An underwater multi-source navigation positioning method based on submarine oil and gas pipeline detection

Abstract

The application provides an underwater multi-source navigation positioning method based on seabed oil and gas pipeline detection, establishes an information filtering model of centralized SINS / DVL / USBL combined navigation, and obtains global optimal prediction estimation of navigation error state; then, a SINS / DVL / USBL combined navigation system state model and a SINS / DVL / USBL combined navigation observation model are established respectively; through the improved DS-UKF algorithm, the delayed measurement information is fully utilized to improve the precision of the nonlinear state estimation; finally, through the outlier detection algorithm of the weighted one-class SVM, the noise and jump point problems of the USBL data are effectively solved, and the underwater navigation precision is improved. The application effectively solves the underwater positioning precision problem of the long-time pipeline detection robot, and greatly improves the underwater positioning precision by comprehensively considering factors such as information time delay and information mutation.

Description

Technical Field

[0001] This invention relates to the field of marine engineering technology, and specifically to an underwater multi-source navigation and positioning method based on the detection of subsea oil and gas pipelines. Background Technology

[0002] Subsea pipeline systems are the lifeline for marine energy transportation, serving as vital facilities connecting underwater wellheads, offshore platforms, and land-based terminals. As an important component of offshore platforms, subsea pipelines are widely used for the efficient and safe long-distance transportation of water, natural gas, oil, and waste.

[0003] Submarine oil and gas pipelines are the most expensive, largest, and most widely installed component of submarine infrastructure. Due to the complex and variable subsea environment, their safety and stability cannot be fully guaranteed. For a long time, submarine pipelines have been affected by the uncertainties of the marine environment, such as changes in geological structure, seawater corrosion, and sediment burial, which can lead to pipeline damage and easily cause accidents such as blockages, corrosion, cracks, suspension, and even leaks. Since pipelines often carry hazardous substances, they may damage the ecological environment and cause huge economic losses. Therefore, regular routine inspections of pipelines are necessary to ensure their safe and stable service life.

[0004] Due to the complexity of the marine environment, locating and pinpointing damage points is often intricate and difficult. Furthermore, the volatile nature of oil and gas makes it challenging to detect leaks through pipeline traces. As pipeline mileage continues to increase, routine underwater pipeline inspection tasks become extremely arduous and expensive. In this context, unmanned underwater vehicles (UUVs) have become a superior choice for pipeline inspection and maintenance due to their high stability, continuity, and autonomy.

[0005] In recent years, underwater robots have demonstrated significant advantages in marine research, the offshore oil and gas industry, and military operations. Therefore, as an essential tool for subsea pipeline inspection, underwater robotic subsea pipeline inspection systems hold important strategic significance. Typically, external inspections of subsea pipelines are performed by remotely operated vehicles (ROVs). However, ROVs require cable support, resulting in low efficiency and high cost. With the rapid development of marine development, AUVs, capable of autonomously performing more complex tasks such as oil extraction and subsea pipeline inspection, have attracted widespread attention. AUVs provide continuous autonomous positioning information for seabed resource exploration, offshore drilling, subsea pipeline laying, and maintenance, becoming an effective tool for seabed exploration and marine research.

[0006] Currently, using autonomous underwater vehicles (AUVs) equipped with acoustic, visual, and magnetic sensors to detect subsea oil and gas pipelines is effective and feasible, featuring miniaturization, economy, and intelligence. For long-endurance underwater operations, the navigation path of an AUV is often linear, requiring highly accurate navigation information. Single navigation methods, such as inertial navigation, accumulate errors over time and with the passage of time, causing the AUV to deviate from its course. Therefore, external navigation equipment is needed for auxiliary positioning. Currently, the combination of a strapdown inertial navigation system (SINS) and a Doppler velocity log (DVL) is the main underwater integrated navigation method, but it still suffers from error accumulation. Using an unmanned surface vessel (USV) equipped with an ultra-short baseline acoustic positioning system (USBL) can effectively suppress the divergence of position errors in the SINS / DVL navigation system, improving underwater positioning accuracy, and is a highly effective technical means.

[0007] Inertial navigation boasts advantages such as comprehensive and omnidirectional information parameters and strong anti-interference capabilities. However, a single sensitive device cannot meet the requirements for enhancing information redundancy and ensuring continuous reliability. Therefore, under long-endurance and long-distance operational conditions, it is necessary to select appropriate auxiliary navigation devices and implement multi-source information fusion technology to achieve target positioning, identification, and tracking. In the field of multi-source information fusion, commonly used estimation algorithms include the Kalman Filtering (KF) algorithm based on Bayesian theory and its derivatives, such as the Extended Kalman Filter (EKF) and Unscented Kalman Filter for nonlinear conditions, as well as centralized or distributed Kalman filters based on different framework structures. Over the years, scholars have proposed many multi-source information fusion algorithms to solve specific problems in the field of underwater navigation and positioning.

[0008] Xu Bo et al. proposed a novel robust filter, the Huber-M estimation delay Kalman filter, for the SINS / USBL integrated navigation system, to handle time-varying delays in underwater acoustic communication and non-Gaussian noise caused by outliers. Based on the time delay characteristics of USBL underwater acoustic communication, a time delay system model based on state inversion was derived. A linear recursive model based on the time delay system model was constructed by combining posterior estimation and covariance matrix with Huber-M estimation theory. This model is robust to outliers. Zhao L proposed a gridded polar region SINS / DVL / USBL integrated navigation algorithm. Addressing the low update frequency of USBL, a multi-frequency Kalman filter was proposed to provide a higher and more stable update frequency. Liu H proposed a SINS / DVL / USBL integrated navigation system based on distance / depth measurement for TUV deep-sea operations. When velocity information from the DVL sensor is missing or the seabed depth exceeds the DVL bottom lock range, high-precision position information can be obtained through the USBL system and pressure sensors. For data fusion, an integrated filtering algorithm based on interactive multi-model filtering and adaptive robust square root filtering was proposed. To improve the positioning accuracy of USBL, Xu Y designed a novel positioning system based on an adaptive Kalman filter. AKF utilizes the latest measurement data to adjust the measurement covariance matrix and process noise covariance matrix. Furthermore, a forgetting factor is set in the filter to limit the filter's storage length, thereby making full use of the current measurement data and reducing the impact of USBL measurement anomalies. To reduce USBL positioning errors, Chenglong Xu designed a conditional adaptive gain extended Kalman filter algorithm, which mainly introduces conditional constraints and confidence metric operators, and has a certain universality for USBL variation noise. To suppress the impact of USBL outliers and long-term data interruptions on underwater positioning accuracy, Wang J proposed an outlier detection and classifier algorithm based on the median absolute deviation outlier detection filter, and proposed a Kalman filter based on Student t, which provides robustness to measurement uncertainties caused by the existence of outliers.

[0009] Drawing on research experience from other methods, this invention proposes a multi-source positioning method for subsea oil and gas pipeline inspection, enabling precise underwater navigation and positioning of pipeline inspection robots under long-endurance conditions. Summary of the Invention

[0010] This invention addresses the need for accurate fault location and pipeline deviation in subsea oil and gas pipeline inspection, proposing an underwater multi-source navigation and positioning method for this purpose. This method effectively solves the underwater positioning accuracy problem for long-endurance pipeline inspection robots, and by considering factors such as information time delay and sudden information changes, it significantly improves underwater positioning accuracy.

[0011] The objective of this invention is achieved as follows:

[0012] A seabed pipeline navigation and positioning system was constructed based on three sensors: Strapdown Inertial Navigation System (SINS), Strapdown Inertial Navigation System (DVL), and Ultra-Short Baseline (USBL). A centralized SINS / DVL / USBL integrated navigation information filtering model was established, using a Kalman filter to centrally process information from each navigation subsystem to obtain the globally optimal prediction and estimation of navigation error states. Then, separate state models and observation models for the SINS / DVL / USBL integrated navigation system were established. The accuracy of trend-following nonlinear state estimation was improved by fully utilizing delay measurement information through an improved DS-UKF algorithm. Finally, a weighted one-class SVM outlier detection algorithm was used to effectively solve the noise and jump point problems of USBL data, thereby improving underwater navigation accuracy.

[0013] The system consists of a strapdown inertial navigation system (SINS), a strapdown inertial navigation system (DVL), and an ultra-short baseline (USBL) system. The strapdown inertial navigation system is the core of the entire system. A Doppler velocimeter obtains velocity information, while GPS determines the position of the support vessel on the water surface. This, combined with the USBL underwater positioning system, enables high-precision positioning within the designated area. Before operation, a path planning trajectory is generated based on the pre-set coordinates of the pipeline's start and end points and the inspection requirements. The subsea pipeline inspection robot, carrying the integrated navigation and positioning system, moves to the seabed to conduct subsea pipeline inspection operations. All data is transmitted back to the surface-based support system in real time via a wireless communication bridge.

[0014] This information filtering model uses a Kalman filter to centrally process the information of each navigation subsystem in order to obtain the globally optimal prediction estimate of the navigation error state. The SINS serves as the fusion center and is fused with the velocity information independently output by the DVL and the position information independently output by the USBL using Kalman filtering, enabling the two external sensors to correct the solution error of the SINS in real time.

[0015] SINS utilizes an IMU (Inertial Measurement Unit) to measure the angular and linear motion of the carrier, obtaining angular and specific force increments. It then updates the attitude, velocity, and position to obtain the carrier's real-time navigation information. The 15-dimensional error state variables of the SINS subsystem are selected as the filtered state variables to be estimated.

[0016]

[0017] Where, φ=[φ N φ U φ E ] is the SINS attitude error quaternion; δV = [δVN δV U δV E These represent the eastward, northward, and celestial velocity errors, respectively; δP = [δP N δP U δP E [ represents latitude, longitude, and altitude errors; gyroscope drift error ε = [ε] x ε y ε z ] with accelerometer bias

[0018] The error state variables of the DVL subsystem are:

[0019] X DVL =[δcδv d δΔ] T (2)

[0020] Where δC represents the calibration coefficient error, δv d δΔ represents the offset error, and δΔ represents the flow deviation error.

[0021] The error state variables of the USBL subsystem are:

[0022] X USBL =[θ x θ y θ z δ α δ β δK μ ] T (3)

[0023] Where θ=[θ x θ y θ z ] T These represent the three mounting angle errors from USBL to the inertial navigation system; [δ α δ β ] T The following are the azimuth and elevation errors measured by USBL itself; δK μ This represents the distance measurement scale coefficient error of USBL.

[0024] The state equation for the centralized Kalman filter in the SINS / DVL / USBL integrated navigation system is established as follows:

[0025]

[0026] Among them, the lumped Kalman filter state variables are

[0027]

[0028] The matrix of the lumped Kalman filter system is G and W are the correlation noise matrix terms of a lumped Kalman filter.

[0029] The observation models of the integrated navigation filter include the SINS / DVL velocity observation model and the SINS / USBL position observation model. For the SINS / DVL integrated navigation system, DVL can provide the underwater vehicle's velocity information, so the velocity error between SINS and DVL is chosen as the system measurement. The specific measurement equations are as follows:

[0030]

[0031] The measurement noise is V. D =[η x η y ] T The system observation matrix is: H D =[-V D ×I 3×3 0 3×9 ] T .

[0032] USBL can provide the relative position information of the carrier. After coordinate system transformation, the difference between USBL and the position information of SINS is calculated and used as the system's measurement.

[0033]

[0034] Among them, V U =[δL U δλ U δh U ] T H represents the measurement noise of USBL. U The measurement matrix is ​​represented as: H U =[0 3× 4I 3×3 0 3×9 ].

[0035] Therefore, the lumped Kalman filter observation equation is:

[0036] Z = HX + V (8)

[0037] in, V is the system-related noise matrix, and H is the related measurement matrix.

[0038] The standard UKF algorithm does not account for measurement delays at each sampling time. When the system includes delayed measurements, both its state model and measurement model change. In subsea pipeline monitoring systems, certain key process variables in sensors may arrive at the fusion center after a random time delay during sampling. An improved DS-UKF algorithm is introduced to address the measurement delay problem by fully utilizing delayed measurement information to improve the accuracy of trend-following nonlinear state estimation.

[0039] Assuming a sensor's measurement is sampled at time s1 and arrives at the fusion center at time s after k time periods, the nonlinear system model including hysteresis measurements can be described as follows:

[0040] X k =f k-1 (X k-1 ,u k-1 )+W k-1 (9)

[0041]

[0042]

[0043] For nonlinear state estimation, the sampling point state dimension expansion method is usually used to handle measurement lag. The main idea is to expand the system state matrix by the delay period, redefine the system state space model, and then apply the lag data to the UKF information fusion system for state estimation.

[0044] Considering the above nonlinear system hysteresis measurement model, the model is extended in state dimension and redefined as follows:

[0045]

[0046]

[0047] Assuming the system is sampled at time s, the augmented state estimate at time k can be expressed as:

[0048]

[0049] Based on Bayesian optimal estimation theory, if the system has measurement lag, assuming that the measurement information at time s is delayed and reaches the fusion center at time k after N sampling intervals, the nonlinear filtering algorithm has the following conclusions:

[0050]

[0051]

[0052]

[0053] The UKF fusion algorithm includes:

[0054]

[0055] Within the UKF algorithm framework, the error covariance matrix is ​​obtained by solving the augmented state matrix through a time update process:

[0056]

[0057]

[0058] After introducing hysteresis measurements for dimensionality expansion within the UKF framework, it is essential to ensure that the system state and covariance of the hysteresis measurement information remain unchanged during the fusion period. To reduce algorithm complexity, a selection factor M is introduced, and the Schmidt-Kalman approach of the measurement update equation is adopted to address this issue. The improved UKF gain is shown in the following equation:

[0059]

[0060]

[0061] Here, M is the selection factor matrix, and the corresponding state estimate will only be updated after interacting with the identity matrix of the block matrix M. For the covariance matrix, the lower right sub-block retains the error covariance state at the initial sampling time s, and the upper left sub-block is the error covariance matrix for updating the state, which is the same as the original covariance value under the UKF algorithm framework.

[0062] To reduce classification errors caused by different feature vector concentration regions, this invention introduces a weighted mechanism to improve the algorithm. When the one-class SVM classifier classifies near the hyperplane or in overlapping regions, a support vector is selected as a representative pivot for each sample point, thus improving the model's classification accuracy. The distance from the unknown sample to the support vector pivot φ(x) in the feature space is calculated separately. + With φ(x) - The distance between the two distance values ​​is calculated, and the difference g(x) is obtained by subtracting the two distance values. This difference is then compared with a predefined threshold. If the difference is greater than the threshold, it indicates that the unknown sample is far from the optimal hyperplane, such as φ(x)1. In this case, the classic SVM classifier can be used directly to classify the unknown sample. If the difference is less than the threshold, it indicates that the unknown sample is close to the optimal hyperplane, such as φ(x)2. If the SVM classifier is used to classify the data, it is easy to lead to misclassification. In this case, an improved weighted classifier should be used for classification prediction.

[0063] The weighting factor is shown in the following formula:

[0064]

[0065] By comparing the distances from the support vector pivots to the sample points, different weighting factors are selected and fused into the algorithm to obtain an improved one-clss SVM classifier, as shown in the following equation:

[0066]

[0067] By injecting 10% outlier samples into the slant range variation data output by USBL, an imbalanced sample set is provided as a classification premise. An improved weighted one-class SVM classifier is then introduced to verify the algorithm's performance. The optimization process for the objective function involves establishing an optimal decision hyperplane that maximizes the distance between the two classes of samples closest to the hyperplane on either side, thus providing good generalization ability. During training, a hyperplane is randomly generated and continuously moved. Bayesian optimization is used to search for the optimal relationship between the parameters box and sigma of the support vector machine, ensuring that each observation sample can be represented by best and accept as it moves along the hyperplane. The classification results of the weighted one-class SVM classifier on the training set are shown. The training set consists of 353 positive samples (denoted as "+1") and 47 outlier samples (denoted as "-1"). Through training, a suitable optimal decision plane is found between the two classes of samples. The weighted one-class SVM classification results on the test set, which consists of 547 positive samples and 53 outlier samples, demonstrate how the weighting mechanism further compresses the envelope space of positive samples. This not only enables the model to classify correctly but also improves the classification margin, keeping outliers as far away from the hyperplane as possible and minimizing their interference with reliable samples. Analysis of the classification results for both classes shows that the improved weighted one-class SVM strengthens the optimization of the hyperplane near the decision boundary, further shortening the classification region for reliable values ​​and significantly reducing the risk of misclassification or false detection of outliers.

[0068] Compared with the prior art, the present invention has the following advantages:

[0069] (1) This invention has good versatility and can be widely applied to the navigation system of various underwater robots for detecting submarine oil and gas pipelines and submarine cables.

[0070] (2) This invention utilizes the State Delay Kalman Filter (DS-UKF) algorithm to effectively solve the USBL information delay problem.

[0071] (3) This invention utilizes an outlier detection algorithm based on weighted one-class SVM to effectively solve the problem of information jump points caused by noise in the information. Attached Figure Description

[0072] Figure 1 Flowchart of a multi-source localization method based on submarine oil and gas pipeline detection;

[0073] Figure 2 Schematic diagram of a combined navigation system based on submarine pipeline detection;

[0074] Figure 3 An INS / DVL / USBL integrated navigation system based on subsea pipeline detection;

[0075] Figure 4 Schematic diagram of the improved weighted classifier;

[0076] Figure 5 The optimization process of weighted one-class SVM;

[0077] Figure 6 Weighted one-class SVM training results;

[0078] Figure 7 The classification results of the weighted one-class SVM on the test set. Detailed Implementation

[0079] This invention addresses the shortcomings of current active wave compensation systems by proposing an active wave compensation analysis method for hybrid boarding systems based on multi-task motion planning. This method effectively solves the redundancy problem of hybrid mechanisms while maximizing the optimization of spatial motion allocation. By utilizing the synergistic effect between the tandem gangways and parallel platforms in their motion trajectories, active wave compensation is achieved, providing crucial control parameters for the active wave compensation system.

[0080] The present invention will now be described in more detail with reference to the accompanying drawings:

[0081] Implementation 1: As attached Figure 1 As shown, this invention establishes a multi-source localization method based on the detection of subsea oil and gas pipelines. The specific steps are as follows:

[0082] Step 1: Construct a submarine pipeline navigation and positioning system composed of SINS, DVL, and USBL;

[0083] Step 2: Establish a centralized SINS / DVL / USBL integrated navigation information filtering model, and use a Kalman filter to centrally process the information of each navigation subsystem in order to obtain the globally optimal prediction estimate of the navigation error state;

[0084] Step 3: Establish a state model for the SINS / DVL / USBL integrated navigation system;

[0085] Step 4: Establish the SINS / DVL / USBL integrated navigation observation model;

[0086] Step 5: Improve the DS-UKF algorithm by making full use of delay measurement information to improve the accuracy of the following nonlinear state estimation;

[0087] Step 6: The outlier detection algorithm of weighted one-class SVM effectively solves the noise and jump point problems of USBL data, thereby improving the accuracy of underwater navigation.

[0088] Implementation 2: As attached Figure 2 As shown, this invention utilizes an unmanned surface support vessel (AUV) to perform inspection tasks above seabed oil and gas pipelines. A combined navigation system was designed, consisting of a strapdown inertial navigation system (SINS), a strapdown inertial navigation system (DVL), and an ultra-short baseline (USBL) system. The strapdown inertial navigation system is the core of the entire system; a Doppler velocimeter obtains velocity information, while GPS determines the support vessel's position on the surface, working in conjunction with the USBL underwater positioning system for high-precision positioning within the area. Before operation, a path planning trajectory is generated based on the pre-set coordinates of the pipeline's start and end points and inspection requirements. The subsea pipeline inspection robot, carrying the combined navigation and positioning system, moves to the seabed to perform subsea pipeline inspection operations. All data is transmitted back to the surface-based support system in real time via a wireless communication bridge.

[0089] Implementation 3: In conjunction with the appendix Figure 3 A centralized information filtering model for SINS / DVL / USBL integrated navigation was designed. A Kalman filter is used to centrally process the information from each navigation subsystem to obtain the globally optimal prediction estimate of the navigation error state. The block diagram of the integrated navigation system is attached. Figure 3 As shown, SINS acts as the fusion center, fusing it with the velocity information independently output by DVL and the position information independently output by USBL using Kalman filtering, enabling the two external sensors to correct the SINS's calculation errors in real time.

[0090] Implementation 4: State Model of SINS / DVL / USBL Integrated Navigation System

[0091] SINS utilizes an IMU (Inertial Measurement Unit) to measure the angular and linear motion of the carrier, obtaining angular and specific force increments. It then updates the attitude, velocity, and position to obtain the carrier's real-time navigation information. The 15-dimensional error state variables of the SINS subsystem are selected as the filtered state variables to be estimated.

[0092]

[0093] Where, φ=[φ N φ U φ E ] is the SINS attitude error quaternion; δV = [δV NδV U δV E These represent the eastward, northward, and celestial velocity errors, respectively; δP = [δP N δP U δP E [ represents latitude, longitude, and altitude errors; gyroscope drift error ε = [ε] x ε y ε z ] with accelerometer bias

[0094] The error state variables of the DVL subsystem are:

[0095] X DVL =[δcδv d δΔ] T (2)

[0096] Where δC represents the calibration coefficient error, δv d δΔ represents the offset error, and δΔ represents the flow deviation error.

[0097] The error state variables of the USBL subsystem are:

[0098] X USBL =[θ x θ y θ z δ α δ β δK μ ] T (3)

[0099] Where θ=[θ x θ y θ z ] T These represent the three mounting angle errors from USBL to the inertial navigation system; [δ α δ β ] T The following are the azimuth and elevation errors measured by USBL itself; δK μ This represents the distance measurement scale coefficient error of USBL.

[0100] The state equation for the centralized Kalman filter in the SINS / DVL / USBL integrated navigation system is established as follows:

[0101]

[0102] Among them, the lumped Kalman filter state variables are

[0103]

[0104] The matrix of the lumped Kalman filter system is G and W are the correlation noise matrix terms of a lumped Kalman filter.

[0105] Implementation 5: SINS / DVL / USBL integrated navigation observation model.

[0106] The observation models of the integrated navigation filter include the SINS / DVL velocity observation model and the SINS / USBL position observation model. For the SINS / DVL integrated navigation system, DVL can provide the underwater vehicle's velocity information, so the velocity error between SINS and DVL is chosen as the system measurement. The specific measurement equations are as follows:

[0107]

[0108] The measurement noise is V. D =[η x η y ] T The system observation matrix is: H D =[-V D ×I 3×3 0 3×9 ] T .

[0109] USBL can provide the relative position information of the carrier. After coordinate system transformation, the difference between USBL and the position information of SINS is calculated and used as the system's measurement.

[0110]

[0111] Among them, V U =[δL U δλ U δh U ] T H represents the measurement noise of USBL. U The measurement matrix is ​​represented as: H U =[0 3× 4I 3×3 0 3×9 ].

[0112] Therefore, the lumped Kalman filter observation equation is:

[0113] Z = HX + V (8)

[0114] in, V is the system-related noise matrix, and H is the related measurement matrix.

[0115] Implementation 6: State Delay Kalman Filter (DS-UKF) Algorithm

[0116] The standard UKF algorithm does not account for measurement delays at each sampling time. When the system includes delayed measurements, both its state model and measurement model change. In subsea pipeline monitoring systems, certain key process variables in sensors may arrive at the fusion center after a random time delay during sampling. An improved DS-UKF algorithm is introduced to address the measurement delay problem by fully utilizing delayed measurement information to improve the accuracy of trend-following nonlinear state estimation.

[0117] Assuming a sensor's measurement is sampled at time s1 and arrives at the fusion center at time s after k time periods, the nonlinear system model including hysteresis measurements can be described as follows:

[0118] X k =f k-1 (X k-1 ,u k-1 )+W k-1 (9)

[0119]

[0120]

[0121] For nonlinear state estimation, the sampling point state dimension expansion method is usually used to handle measurement lag. The main idea is to expand the system state matrix by the delay period, redefine the system state space model, and then apply the lag data to the UKF information fusion system for state estimation.

[0122] Considering the above nonlinear system hysteresis measurement model, the model is extended in state dimension and redefined as follows:

[0123]

[0124]

[0125] Assuming the system is sampled at time s, the augmented state estimate at time k can be expressed as:

[0126]

[0127] Based on Bayesian optimal estimation theory, if the system has measurement lag, assuming that the measurement information at time s is delayed and reaches the fusion center at time k after N sampling intervals, the nonlinear filtering algorithm has the following conclusions:

[0128]

[0129]

[0130]

[0131] The UKF fusion algorithm includes:

[0132]

[0133] Within the UKF algorithm framework, the error covariance matrix is ​​obtained by solving the augmented state matrix through a time update process:

[0134]

[0135]

[0136] After introducing hysteresis measurements for dimensionality expansion within the UKF framework, it is essential to ensure that the system state and covariance of the hysteresis measurement information remain unchanged during the fusion period. To reduce algorithm complexity, a selection factor M is introduced, and the Schmidt-Kalman approach of the measurement update equation is adopted to address this issue. The improved UKF gain is shown in the following equation:

[0137]

[0138]

[0139] Here, M is the selection factor matrix, and the corresponding state estimate will only be updated after interacting with the identity matrix of the block matrix M. For the covariance matrix, the lower right sub-block retains the error covariance state at the initial sampling time s, and the upper left sub-block is the error covariance matrix for updating the state, which is the same as the original covariance value under the UKF algorithm framework.

[0140] Implementation 7: Outlier Detection Algorithm Based on Weighted One-Class SVM

[0141] To reduce classification errors caused by different feature vector concentration regions, this invention introduces a weighted mechanism to improve the algorithm. When the one-class SVM classifier classifies near the hyperplane or in overlapping regions, a support vector is selected as a representative fulcrum for each sample point, thereby improving the model's classification accuracy. (See attached diagram.) Figure 4 As shown, the distance from the unknown sample to the support vector pivot φ(x) in the feature space is calculated respectively. + With φ(x) -The distance between the two distance values ​​is calculated, and the difference g(x) is obtained by subtracting the two distance values. This difference is then compared with a predefined threshold. If the difference is greater than the threshold, it indicates that the unknown sample is far from the optimal hyperplane, such as φ(x)1. In this case, the classic SVM classifier can be used directly to classify the unknown sample. If the difference is less than the threshold, it indicates that the unknown sample is close to the optimal hyperplane, such as φ(x)2. If the SVM classifier is used to classify the data, it is easy to lead to misclassification. In this case, an improved weighted classifier should be used for classification prediction.

[0142] The weighting factor is shown in the following formula:

[0143]

[0144] By comparing the distances from the support vector pivots to the sample points, different weighting factors are selected and fused into the algorithm to obtain an improved one-clss SVM classifier, as shown in the following equation:

[0145]

[0146] By injecting 10% outlier samples into the slope range variation data output by USBL, an imbalanced sample set is provided as a classification premise. An improved weighted one-class SVM classifier is then introduced to verify the algorithm's performance. (See attached...) Figure 5 The optimization process for the objective function involves establishing an optimal decision hyperplane that maximizes the distance between the two classes of samples closest to the hyperplane on either side, thus providing good generalization ability. During training, the system randomly generates a hyperplane and continuously moves it. Bayesian optimization is used to search for the optimal relationship between the parameters `box` and `sigma` of the support vector machine, ensuring that each observed sample can be represented by `best` and `accept` as it moves along the hyperplane. (Combined with the appendix...) Figure 6 This demonstrates the classification results of a weighted one-class SVM classifier on the training set, which consists of 353 positive samples (denoted as "+1") and 47 outlier samples (denoted as "-1"). The optimal decision plane between the two classes is found through training. (See attached diagram.) Figure 7The classification results of the weighted one-class SVM on the test set are presented. The test set consists of 547 positive samples and 53 outlier samples. Due to the introduction of the weighting mechanism, the space of the envelope of positive samples is further compressed, enabling the model not only to classify correctly but also to improve the classification margin, keeping outliers as far away from the hyperplane as possible and minimizing the interference of outliers on reliable samples. By analyzing the classification results of the two classes, it can be seen that the improved weighted one-class SVM strengthens the optimization of the hyperplane near the decision boundary, further shortens the classification region of reliable values, and greatly reduces the risk of misclassification or false detection of outliers.

[0147] This invention addresses the need for accurate fault location and subsea pipeline position deviation in subsea oil and gas pipeline inspection, designing an underwater multi-source navigation and positioning method for this purpose. The invention constructs a subsea pipeline navigation and positioning system composed of three sensors: Strapdown Inertial Navigation System (SINS), DVL (Dual Vessel Loading) system, and Ultra-Short Baseline (USBL). Based on this, a centralized SINS / DVL / USBL integrated navigation information filtering model is established, using a Kalman filter to centrally process information from each navigation subsystem to obtain a globally optimal prediction and estimate of the navigation error state. Then, separate state models and observation models for the SINS / DVL / USBL integrated navigation system are established. The DS-UKF algorithm is improved by fully utilizing delay measurement information to enhance the accuracy of the following nonlinear state estimation. Finally, a weighted one-class SVM outlier detection algorithm effectively solves the noise and jump point problems of USBL data, thereby improving underwater navigation accuracy. This method effectively solves the problem of underwater positioning accuracy for long-endurance pipeline inspection robots, and by taking into account factors such as information time delay and information abrupt changes, it greatly improves underwater positioning accuracy.

Claims

1. An underwater multi-source navigation and positioning method based on the detection of subsea oil and gas pipelines, characterized by: A seabed pipeline navigation and positioning system was constructed based on three sensors: a strapdown inertial navigation system (SINS), a Doppler log (DVL), and an ultra-short baseline (USBL) system. A centralized SINS / DVL / USBL integrated navigation information filtering model was then established, using a Kalman filter to centrally process information from each navigation subsystem to obtain the globally optimal prediction and estimation of navigation error states. Next, separate state models and observation models for the SINS / DVL / USBL integrated navigation system were established. The DS-UKF algorithm was improved to fully utilize delay measurement information. Finally, a weighted one-class SVM outlier detection algorithm was employed. The UKF algorithm is as follows: assuming the measurement value of a certain sensor is at time... Sampling was performed, and after Each time period in If the time reaches the fusion center, the nonlinear system model including the hysteresis measurement can be described as: (9) (10) (11) Considering the above nonlinear system hysteresis measurement model, the model is extended in state dimension and redefined as follows: (12) (13) Assuming the system is sampled at time s, the augmented state estimate at time k can be expressed as: (14) Based on Bayesian optimal estimation theory, if the system has measurement lag, assume that... There is a delay in the time measurement information, after which... Each sampling interval is Upon reaching the fusion center at a given time, the nonlinear filtering algorithm yields the following conclusions: (15) (16) (17) The UKF fusion algorithm includes: (18) Within the UKF algorithm framework, the error covariance matrix is ​​obtained by solving the augmented state matrix through a time update process: (19) (20) After introducing hysteresis measurements for dimension expansion within the UKF framework, and introducing a selection factor M, the improved UKF gain is as follows: (21) (22) Where M is the selection factor matrix, and the corresponding state estimate will only be updated after interacting with the identity matrix of the block matrix M; for the covariance matrix, the lower right sub-block maintains the error covariance state at the initial sampling time s, and the upper left sub-block is the error covariance matrix for updating the state, which is the same as the original covariance value under the UKF algorithm framework.

2. The underwater multi-source navigation and positioning method based on subsea oil and gas pipeline detection according to claim 1, characterized in that: The submarine pipeline navigation and positioning system consists of a strapdown inertial navigation system (SINS), a Doppler log (DVL), and an ultra-short baseline (USBL). The strapdown inertial navigation system is the core of the entire system. The Doppler velocimeter obtains velocity information, while the surface portion is... The location of the support vessel is determined, and an ultra-short baseline underwater positioning system is used for high-precision positioning within the area. Before the operation, a path planning trajectory is generated based on the pre-set coordinates of the pipeline start and end points and the inspection requirements. The subsea pipeline inspection robot carries the integrated navigation and positioning system to the seabed to carry out the subsea pipeline inspection operation. All data is transmitted back to the surface shore-based support system in real time via a wireless communication bridge.

3. The underwater multi-source navigation and positioning method based on subsea oil and gas pipeline detection according to claim 1, characterized in that: The centralized The integrated navigation information filtering model uses a Kalman filter to centrally process the information of each navigation subsystem in order to obtain the globally optimal prediction estimate of the navigation error state. The SINS, as the fusion center, is fused with the velocity information output independently by the DVL and the position information output independently by the USBL using Kalman filtering, so that the two external sensors can correct the solution error of the SINS in real time.

4. The underwater multi-source navigation and positioning method based on subsea oil and gas pipeline detection according to claim 1, characterized in that: SINS uses an IMU (Inertial Measurement Unit) to measure the angular and linear motion of the carrier, obtaining angular and specific force increments. It then updates the attitude, velocity, and position to obtain the carrier's real-time navigation information. The 15-dimensional error state variables of the SINS subsystem are selected as the filtered state variables to be estimated. (1) in, It is the SINS attitude error quaternion; These are the speed errors in the east, north, and sky directions, respectively. These are errors in latitude, longitude, and altitude; gyroscope drift error. biased with accelerometer ; The error state variables of the DVL subsystem are: (2) in, Indicates the scale coefficient error. Indicates offset error. Indicates the eccentricity error; The error state variables of the USBL subsystem are: (3) in These represent the three installation angle errors from USBL to the inertial navigation system; The azimuth and elevation errors measured by USBL itself are, in order. This refers to the ranging scale coefficient error of USBL; The state equation for the centralized Kalman filter in the SINS / DVL / USBL integrated navigation system is established as follows: (4) Among them, the lumped Kalman filter state variables are (5) The matrix of the lumped Kalman filter system is G and W are the correlation noise matrix terms of the lumped Kalman filter.

5. The underwater multi-source navigation and positioning method based on subsea oil and gas pipeline detection according to claim 1, characterized in that: The observation model of the integrated navigation filter includes Velocity observation model and Position observation model; For the SINS / DVL integrated navigation system, DVL can provide the speed information of the underwater vehicle, so the speed error between SINS and DVL is selected as the system measurement. The specific measurement equation is as follows: (6) The measurement noise is The system observation matrix is: ; USBL can provide the relative position information of the carrier. After coordinate system transformation, the difference between USBL and the position information of SINS is calculated and used as the system's measurement. (7) in, Indicates the measurement noise of USBL; The measurement matrix is ​​represented as follows: ; Therefore, the lumped Kalman filter observation equation is: (8) in, , The system's correlated noise matrix, This is the relevant measurement matrix.

6. The underwater multi-source navigation and positioning method based on subsea oil and gas pipeline detection according to claim 1, characterized in that: The outlier detection algorithm based on weighted one-class SVM is as follows: A weighting mechanism is introduced to improve the algorithm. When the one-class SVM classifier classifies near the hyperplane or in the overlapping region, a support vector is selected as the representative fulcrum for each sample point. Calculate the distance from the unknown sample to the support vector pivot in the feature space. and The distance is calculated by subtracting the two distance values ​​to obtain the distance difference. The distance difference is compared with a predefined threshold. If the difference is greater than the threshold, the classic SVM classifier is used to classify the unknown sample directly. If the difference is less than the threshold, an improved weighted classifier is used for classification prediction. The weighting factor is shown in the following formula: (23) By comparing the distances from the support vector pivots to the sample points, different weighting factors are selected and fused into the algorithm to obtain an improved one-clss SVM classifier, as shown in the following equation: (24)。

Images