

[0027] Specifically, such as Figure 1 As shown, a gravity gradient matching localization method based on multidimensional MatchNet includes the following steps: S1: Obtain the current position of the underwater vehicle from the INS and the current gravity gradient data of the underwater vehicle from the gravity gradiometer. S2: Obtain the real-time gravity gradient map and the baseline gravity gradient map corresponding to the five independent gravity gradient components in the gravity gradient data respectively; For example, such as Figure 2 As shown, the gravity gradiometer is a rotating accelerometer gravity gradiometer composed of three mutually perpendicular gravity gradiometers, and the nine gravity gradient components are:

[0028] Among them, the five independent gravity gradient components are as follows: , , , , ,in, For located x Gravity gradient data measured by a gravity gradiometer on the axis x Components on the axis, For located x Gravity gradient data measured by a gravity gradiometer on the axis y Components on the axis, For located x Gravity gradient data measured by a gravity gradiometer on the axis zComponents on the axis, For located y Gravity gradient data measured by a gravity gradiometer on the axis y Components on the axis, For located y Gravity gradient data measured by a gravity gradiometer on the axis z Components on the axis.

[0029] S3: Input the current indicated position, the confidence interval corresponding to the indicated position, the five real-time gravity gradient maps corresponding to the five independent gravity gradient components, and the five gravity gradient baseline maps into the trained GGIMN network to obtain the predicted position of the underwater vehicle at the current time. It should be noted that when acquiring the GGIMN network, firstly, the network framework is determined. The GGIMN network consists of input samples, a feature network, a metric network, and a loss function. Secondly, the GGIMN network parameters are defined. It is composed of a five-dimensional MatchNet network, including five convolutional layers, three pooling layers, and preprocessing layers and unfolding layers at both ends, using the cross-entropy loss function. Thirdly, positive and negative samples are prepared to construct a gravity gradient map matching dataset. Finally, equal amounts of positive and negative samples are input to train GGIMN.

[0030] Specifically, the method for obtaining the GGIMN network is as follows: S31: INS locates the underwater vehicle in real time as it traverses the underwater environment along a set path, and obtains the real-time indicated position of the underwater vehicle; at the same time, the real-time actual position of the underwater vehicle is a known quantity; S32: The gravity gradiometer acquires gravity gradient data of the underwater vehicle in real time as the vehicle traverses the underwater environment along a set path. At the same time, it extracts the real-time gravity gradient map corresponding to the five independent gravity gradient components in the gravity gradient data at each moment, as well as the gravity gradient reference map corresponding to the same gradient component at different moments. S33: Perform a sample acquisition operation on each gravity gradient component at each time moment as the current gravity gradient component to obtain the positive and negative samples corresponding to each gravity gradient component at each time moment. The sample acquisition operation is as follows: S33a: Using the position of the current indicated position on the current gravity gradient reference map corresponding to the current gravity gradient component as the center and the confidence interval corresponding to the current indicated position as the radius, a search space is defined on the current gravity gradient reference map. S33b: The current real-time gravity gradient map traverses the search space according to a set step size until the center position of the current real-time gravity gradient map coincides with the actual position of the underwater body on the current gravity gradient reference map. The reference map that overlaps with the current gravity gradient reference map and the current real-time gravity gradient map together constitute a positive sample. At the same time, a patch reference map of the same size as the current real-time gravity gradient map is randomly selected in the search space to form a negative sample together with the current real-time gravity gradient map. Therefore, the positive and negative sample data of this invention are constructed by strictly segmenting and encoding the gravity gradient map in a sequential manner. The number of matching images within the INS confidence interval (i.e., the center position of the current real-time gravity gradient map coincides with the actual position of the underwater body on the current gravity gradient reference map) is 1, labeled 1, and is used as a positive sample; the rest are non-matching images. Therefore, the positive and negative samples are unbalanced. A local image of the gravity gradient reference image is randomly selected in each search space as a patch. It is probabilistically non-matching with the real-time gravity gradient map and is used as a negative sample with a label of 0 to ensure the balance of training samples.

[0031] S34: The initial GGIMN network is trained using five gravity gradient reference maps corresponding to five independent gravity gradient components, five independent gravity gradient components at each time step, the indicated position at each time step, the confidence interval corresponding to the indicated position at each time step, and an equal number of positive and negative samples as inputs to the GGIMN network. The labels corresponding to each positive and negative sample are used as the theoretical outputs of the GGIMN network. The label corresponding to the positive sample is 1, indicating that the true position of the underwater object has been obtained, and the label of the negative sample is 0, indicating that the true position of the underwater object has not been obtained. S35: Determine whether the cross-entropy loss function used by the GGIMN network is less than the set threshold. If yes, the current GGIMN network is the GGIMN network obtained by the final training. If no, reset the path for the underwater vehicle to traverse the underwater environment, and then repeat steps S31 to S35 until the cross-entropy loss function is less than the set threshold.

[0032] Specifically, the GGIMN network uses the cross-entropy loss function. as follows:

[0033] in, For the water-based body in the first i The actual location at any given moment The water download body output by the GGIMN network in the first i Predicted location at each moment, This represents the total number of time points corresponding to the sample data used when training the GGIMN network.

[0034] Therefore, unlike the MatchNet network, the feature network of the gravity gradient image matching network GGIMN of this invention has a more complex five-dimensional dual-tower structure than the MatchNet dual-tower structure, consisting of five independent components of the gravity gradient, such as... Figure 3 As shown, pink, green, purple, blue, and yellow represent these five components respectively. , , , , .

[0035] Meanwhile, the parameters of each layer of the constructed GGIMN network are shown in Table 1.

[0036] Table 1

[0037] S4: Determine whether the curvature of the INS indication trajectory formed by the current indication position and the previous indication position is greater than a set threshold. If yes, it means that the trajectory at this time is the initial matching trajectory or the trajectory at the turning point. Then, the adaptive accompanying smoothing method is used to smooth the predicted trajectory formed by the current prediction position and the previous prediction position. If no, it means that the trajectory at this time is the trajectory of a flat segment. Then, the least squares method is used to smooth the predicted trajectory. Furthermore, the method for obtaining the curvature of the INS indication trajectory is as follows: Record the current position as ( x 3, y 3) The indicated positions at the previous two moments are respectively denoted as ( x 1, y 1) ( x 2, y 2); Obtaining INS tracking data x Axis coordinates and y Axis coordinates The expression is as follows:

[0038] in, For trajectory parameters, a 1. a 2. a 3. b 1. b 2. b All three are undetermined coefficients; Obtain the vector length between three time points and as follows:

[0039] Right now

[0040] but

[0041] The above formula can be written in the following matrix form.

[0042] Based on this, the INS indicates the trajectory x Axis coordinates and y Axis coordinates The expression is converted into a matrix solution of the linear equation system as follows:

[0043] Indicate location ( x 1, y 1) ( x 2, y 2), ( x 3, y 3) Substitute the matrix into the system of linear equations to obtain... a 1. a 2. a 3. b 1. b 2. b The possible values ​​of 3; The formula for calculating curvature is as follows:

[0044] In the formula r It is a curve.

[0045] If the curve is represented by a parametric equation, then

[0046] Track indicated by INS x Axis coordinates and y Axis coordinates The expression can be obtained

[0047] Then according to a 1. a 2. a 3. b 1. b 2. b 3. The curvature of the INS indicator track is calculated as follows:

[0048] in, The curvature of the INS indicator trajectory.

[0049] It should be noted that, A smaller value indicates a smoother INS trajectory. Since the turning changes of INS are relatively gentle, this value is only considered when the absolute value of the curvature is greater than the curvature threshold. T k (0.5) indicates a turn.

[0050] Firstly, the method for smoothing the predicted trajectory using an adaptive adjoint smoothing method is as follows: Get the indication position of the previous L-1 time steps. and ;in, and These are the longitude and latitude at the current moment, respectively. These are the longitudes of the L-1 time points preceding the current time. These are the dimensions of the L-1 time steps preceding the current time step; The difference in indicated position between adjacent time points is obtained as follows:

[0051] in, ~ These represent the longitude differences between each adjacent time point in the previous L time points, including the current time point. ~ These are the latitude differences between each adjacent time point in the previous L time points, including the current time point; The predicted trajectory is smoothed based on the indicated position difference as follows:

[0052] in, ~ These are the longitude values ​​of the smoothed predicted trajectory for the previous L time points, including the current time. ~ These are the latitude values ​​of the smoothed predicted trajectory for the previous L time steps, including the current time step. and The predicted trajectories before smoothing are respectively the first... Longitude and latitude values ​​at any given time; Will ~ and ~ The resulting trajectory serves as the smoothed predicted trajectory.

[0053] It should be noted that the adaptive adjoint smoothing method proposed in this invention refers to further estimating the gravity gradient map matching result following the positional changes of INS longitude and latitude. Furthermore, the adaptive adjoint smoothing method of this invention smooths the predicted trajectory using the previous L time points, including the current time point. This is equivalent to setting the width of the filtering window to L, i.e., smoothing the positioning result using a sequence sliding window of length L. Figure 4 As shown.

[0054] Secondly, the method for smoothing the predicted trajectory using the least squares method is as follows: The second-order polynomial representing the predicted trajectory is constructed as follows:

[0055] in, , and All are undetermined coefficients; The longitude values ​​of the predicted trajectory before smoothing, including the current time, for the previous L time points. ~ The latitude values ​​of the predicted trajectory before smoothing, including the current time step, for the previous L time steps. ~ Solve for the undetermined coefficients , and ; Will , and The known second-order polynomial is used as the smoothed predicted trajectory.

[0056] Furthermore, the estimation effect of the underwater vehicle position can be experimentally verified using the above method. The results show that the proposed GGIMN network meets the accuracy requirements for underwater vehicle navigation and positioning.

[0057] S5: Use the position of the smoothed predicted trajectory at the current moment as the final estimated position of the underwater vehicle.

[0058] Therefore, this invention utilizes the feature extraction and measurement capabilities of MatchNet, combined with the five independent components of the gravity gradient, to construct a gravity gradient image matching network (GGIMN); it constructs an equal number of positive and negative samples from a probabilistic perspective to train the gravity gradient image matching network; furthermore, this invention also proposes an adaptive adjoint smoothing method to smooth the GGIMN matching results, thereby achieving high-precision gravity gradient-assisted navigation and positioning of underwater vehicles.

[0059] The present invention will be further described below with reference to specific embodiments.

[0060] Step 1: Constructing as follows Figure 3 The gravity gradient image matching network GGIMN shown uses the Keras deep learning framework.

[0061] Step 2: Define the GGIMN network parameters, as shown in Table 1.

[0062] Step 3: Construct a gravity gradient image matching dataset.

[0063] For example, in this embodiment, a baseline map of a sea area in the southern part of a certain sea area with a latitude and longitude span of 1° is selected for simulation experiments. The resolution is 1′×1′, the sampling frequency of gravity anomalies is 1 Hz, and the gravity gradient baseline map is gridded to a size of 600×600 grids, that is, the resolution is interpolated to 6″×6″ (1″≈30.8 meters). The size of the real-time gravity gradient map is set to 32×32 grids, with a step size of... d =4 grids were used to divide the gravity gradient baseline map, resulting in an equal number (3844) of positive and negative samples.

[0064] Step 4: Input equal amounts of positive and negative samples to train GGIMN.

[0065] Step 5: Use the trained GGIMN network to estimate the matching position.

[0066] For example, in this embodiment, a straight line trajectory along the northeast direction is set, the number of sampling points is 100, and the length of the smoothing sequence is [missing information]. L =8, the confidence radius of INS is 40, and the confidence interval is a rectangular area with a radius of 40 centered on the INS indication position.

[0067] Step 6: Calculate the curvature of the INS indicator trajectory to determine the underwater vehicle's trajectory.

[0068] For example, in this embodiment, the running trajectory is divided into two types based on the curvature: the initial segment and the normal trajectory segment. The initial segment and the normal trajectory segment are smoothed using the adaptive adjoint smoothing method, which utilizes the INS indicator trajectory and the least squares method, respectively.

[0069] Step 7: Smooth the matching trajectory output in Step 5 using the adaptive adjoint smoothing method.

[0070] For example, in this embodiment, the initial segment and the ordinary trajectory segment are smoothed using the INS indicator trajectory and the least squares method, respectively, according to the adaptive adjoint smoothing method.

[0071] Step 8: Output the position estimation results of the underwater body.

[0072] For example, in this embodiment, the minimum error between the INS and the actual trajectory is 3 grids, and the maximum is 13 grids. GGIMN demonstrates superior performance in learning and measuring the features of the gravity gradient reference map, with a minimum positioning error of 0 grids and a maximum of 10 grids. Since the GGIMN matching result for each sampling point is independent, although the overall accuracy of the GGIMN matching result is high, there are still a few individual points where the positioning error reaches 10 grids. The maximum, minimum, and average values ​​of the INS indicated track position error are 13.150 grids, 3.162 grids, and 7.546 grids, respectively; the maximum, minimum, and average values ​​of the GGIMN matching position error are 10 grids, 0 grids, and 1.816 grids, respectively; and the maximum, minimum, and average values ​​of the GGIMN matching position error with adaptive accompanying smoothing are 3.606 grids, 0 grids, and 1.335 grids, respectively. GGIMN achieves an average error correction of 75.9% for INS, and GGIMN with adaptive accompanying smoothing achieves an average error correction of 82.3% for INS.

[0073] Compared with existing technologies, the beneficial effects of this invention are as follows: The gravity gradient-assisted navigation matching network GGIMN based on multidimensional MatchNet fully utilizes the advantages of deep learning in image matching, possessing more accurate and efficient representation and measurement capabilities than traditional mathematical statistical features, grayscale features, and other image features, thus solving the problems of computational complexity and high requirements for image feature changes in previous matching methods. Simultaneously, an adaptive adjoint smoothing method combining single-point matching and sequence matching is proposed to further smooth the matching results, thereby improving matching and positioning accuracy.

[0074] Simulation results show that the present invention has high positioning accuracy and can perform matching across the entire domain.

[0075] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A gravity gradient matching localization method based on multidimensional MatchNet, characterized in that, Includes the following steps: S1: Obtain the current position of the underwater vehicle from the INS and the current gravity gradient data of the underwater vehicle from the gravity gradiometer. S2: Obtain the real-time gravity gradient map and the baseline gravity gradient map corresponding to the five independent gravity gradient components in the gravity gradient data respectively; S3: Input the current indicated position, the confidence interval corresponding to the indicated position, the five real-time gravity gradient maps corresponding to the five independent gravity gradient components, and the five gravity gradient baseline maps into the trained GGIMN network to obtain the predicted position of the underwater vehicle at the current time. S4: Determine whether the curvature of the INS indication trajectory formed by the current indication position and the previous indication position in the INS output is greater than a set threshold. If yes, then use an adaptive adjoint smoothing method to smooth the prediction trajectory formed by the current prediction position and the previous prediction position. If no, then use the least squares method to smooth the prediction trajectory. S5: Use the position of the smoothed predicted trajectory at the current moment as the final estimated position of the underwater vehicle; The method for obtaining the GGIMN network is as follows: S31: INS locates the underwater vehicle in real time as it traverses the underwater environment along a set path, and obtains the real-time indicated position of the underwater vehicle; at the same time, the real-time actual position of the underwater vehicle is a known quantity; S32: The gravity gradiometer acquires gravity gradient data of the underwater vehicle in real time as the vehicle traverses the underwater environment along a set path. At the same time, it extracts the real-time gravity gradient map corresponding to the five independent gravity gradient components in the gravity gradient data at each moment, as well as the gravity gradient reference map corresponding to the same gradient component at different moments. S33: Perform a sample acquisition operation on each gravity gradient component at each time moment as the current gravity gradient component to obtain the positive and negative samples corresponding to each gravity gradient component at each time moment. The sample acquisition operation is as follows: S33a: Using the position of the current indicated position on the current gravity gradient reference map corresponding to the current gravity gradient component as the center and the confidence interval corresponding to the current indicated position as the radius, a search space is defined on the current gravity gradient reference map. S33b: The current real-time gravity gradient map traverses the search space according to a set step size until the center position of the current real-time gravity gradient map coincides with the actual position of the underwater body on the current gravity gradient reference map. The reference map that overlaps with the current gravity gradient reference map and the current real-time gravity gradient map together constitute a positive sample. At the same time, a patch reference map of the same size as the current real-time gravity gradient map is randomly selected in the search space to form a negative sample together with the current real-time gravity gradient map. S34: The initial GGIMN network is trained using five gravity gradient reference maps corresponding to five independent gravity gradient components, five independent gravity gradient components at each time step, the indicated position at each time step, the confidence interval corresponding to the indicated position at each time step, and an equal number of positive and negative samples as inputs to the GGIMN network. The labels corresponding to each positive and negative sample are used as the theoretical outputs of the GGIMN network. The label corresponding to the positive sample is 1, indicating that the true position of the underwater object has been obtained, and the label of the negative sample is 0, indicating that the true position of the underwater object has not been obtained. S35: Determine whether the cross-entropy loss function used by the GGIMN network is less than the set threshold. If yes, the current GGIMN network is the GGIMN network obtained by the final training. If no, reset the path for the underwater vehicle to traverse the underwater environment, and then repeat steps S31 to S35 until the cross-entropy loss function is less than the set threshold.

2. The gravity gradient matching localization method based on multidimensional MatchNet as described in claim 1, characterized in that, The cross-entropy loss function used in the GGIMN network as follows: in, For the water-based body in the first i The actual location at any given moment The water download body output by the GGIMN network in the first i Predicted location at each moment, This represents the total number of time points corresponding to the sample data used when training the GGIMN network.

3. The gravity gradient matching localization method based on multidimensional MatchNet as described in claim 1, characterized in that, The method for obtaining the curvature of the INS indicator trajectory in step S4 is as follows: Record the current position as ( x 3, y 3) The indicated positions at the previous two moments are respectively denoted as ( x 1, y 1) ( x 2, y 2); Obtaining INS tracking data x Axis coordinates and y Axis coordinates The expression is as follows: in, For trajectory parameters, a 1. a 2. a 3. b 1. b 2. b All three are undetermined coefficients; Obtain the vector length between three time points and as follows: INS-indicated trajectory x Axis coordinates and y Axis coordinates The expression is converted into a matrix solution of the linear equation system as follows: Indicate location ( x 1, y 1) ( x 2, y 2), ( x 3, y 3) Substitute the matrix into the system of linear equations to obtain... a 1. a 2. a 3. b 1. b 2. b The possible values ​​of 3; according to a 1. a 2. a 3. b 1. b 2. b 3. The curvature of the INS indicator track is calculated as follows: in, The curvature of the INS indicator trajectory.

4. The gravity gradient matching localization method based on multidimensional MatchNet as described in claim 1, characterized in that, The method for smoothing the predicted trajectory using the adaptive adjoint smoothing method in step S4 is as follows: Get the indication position of the previous L-1 time steps. and ;in, and These are the longitude and latitude at the current moment, respectively. These are the longitudes of the L-1 time points preceding the current time. These are the dimensions of the L-1 time steps preceding the current time step; The difference in indicated position between adjacent time points is obtained as follows: in, ~ These represent the longitude differences between each adjacent time point in the previous L time points, including the current time point. ~ These are the latitude differences between each adjacent time point in the previous L time points, including the current time point; The predicted trajectory is smoothed based on the indicated position difference as follows: in, ~ These are the longitude values ​​of the smoothed predicted trajectory for the previous L time points, including the current time. ~ These are the latitude values ​​of the smoothed predicted trajectory for the previous L time steps, including the current time step. and The predicted trajectories before smoothing are respectively the first... Longitude and latitude values ​​at any given time; Will ~ and ~ The resulting trajectory serves as the smoothed predicted trajectory.

5. The gravity gradient matching localization method based on multidimensional MatchNet as described in claim 1, characterized in that, The method for smoothing the predicted trajectory using the least squares method in step S4 is as follows: The second-order polynomial representing the predicted trajectory is constructed as follows: in, , and All are undetermined coefficients; The longitude values ​​of the predicted trajectory before smoothing, including the current time, for the previous L time points. ~ The latitude values ​​of the predicted trajectory before smoothing, including the current time step, for the previous L time steps. ~ Solve for the undetermined coefficients , and ; Will , and The known second-order polynomial is used as the smoothed predicted trajectory.

6. The gravity gradient matching localization method based on multidimensional MatchNet as described in claim 1, characterized in that, The gravity gradiometer is a rotating accelerometer gravity gradiometer composed of three mutually perpendicular gravity gradiometers, and the five independent gravity gradient components are as follows: , , , , ,in, For located x Gravity gradient data measured by a gravity gradiometer on the axis x Components on the axis, For located x Gravity gradient data measured by a gravity gradiometer on the axis y Components on the axis, For located x Gravity gradient data measured by a gravity gradiometer on the axis z Components on the axis, For located y Gravity gradient data measured by a gravity gradiometer on the axis y Components on the axis, For located y Gravity gradient data measured by a gravity gradiometer on the axis z Components on the axis.

Images