Underwater navigation adaptation area connectivity improving method based on correction domain boundary screening

Abstract

The invention relates to an underwater navigation adaptation area connectivity improving method based on correction domain boundary screening, which comprises the following steps: constructing a gravity anomaly feature and adaptation area correlation model EBMAS, and completing underwater navigation adaptation area calibration by using an EBMAS model, firstly, making up an evaluation blank of a navigation reserved boundary of a TERCOM algorithm by using a correction square domain migration mechanism and an out-of-domain sequence reduction mechanism; an adaptation area is calibrated; secondly, obtaining a training data set in combination with adaptive region class labels and gravity anomaly features; then, a range control synthetic minority class oversampling technology RCSMOTE is combined to control the position of a newly synthesized adaptive region sample in a data space so as to identify a potential adaptive region sample and enhance the classification identification performance of the calibration model; afterwards, a key gravity anomaly feature set with comprehensive information representation is screened in combination with a maximum correlation minimum redundancy algorithm mRMR, and compared with the prior art, the comprehensiveness and accuracy of calibration of the adaptation area can be effectively improved, and high applicability is still shown when the model is migrated to a new area.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of underwater navigation and ocean surveying, and particularly relates to an underwater navigation adaptation area connectivity improvement method based on a correction domain boundary screening method, which can effectively improve the calibration accuracy of an adaptation area. BACKGROUND

[0002] An inertial navigation system (INS) is a core navigation system capable of realizing short-time high-precision positioning and is widely applied in the field of autonomous navigation of underwater vehicles. However, due to the inherent errors of inertial components and the integral accumulation errors in position calculation, the navigation precision and reliability are reduced. Therefore, the INS needs to be periodically adjusted and calibrated to meet the long-time and high-precision positioning requirements of underwater vehicles. As one of the inherent properties of the earth, the gravity field information is stable for a long time and is not easily affected by the marine environment, and thus becomes a relatively reliable auxiliary information source in underwater vehicle navigation. At present, gravity-based auxiliary navigation has always been a hot research topic in underwater auxiliary INS navigation.

[0003] As an important component of the gravity aided navigation system, the gravity matching algorithm matches the measured gravity anomaly sequence with the gravity anomaly reference region to realize accurate positioning during navigation. The TERCOM algorithm is a high-performance gravity matching algorithm and is widely applied due to its advantages of being insensitive to initial errors, high matching precision and strong robustness. However, when the gravity field in the matching domain has similar gravity contour changes or some regions show gravity flat areas, the positioning precision of the TERCOM algorithm will be greatly affected. Therefore, identifying key gravity anomaly features and constructing appropriate navigation region calibration criteria play an important role in screening underwater navigation adaptation areas, improving the positioning precision of underwater vehicles and ensuring the stability of navigation.

[0004] Generally speaking, the more intense the gravity anomaly changes in the matching domain, the better the matching performance. In the gravity flat area, due to the similarity of the positions indicated by the INS to the gravity features, the gravity aided navigation system will produce a large matching error. In order to ensure the accuracy and stability of underwater gravity aided navigation, the research on matching domain adaptation analysis has attracted widespread attention in recent years. The existing research usually takes the maximization of the information richness of the gravity anomaly features of each region as the target to calibrate the adaptation area, which is a model-driven analysis method. This method has a relatively low computational complexity, but the performance of the model depends more on the quality and number of the gravity anomaly features selected empirically, and it is difficult to construct a matching model associated with the gravity anomaly features and the adaptation area. The data-driven learning method can model without prior knowledge and noise statistical information, avoiding artificial intervention and empirical judgment, and has better performance than the model-driven analysis method. SUMMARY:

[0005] The application aims at the defects and deficiencies in the prior art, and proposes an underwater navigation adaptation area connectivity improvement method based on modified domain boundary screening, which is driven by navigation calibration data, uses a modified domain migration mechanism and an out-of-domain sequence reduction mechanism to compensate for the evaluation blank of the TERCOM algorithm navigation reserved boundary, and can effectively improve the calibration accuracy of the adaptation area.

[0006] The application achieves the above-mentioned purposes by the following measures:

[0007] An underwater navigation adaptation area connectivity improvement method based on modified domain boundary screening, characterized by constructing a gravity anomaly feature and adaptation area association model EBMAS, and using the EBMAS model to complete underwater navigation adaptation area calibration, wherein in the EBMAS model construction process, first, the modified domain migration mechanism and the out-of-domain sequence reduction mechanism are used to compensate for the evaluation blank of the TERCOM algorithm navigation reserved boundary, and the adaptation area is calibrated; second, the training data set is obtained by combining the adaptation area class label and the gravity anomaly feature; then, the range control synthetic minority over-sampling technique RCSMOTE is combined to control the position of the newly synthesized adaptation area sample in the data space, to identify potential adaptation area samples and enhance the classification and identification performance of the calibration model; thereafter, on the balanced calibration data set, the maximum relevance minimum redundancy algorithm mRMR is combined to screen a comprehensive key gravity anomaly feature set to improve the calibration accuracy of the adaptation area.

[0008] The application improves the TERCOM algorithm by the modified domain migration mechanism and the out-of-domain sequence reduction mechanism, eliminates the navigation reserved boundary, and ensures the comprehensiveness of the matching area gravity evaluation, which can effectively enhance the connectivity of the adaptation area calibration without shortening the sliding step length, wherein the modified domain migration mechanism includes the following contents: let the INS indicated position at the t-th time be ω is the side length of the square modified domain grid, D is the set of square modified domain grid point positions, and R is the set of matching domain grid points;

[0009] If (x,y) is in R, then the error domain generated by is not out of bounds; if (x,y) is not in R, then indicates that the error point (x,y) is out of bounds, and the out-of-bound mode is divided into single-edge out-of-bound and double-edge out-of-bound, when the single-edge out-of-bound, the out-of-bound grid (x out ,y out ) is parallelly migrated to the in-domain grid point symmetrical about the INS indicated position, and the single-edge migration mechanism is

[0010]

[0011] When the bilateral out-of-boundary, the out-of-boundary grid points will be migrated twice to the corresponding area, and the bilateral migration mechanism is

[0012]

[0013] The present application takes each grid point of the newly revised square field as a starting point, generates a sequence in reverse order parallel to the real track, and uses mean square error as a correlation evaluation index and determines the best matching position according to the optimal matching index principle. However, when the error grid point is far away from the real end point position, the starting position of part of the revised sequence will exceed the matching domain, resulting in matching failure. The revised sequence 1 is the revised sequence closest to the real end point position, the revised sequences 2 and 3 are the non-out-of-boundary sequence and the out-of-boundary sequence respectively, the corresponding end point revised position of the out-of-boundary sequence 3 is simplified to obtain a secondary revised square field,

[0014] Let the revised sequence be Where sample is the sampling number, is the navigation starting point of the revised sequence, is the navigation end point of the revised sequence,

[0015] When When the real end point position has a large deviation, the navigation track generated by the end point has part of the sampling points exceeding the boundary. At this time, the part of the sampling points in the revised square field become the best matching position of the end point with a lower probability, and the out-of-boundary sequence generated by the sampling points leads to the matching failure of the TERCOM algorithm. Based on this, the out-of-domain sequence simplification mechanism eliminates the error grid points in the revised square field to ensure the executability of the matching algorithm and improve the matching positioning efficiency, that is, the sequence Seq out is generated in reverse order from any grid point in the revised sequence D, Then

[0016]

[0017] In the present application, the gravity anomaly characteristics can reflect the gravity anomaly change of the matching area, and the more intense the change of the matching area characteristics, the better the matching effect. In order to study the correlation between the gravity anomaly characteristics and the navigation adaptation area, the gravity anomaly characteristic value of the sub-matching domain is used as the characteristic data of the training calibration model, wherein g(i,j) represents the gravity anomaly value at the grid (i,j) position, is the average value of the gravity anomaly in the matching domain, M and N are the number of grid points across the matching domain in longitude and latitude respectively, and the expression of the gravity anomaly characteristics is

[0018] (1) Gravity anomaly standard deviation (σ):

[0019]

[0020] (2) the coefficient of kurtosis (C e ) :

[0021]

[0022] (3) the coefficient of skewness (C s ) :

[0023]

[0024] (4) the roughness (R) :

[0025]

[0026] R = (R λ + R φ ) / 2 (8), where R λ and R φ are the roughness in the east-west and north-south directions, respectively; g

[0027] (6) the standard deviation of slope (σ s ) :

[0028]

[0029] (7) the differential entropy of gravity anomaly (H) : where the local differential probability is the differential value of gravity anomaly

[0030] (8) the correlation coefficient of gravity anomaly (RC) :

[0031] RR = (RC λ + RC φ ) / 2 (16),

[0032]

[0033] (9) the average gravity difference (AGD) :

[0034]

[0035] (10) the integral of the standard deviation of gravity anomaly and the absolute roughness (Γ) :

[0036]

[0037] S Γφ = [g(i,j) - g(i,j-1)] 2 + [g(i,j) - g(i,j+1)] 2 (22),​

[0038] S Γλ = [g(i,j) - g(i-1,j)] 2 + [g(i,j) - g(i+1,j)] 2 (23),

[0039] (11) Slope in the direction of longitude and latitude (S) :

[0040]

[0041] (12) Coefficient of variation of gravity anomaly (C v ) :

[0042] (13) Local energy (E) : The 13 gravity anomaly features in each matching field are normalized to remove the dimensional influence between different features, and the navigation result is obtained to obtain the training calibration data set.

[0043] The application also includes balancing and feature selection of the training calibration data, which includes: (1) RCSMOTE balancing method, the execution steps are:

[0044] Step (1-1): separate the calibration data set into adaptive region samples and non-adaptive region samples;

[0045] Step (1-2): in the adaptive region samples, identify the boundary samples, noise samples and safe samples;

[0046] Step (1-3): oversample based on range control for majority class, boundary and noise samples;

[0047] (2) mRMR feature screening method, the execution steps are:

[0048] Step (2-1): using the balanced calibration data set, the first gravity anomaly feature is screened out by using the mutual information maximization method;

[0049] Step (2-2): according to the mRMR criterion, the contribution value of each feature is calculated, and the gravity anomaly feature with the largest information is selected as the selected feature;

[0050] Step (2-3): the selected gravity anomaly feature set is used to train the KNN (K=5) classifier to obtain the adaptive region self-calibration model.

[0051] Compared with the prior art, the application can effectively improve the comprehensiveness and accuracy of the adaptive region calibration, and still show strong applicability when the model is migrated to a new region. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1Fig. 1 is a schematic diagram of the navigation reserved boundary in the present application.

[0053] Figure 2 Fig. 2 is a schematic diagram of the modified square domain migration mechanism in the present application.

[0054] Figure 3 Fig. 3 is a schematic diagram of the modified sequence out-of-boundary and reduction in the present application.

[0055] Figure 4 Fig. 4 is the gravity distribution of the research area in the embodiment of the present application, wherein Figure 4 Fig. 4(a) is a water depth reference map, and Fig. 4(b) is a gravity anomaly reference map.

[0056] Figure 5 Fig. 5 is the underwater gravity matching mechanism of the EBMAS algorithm in the embodiment of the present application, wherein Figure 5 Fig. 5(a) is the modified square domain migration, and Fig. 5(b) is the out-of-boundary sequence reduction.

[0057] Figure 6 Fig. 6 is the calibration comparison results of the EBMAS algorithm in the embodiment of the present application, wherein Figure 6 Fig. 6(a) is the TERCOM adaptation zone calibration, s=100; Fig. 6(b) is the TERCOM adaptation zone calibration, s=50; Fig. 6(c) is the adaptation zone calibration of the algorithm, s=100; and Fig. 6(d) is the adaptation zone calibration of the algorithm, s=50.

[0058] Figure 7 Fig. 7 is the multi-azimuth calibration results in the embodiment of the present application, wherein Figure 7 Fig. 7(a) is the northeast azimuth calibration results; Fig. 7(b) is the due east azimuth calibration results; Fig. 7(c) is the southeast azimuth calibration results; and Fig. 7(d) is the due south azimuth calibration results.

[0059] Figure 8 Fig. 8 is the calibration results of different gravity anomaly characteristics in the embodiment of the present application, wherein Figure 8 Fig. 8(a) is the model performance of different characteristic numbers; and Fig. 8(b) is the calibration results when the number of characteristics is 6.

[0060] Figure 9 Fig. 9 is the migration calibration results of the EBMAS algorithm in the embodiment of the present application, wherein Figure 9 Fig. 9(a) is the new area calibration and track line; Fig. 9(b) is the adaptation zone matching test effect; Fig. 9(c) is the new area calibration and track line; and Fig. 9(d) is the adaptation zone matching test effect. DETAILED DESCRIPTION

[0061] The present application will be further described below in conjunction with the drawings and embodiments.

[0062] The example proposes a new type of correction domain boundary balancing screening method to improve the connectivity of underwater gravity navigation adaptation area, including: first, in order to ensure the feasibility of TERCOM algorithm in the navigation evaluation of the boundary of the region, the correction square migration mechanism and the sequence reduction mechanism outside the domain are proposed, and the navigation calibration adaptation area is proposed. Second, the traditional gravity anomaly characteristics are briefly introduced, and the training data set is obtained combined with the adaptation area class label. Finally, the RCSMOTE algorithm and the mRMR algorithm are used to process the unbalanced calibration data and the redundant gravity anomaly characteristics respectively.

[0063] The RCSMOTE algorithm and the mRMR algorithm are used to process the unbalanced calibration data and the redundant gravity anomaly characteristics respectively.

[0064] In order to comprehensively evaluate the gravity anomaly change in the matching domain, the total matching domain is divided into multiple sub-matching domains by using a fixed size sliding window and a sliding step, and then the TERCOM algorithm is used to simulate navigation and calibrate the adaptation area one by one. As a classic terrain matching algorithm, TERCOM algorithm greatly reduces the error of single point matching algorithm using sequence correlation analysis method, and has the advantages of insensitive to initial error, high positioning accuracy and strong stability. At the same time, TERCOM algorithm is sensitive to the characteristics of gravity field change, which is beneficial to the analysis and training calibration of regional adaptability. When using TERCOM algorithm for calibration, in order to comprehensively evaluate the overall gravity anomaly change of the matching domain, the navigation track should cover the matching domain as much as possible. At the same time, evaluating the gravity anomaly change of the boundary of each matching domain is also conducive to enhancing the connectivity between the calibration areas, ensuring the navigation continuity of the selected adaptation area in the field of underwater obstacle detection and path planning, and then enhancing the diversity of the task route of underwater vehicle, and ensuring its navigation safety and reliability. The traditional TERCOM algorithm takes the end point of the INS output track as the domain center, and takes 3 times the cumulative drift error of the inertial navigation as the half side length to form a square matching domain, and then searches the grid points to get the best matching position, realizes the error correction and positioning calibration of the INS system of underwater vehicle. However, in order to ensure the feasibility of navigation, part of the boundary area of the sub-matching domain needs to be reserved for generating the correction square, as shown in Figure 1 This will cause a large area of horizontal and vertical navigation reserved boundary between adjacent sub-matching domains, which will cause the adaptation area search not comprehensive and the connectivity between adaptation areas reduced.

[0065] In order to solve the above problems, the example proposes an improved TERCOM algorithm to eliminate the navigation reserved boundary to ensure the comprehensiveness of the matching area gravity evaluation, which can effectively enhance the connectivity of the adaptation area calibration without shortening the sliding step. The matching domain calibration method with enhanced connectivity will respectively solve the problems of correction square out-of-bound and matching sequence out-of-bound, and propose two solutions: correction square migration mechanism and sequence reduction mechanism outside the domain.

[0066] In the process of the calibration of the adaptive region, the track should cover the matching region as much as possible to ensure the comprehensiveness of the current heading evaluation. At this time, the real position of the end point and the position of the end point indicated by the INS are both located at the edge of the matching region with a high probability, which leads to the out-of-bound of the square correction region generated by the center of the indicated end point position, and further limits the TERCOM algorithm to sail near the edge of the matching region.

[0067] Let the INS indicated position at the tth time be ω is the side length of the square correction region grid, D is the set of the square correction region grid points, R is the set of the matching region grid points, and if (x, y) ∈ R, then the error square region generated by is not out-of-bound; if (x, y) ∉ R, it is illustrated that the error point (x, y) is out-of-bound, which can be divided into single-edge out-of-bound and double-edge out-of-bound, as shown in Figure 2

[0068] When the single-edge out-of-bound occurs, as shown in Figure 2 (a), the out-of-bound grid (x out ,y out ) is parallelly migrated to the in-bound grid point symmetric about the indicated position of the INS, and the single-edge migration mechanism is When the double-edge out-of-bound occurs, as shown in Figure 2 (b), the out-of-bound grid point will be migrated twice to the corresponding region, and the double-edge migration mechanism is

[0069] The migration mechanism of the proposed correction square region is to parallelly migrate the out-of-bound error point to the in-bound region, which not only ensures the feasibility of the TERCOM algorithm in the entire matching region, but also enhances the matching accuracy in the model region by expanding the error matching area in the region by probability.

[0070] Taking each grid point of the new correction square region as the starting point, a sequence parallel to the real track is generated in reverse order, and the mean square deviation is used as the correlation evaluation index to determine the optimal matching position according to the optimal matching index principle. However, when the distance between the error grid point and the real end point position is far, the starting position of part of the correction sequence will be out of the matching region, resulting in matching failure, as shown in Figure 3 The correction sequence ① is the correction sequence closest to the real end point position, and the correction sequences ② and ③ are the in-bound and out-of-bound sequences, respectively. The out-of-bound sequence ③ is far away from the real end point and the indicated position of the INS, and the corresponding end point correction position is simplified to obtain the secondary correction square region, i.e., the green grid point set of Figure 3 (b).

[0071] Let the correction sequence be where sample is the sampling number, is the sailing starting point of the correction sequence, ​To amend the sequence of the end of the voyage.

[0072] According to Figure 3 (a) can be known, when When there is a large deviation from the true end position, the generated voyage trajectory with this end position has some sampling points exceeding the boundary. At this time, this part of the sampling points in the amendment domain has a lower probability of becoming the best matching position of the end, and the sequence generated by the sampling points is out of the boundary, which leads to

[0073] TERCOM algorithm matching failure. Based on this, the out-of-domain sequence reduction mechanism eliminates the error grid points of the matching sequence out of the boundary in the amendment domain Figure 3 (b)), to ensure the executability of the matching algorithm and improve the matching positioning efficiency. That is, the sequence Seq out in the amendment sequence D is generated in reverse order

[0074] Then

[0075] The out-of-domain sequence reduction mechanism eliminates the error correction grid points far from the true end position, which not only ensures the feasibility of the TERCOM algorithm in the region navigation, but also improves the navigation matching efficiency to some extent.

[0076] The gravity anomaly characteristics can reflect the gravity anomaly changes of the matching region to some extent, and the more dramatic the change of the matching region characteristics, the better the matching effect. To study the correlation between the gravity anomaly characteristics and the navigation adaptation area, the gravity anomaly characteristic value of the sub-matching domain is used as the feature data for training the calibration model; g(i,j) represents the gravity anomaly value at the grid (i,j) position, is the average value of the gravity anomaly in the matching domain, and M and N are the number of grid points across the matching domain in longitude and latitude, respectively. The expression of the gravity anomaly characteristics is

[0077] (1) Gravity anomaly standard deviation (σ):

[0078]

[0079] (2) Kurtosis coefficient (C e ): (4);

[0080] (3) Skewness coefficient (C s ):

[0081]

[0082] (4) Roughness (R):

[0083] R = (R λ + R φ ) / 2 (8), where R λ and R φ are the roughness in the longitude and latitude directions, respectively;

[0084] (5) Gravity anomaly feature abundance (A g ):

[0085] (6) Slope standard deviation (σ s ):

[0086]

[0087] (7) Gravity anomaly difference entropy (H): where the local difference probability is the gravity anomaly difference value

[0088] (8) Gravity anomaly correlation coefficient (RC):

[0089] RR = (RC λ + RC φ ) / 2 (16),

[0090]

[0091] (9) Average gravity difference (AGD):

[0092]

[0093] (10) Integral of gravity anomaly standard deviation and absolute roughness (Γ):

[0094]

[0095] S Γφ = [g(i,j) - g(i,j-1)] 2 + [g(i,j) - g(i,j+1)] 2 (22),

[0096] S Γλ = [g(i,j) - g(i-1,j)] 2 + [g(i,j) - g(i+1,j)] 2 (23),

[0097] (11) Longitude and latitude direction slope (S):

[0098]

[0099]

[0100] (12) Coefficient of variation of gravity anomaly (C v ) :

[0101] (13) Local energy (E) :

[0102] The 13 gravity anomaly features in each matching field are normalized to remove the dimensional influence between different features, and the navigation results are obtained to train the calibration dataset.

[0103] The balancing and feature selection of the training calibration data include the following contents:

[0104] (1) RCSMOTE balancing method

[0105] The number of samples in the adaptation area in the total matching field is much smaller than that in the non-adaptation area, and it is easy to have a data imbalance problem. If the unprocessed non-balanced dataset is directly used, the calibration model will be inclined to the non-adaptation area samples, resulting in the adaptation area samples being ignored and misclassified. In view of the problem that the calibration model is only applicable to balanced data sets, the RCSMOTE algorithm is used to oversample the non-balanced data, realize the balancing of the training dataset, and improve the classification recognition performance of the calibration model.

[0106] The RCSMOTE algorithm is used to solve the imbalance problem in the dataset, which can effectively reduce the probability of misclassification of the adaptation area samples. In order to overcome the problems of oversampling of noise samples, oversampling of non-information samples and excessive overlap of boundary increased samples in traditional SMOTE algorithm, RCSMOTE realizes the oversampling technology by accurately controlling the position of samples in the data space, and the main execution steps are as follows:

[0107] Step (1-1): Separate the calibration dataset into adaptation area samples and non-adaptation area samples.

[0108] Step (1-2): In the adaptation area samples, identify the boundary samples, noise samples and safe samples.

[0109] Step (1-3): Range control-based oversampling of majority class, boundary and noise samples.

[0110] (2) mRMR feature selection method

[0111] The regional gravity anomaly change can generally be reflected by multiple gravity anomaly features, but the redundancy or collinearity between multiple features will affect the accuracy and comprehensiveness of the adaptation area calibration. In order to improve the performance of the underwater matching area self-calibration method, the mRMR feature selection algorithm is used to screen the key gravity anomaly features and remove redundant features.

[0112] mRMR is a filter feature selection algorithm that can significantly improve the classification accuracy and has lower computational complexity. The mRMR algorithm uses the feature contribution evaluation criterion of maximizing feature correlation and minimizing feature redundancy, and obtains the feature set with the highest correlation with the matching area in the previous sequence search method. The main execution steps are as follows:

[0113] Step (2-1): The first gravity anomaly feature is selected by using the mutual information maximization method with the balanced calibration data set.

[0114] Step (2-2): The contribution value of each feature is calculated according to the mRMR criterion, and the gravity anomaly feature with the maximum information is selected as the selected feature.

[0115] Step (2-3): The selected gravity anomaly feature set is used to train the KNN (K=5) classifier to obtain the self-calibration model of the matching area.

[0116] In order to verify the feasibility and effectiveness of the new matching area calibration method of underwater gravity aided navigation, the example data in the South China Sea is used as the simulation sample, and the calibration method of enhancing the connectivity of the boundary, the calibration data balancing, the key gravity feature screening and the training transfer of the calibration model are verified respectively.

[0117] The example data is from the website of the University of California, San Diego (http: / / topex.ucsd.edu / ) with a resolution of 1'x1' regional water gravity anomaly data, and the longitude and latitude value range is 114°E-116°E, 10°N-12°N; the gravity anomaly reference data is converted to 100m x 100m grid resolution by bilinear interpolation method, as shown in Figure 4 The average gravity anomaly of the region is 17.97 mGal, the maximum value is 136.40 mGal, and the minimum value is -36.43 mGal.

[0118] All experiments are run on a computer with a CPU of 2.50 GHz and a memory of 16 G using Matab2020a software. In the simulation sample area, the area is divided into 484 sub-areas by sliding window, and the sliding window size is 100x100 grid range, and the step is 100 grid range. In each sub-area, the TERCOM algorithm is used for trajectory simulation.

[0119] In order to explore the reasons for the boundary navigation feasibility and overall evaluation of the EBMAS algorithm relative to TERCOM, the correction domain migration mechanism and the out-of-domain sequence reduction mechanism are drawn when the correction domain is out of boundary and the matching sequence is out of boundary, as shown in Figure 5 .

[0120] According to Figure 5(a) It is known that the EBMAS algorithm migrates the out-of-bound error correction field to the red region, ensuring the feasibility of underwater vehicle navigation at any position in the matching field. Furthermore, the algorithm reduces the error grid corresponding to the out-of-bound sequence in the correction field, such as Figure 5 (b) The yellow region ensures the feasibility of sequence matching and improves matching efficiency.

[0121] The traditional TERCOM algorithm has a small navigation range in the sub-region, so the sliding step can be shortened to compensate for the defects of incomplete search for the adaptation area and reduced inter-regional connectivity caused by the track reservation boundary. Therefore, the same starting point ([2000, 2000]) and heading angle (45°), sampling number 40, and sliding step 100 and 50 grid ranges (a total of 1849 sub-matching fields) were simulated to test the accuracy and comprehensiveness of the improved TERCOM algorithm calibration. The positioning accuracy less than the grid diagonal line ( ) was calibrated as the adaptation area, and the calibration results are shown in Figure 6 .

[0122] From Figure 6 (a) and (c), it can be seen that when the sliding step is the same, the number of adaptation areas calibrated by the proposed algorithm increases from 30 to 105, indicating that the EBMAS model calibration is more comprehensive and has better connectivity. The main reason is that the proposed algorithm solves the navigation reservation boundary problem caused by the error correction process of the TERCOM algorithm, which increases the track coverage area and makes the evaluation of the overall gravity field change more complete. Figure 6 (b) The area of the adaptation area is relatively Figure 6 (a) increased, indicating that shortening the sliding step can compensate for the navigation blank of the matching field to some extent, but the calibration effect is still not as good as the proposed algorithm (s = 100). Figure 6 (d) Further shortening the sliding step finds that the EBMAS algorithm enhances regional connectivity based on more comprehensive coverage of areas with severe gravity anomaly changes, effectively expanding the underwater vehicle activity range and improving task execution efficiency and safety.

[0123] To reduce the computational complexity of the calibration process, the step size s is set to 100 grid ranges. The EBMAS algorithm is used to calibrate the adaptation area from the northeast direction (heading angle 45°), the east direction (heading angle 90°), the southeast direction (heading angle 135°), and the south direction (heading angle 180°), as shown in Figure 7 .

[0124] According to Figure 7It can be seen that the adaptive regions of the proposed algorithm in different directions are not completely the same, but they basically cover the areas where the gravity anomaly changes obviously. In addition, the high connectivity between regions also shows that the proposed algorithm can effectively improve the comprehensiveness and connectivity of the calibration region. In order to train a calibration model with direction adaptability, regions with three or more direction adaptabilities are recorded as comprehensive adaptive regions. The number of adaptive regions in the proposed algorithm in four directions is less than the number of non-adaptive regions, and both show an imbalance. In particular, the comprehensive adaptive region with multi-direction adaptability has an imbalance ratio of 4.56, which needs to be balanced before model training and calibration.

[0125] In order to solve the problems of class label imbalance and gravity feature redundancy in the calibration data, the RCSMOTE algorithm is used to balance the data set, and the mRMR algorithm is used to extract key gravity anomaly features. In the unbalanced classification problem, the classification accuracy cannot effectively evaluate the classification performance of the minority class samples, so AUC and F-measure are selected as evaluation indicators. Figure 8 (a) shows the influence of different gravity anomaly feature numbers on the performance of the balanced calibration model. With the increase of the number of gravity anomaly features, the AUC value and F-measure show a trend of first increasing and then decreasing, indicating that the information interaction of multiple features can enhance the classification and recognition performance of the calibration model, but it will also be reduced by the insertion of redundant features. Obviously, the calibration model can achieve the optimal classification and recognition performance when 6 features are selected. The 6 features are shown in the order of selection, which are the product of the standard deviation of gravity anomaly and the integral of absolute roughness, skewness, local gravity anomaly feature abundance, local energy, average gravity difference, and slope standard deviation.

[0126] The 6-feature model is used to calibrate the training region, and the results are shown in Figure 8 (b). The red calibration box is the consistent prediction label and true classification label of the training model, the green calibration box is the prediction of the non-adaptive region by the prediction model, and the yellow calibration box is the prediction of the adaptive region by the prediction model. The proposed model has high recognition accuracy for adaptive regions, with only a few adaptive regions misclassified. At the same time, the proposed method reduces the sample imbalance rate to 1.0206, and the AUC and F-measure indicators are improved by 16.73% and 19.11% respectively compared with the original calibration data set.

[0127] The trained model is now transferred to a new region, and the calibration results are shown in Figure 9 As shown in Figure 9 (a) and (c), the calibration region basically covers the contour dense area, and most of the regions have certain connectivity. At the same time, in order to test the accuracy of the calibration and its applicability in actual navigation, a long track is simulated in the adaptive region, and the matching accuracy is shown in Figure 9The results of the experiments (b) and (d) show that the underwater vehicle has higher matching accuracy and stability when it is used in long voyage tasks in the adaptation region, which indicates that the proposed calibration model has good regional migration applicability.

Claims

1. A method for improving connectivity of underwater navigation adaptation zones based on modified region border screening, characterized in that, The gravity anomaly feature and adaptation zone correlation model EBMAS is constructed, and the underwater navigation adaptation zone calibration is completed by using the EBMAS model. In the process of constructing the EBMAS model, first, the modified domain migration mechanism and the out-of-domain sequence reduction mechanism are used to compensate for the evaluation blank of the TERCOM algorithm navigation reserved boundary and calibrate the adaptation zone. Second, the training data set is obtained by combining the adaptation zone class label and the gravity anomaly feature. Then, the range control synthetic minority over-sampling technique RCSMOTE is used to control the position of the newly synthesized adaptation zone samples in the data space, to identify potential adaptation zone samples and enhance the classification and identification performance of the calibration model. After that, on the balanced calibration data set, the maximum relevance minimum redundancy algorithm mRMR is used to screen the key gravity anomaly feature set to improve the accuracy of the adaptation zone calibration.

2. The method of claim 1, wherein, The TERCOM algorithm is improved by using the modified domain migration mechanism and the out-of-domain sequence reduction mechanism to eliminate the navigation reserved boundary and ensure the comprehensiveness of the matching area gravity evaluation, which can effectively enhance the connectivity of the adaptation zone calibration without shortening the sliding step. The modified domain migration mechanism includes the following contents: Let the position indicated by the tth INS be ω is the side length of the square correction domain grid, D is the set of square correction domain grid point positions, and R is the set of matching domain grid points. like If (x,y)∈R, then by The error domain generated centered on the center is not out of bounds; if satisfy This indicates that the error point (x, y) has gone out of bounds. There are two types of out-of-bounds errors: one-sided and two-sided. When it is one-sided, the grid outside the boundary (x, y) will be... out ,y out Parallel migration to a grid point within the domain symmetric about the inertial navigation indicator position; its unilateral migration mechanism is as follows: When the bilateral boundary is out of range, the out-of-range grid points will be migrated to the corresponding area twice, and the bilateral migration mechanism is 3. The method of claim 1, wherein, With each grid point of the new revised square field as the starting point, a sequence parallel to the real track is generated in reverse order, and the mean square error is used as the correlation evaluation index and the best matching position is determined according to the optimal matching index principle. However, when the distance between the error grid point and the real end point position is far, the starting position of part of the revised sequence will exceed the matching domain, resulting in matching failure. The revised sequence ① is the revised sequence closest to the real end point position, the revised sequences ② and ③ are respectively the non-out-of-boundary sequence and the out-of-boundary sequence, the corresponding end point revised position of the out-of-boundary sequence ③ is far away from the real end point and the INS indicated position, and the secondary revised square field is obtained by simplifying the corresponding end point revised position. Let the revised sequence be where sample is the sampling number, is the sailing starting point of the revised sequence, is the sailing end point of the revised sequence, When When there is a large deviation from the real end position, the navigation trajectory generated with the end position has some sampling points beyond the boundary. At this time, the part of sampling points in the modified domain become the best matching position of the end position with a lower probability, and the out-of-bound sequence generated by the sampling points leads to the failure of the TERCOM algorithm matching. Based on this, the out-of-bound sequence reduction mechanism eliminates the error grid points of the matching sequence to ensure the executability of the matching algorithm and improve the matching positioning efficiency. That is, the sequence Seq out In the formula, l = 1, 2, …, sample, and l = 1, 2, …, sample, and 4. The method of claim 1, wherein, The gravity anomaly feature can reflect the gravity anomaly change of the matching area, and the more intense the feature change of the matching area is, the better the matching effect is. In order to study the correlation between the gravity anomaly feature and the navigation adaptation area, the gravity anomaly feature value of the sub-matching domain is taken as the feature data of the training calibration model, wherein g(i,j) represents the gravity anomaly value at the grid (i,j) position, is the gravity anomaly average value in the matching domain, M and N are the number of horizontal grids of the matching domain in the longitude and latitude respectively, wherein the expression of the gravity anomaly feature is (1) Gravity anomaly standard deviation (σ): (2) coefficient of kurtosis (C e ): (3) Coefficient of Skewness (C s ): (4) Roughness (R): R = (R λ + R φ ) / 2 (8), where R λ and R φ are the warp and weft roughness, respectively. (5) the gravity anomaly feature abundance (A g ): (6) Slope standard deviation (σ s ): (7) Gravity anomaly difference entropy (H): where the local difference probability Gravity anomaly difference value (8) Gravity anomaly correlation coefficient (RC): RR = (RC λ + RC φ ) / 2 (16), (9) Average gravity difference (AGD): (10) Integral of gravity anomaly standard deviation and absolute roughness (Γ): S Γφ = [g(i,j) - g(i,j - 1)] 2 + [g(i,j) - g(i,j + 1)] 2 (22), S Γλ = [g(i,j) - g(i - 1,j)] 2 + [g(i,j) - g(i + 1,j)] 2 (23), (11) Slope of longitude and latitude direction (S): (12) Coefficient of variation of gravity anomaly (C v ): (13) Local energy (E): The normalized method is used to remove the dimensional influence between different features in the 13 gravity anomaly features in each matching field, and the training calibration dataset is obtained by navigation results.

5. The method of claim 1, wherein, It also includes the balancing and feature selection of the training calibration data, which includes: (1) RCSMOTE balancing method, the execution steps are: Step (1-1): Separate the calibration data set into adaptation zone samples and non-adaptation zone samples; Step (1-2): In the adaptation zone samples, identify the boundary samples, noise samples and safe samples; Step (1-3): Perform over-sampling based on range control on the majority class, boundary and noise samples; (2) mRMR feature selection method, the execution steps are: Step (2-1): Use the balanced calibration data set to select the first gravity anomaly feature by using the mutual information maximization method; Step (2-2): Calculate the contribution value of each feature according to the mRMR criterion, and select the gravity anomaly feature with the largest information amount as the selected feature; Step (2-3): Train the KNN (K=5) classifier with the selected gravity anomaly feature set to obtain the adaptation zone self-calibration model.

Images