DVL speed measurement model training method based on artificial bee colony optimization relevance vector machine

Abstract

The present invention provides a DVL velocity measurement model training method based on artificial bee colony optimization relevance vector machine, comprising: building an AUV test platform, obtaining three-axis gyroscope and three-axis accelerometer data of the inertial navigation system, three-axis DVL velocity information, latitude and longitude output by GNSS, and depth training samples provided by a depth meter; using the heading, pitch, roll, heading change rate and three-axis DVL velocity information output by the inertial navigation system as an RVM model input set, and using the three-axis inertial navigation system velocity information as an RVM model output set; initializing RVM kernel function parameters, mapping standardized data to feature space; calculating the mean and variance of the posterior distribution; updating and calculating hyperparameters and noise variance to obtain a sparse model; using an artificial bee colony algorithm to optimize the hybrid kernel function parameters in each RVM, wherein the GNSS and depth meter are used to constrain the objective function in the artificial bee colony algorithm. This method reduces the positioning error of SINS / DVL integrated navigation from 3‰ of the range to 1.5‰ of the range.

Description

Technical Field

[0001] The present invention relates to the technical field of inertial navigation and combined navigation, and in particular to a DVL speed measurement model training method based on an artificial bee colony optimized relevance vector machine. Background Art

[0002] Underwater navigation technology provides accurate position, velocity, and attitude information for autonomous underwater vehicles (AUVs), crucial for their ability to accurately reach their intended destination and return safely. AUV navigation systems primarily rely on a strapdown inertial navigation system (SINS), supplemented by other navigation systems or sensors to achieve fully autonomous navigation. Currently, the most popular AUV underwater navigation method is a combination of SINS and DVL.

[0003] In SINS / DVL integrated navigation, the DVL's velocity measurement accuracy is a key factor affecting the system's navigation accuracy. Three main factors affect DVL velocity measurement accuracy: the installation error angle between the SINS and DVL; the DVL's own velocity measurement error (primarily consisting of scale factor error and constant error); and the DVL velocity measurement error introduced by attitude angle dynamics. Traditional DVL calibration algorithms can be categorized into two main types: Kalman filter estimation and least squares estimation. Kalman filter estimation requires a small installation error angle and a priori error model, while the vehicle's maneuvering state directly affects filter convergence. Traditional least squares estimation methods typically utilize SVD matrix decomposition for parameter estimation, but this assumes a linear DVL error model and neglects modeling of velocity measurement errors introduced by dynamic attitude angles. This is particularly true under large maneuvers, where traditional linear models cannot accurately describe the DVL's error characteristics, a significant factor affecting DVL model training accuracy.

[0004] With the development of machine learning theory and technology, a Bayesian-based relevance vector machine (RVM) has been developed based on the support vector machine (SVM). This RVM not only shares the advantages of the SVM model, but its kernel function is not constrained by the Mercer condition, resulting in better prediction results compared to the SVM model. Targeting complex underwater applications, this paper introduces a RVM to implement regression estimation of the DVL error model. The artificial bee colony algorithm is used to optimize the uncertain parameters in the RVM model, while also incorporating GNSS and depth sensor information as constraints to improve the accuracy and generalization of the DVL error model. Summary of the Invention

[0005] To address the technical issues raised in the previous article, the present invention provides a DVL velocity measurement model training method based on an artificial bee colony optimized relevance vector machine. This method has excellent approximation capabilities for complex nonlinear functions and aims to address the environmental adaptability and poor generalization capabilities of traditional model training methods. This method can significantly improve the positioning accuracy of underwater integrated navigation and provides an effective means for compensating for DVL velocity measurement errors in underwater integrated navigation systems under complex maneuvering conditions.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A DVL speed measurement model training method based on artificial bee colony optimization relevance vector machine includes the following steps:

[0008] Step 1: Build an AUV test platform to obtain the raw data of the inertial navigation system, the three-axis velocity information output by the DVL, the longitude and latitude output by the GNSS, and the depth training samples provided by the depth meter. The raw data of the inertial navigation system includes the three-axis accelerometer and three-axis gyroscope information;

[0009] Step 2: Initially align the acquired inertial navigation system data training samples. After the initial alignment is completed, perform real-time attitude and velocity calculations to obtain the vehicle's heading, pitch, roll, heading change rate dHeading, and the three-axis vehicle system velocity information V_x, V_y, and V_z;

[0010] Step 3: Get the three-axis velocity information DVL_x, DVL_y and DVL_z of the DVL output carrier, and use DVL_x, DVL_y, DVL_z, Heading, Pitch, Roll and dHeading as the RVM model input set x, and V_x, V_y and V_z as the output set y to form the data training set Where N is the number of sampling points, which are used as training samples for the three RVM models;

[0011] Step 4: Initialize the hybrid kernel function K(x,x) in each RVM i ) in the parameters;

[0012] Step 5: Initialize hyperparameters α and noise variance σ 2 , using kernel function to map the normalized data into feature space;

[0013] Step 6: Calculate the mean μ and variance Σ of the posterior distribution;

[0014] Step 7: Update and calculate the hyperparameter α and noise variance σ 2 , remove the infinite elements in α, corresponding to ω i is 0, ω iis the i-th element of the model weight ω, and we get non-zero ω i The corresponding vector is the relevant vector, thus obtaining a sparse model;

[0015] In step eight, the hybrid kernel function parameters in each RVM are optimized using the artificial bee colony algorithm, wherein the objective function in the artificial bee colony algorithm is adaptively adjusted using GNSS and depth meter to obtain the optimal kernel function parameters.

[0016] Furthermore, in the step 1: under the AUV test platform, the carrier is initially aligned near the water surface, then dives underwater and maintains a uniform speed and straight navigation, the navigation distance is not less than 1 km, and finally surfaces to the water surface; real-time collection of inertial navigation system and DVL, GNSS and depth meter data, wherein the inertial navigation system raw data update rate is 100Hz, and the DVL, GNSS and depth meter data update rate is 1Hz, to ensure that no data frames are lost during the data collection process.

[0017] Furthermore, in step 2: considering that the observability of the heading is poor under non-maneuvering conditions, a SINS / GNSS combined initial alignment is performed near the water surface to improve the observability of the heading; in addition, position and velocity information are used as observation information for the Kalman filter to accelerate the initial alignment process;

[0018] Coning error will affect the accuracy of attitude solution in underwater application environment. In order to further improve the attitude solution accuracy of inertial navigation system, the three-sample rotating vector algorithm is applied to the attitude solution of strapdown inertial navigation.

[0019] Furthermore, in step three, the speed information output by the inertial navigation system is expressed in the geographic coordinate system, while the speed output by the DVL is in the carrier coordinate system. In order to unify the two coordinate systems, the speed information output by the inertial navigation system in the carrier system is converted into the speed in the carrier system through the attitude conversion matrix.

[0020] Furthermore, in step 4, the kernel function selection in the RVM model directly affects the learning and generalization capabilities of the model training. By utilizing the respective advantages of the local kernel function and the global kernel function, a weighted combination is performed to form the following hybrid kernel function:

[0021]

[0022] In the formula, a represents the weight of the Gaussian kernel function, b and c represent the width of the corresponding kernel function, and x i Represents the i-th input value of the model sample.

[0023] Furthermore, in step 5: given a training sample set is Among them, x i∈R d ,y i ∈R; where x i represents the input sample, y i represents the target sample, d represents the dimension of the input sample; the regression model of the target sample can be expressed as:

[0024] y i =f(x i ;ω)+ε n

[0025] Where, ε n is the additional noise and satisfies ε n ∈N(0,σ 2 ), f(x) is defined as follows:

[0026]

[0027] Where N is the number of samples, ω=[ω0,ω1,...,ω N ] represents the weight, K(x,x i ) is the kernel function, Φ(x) is an N×(N+1) order matrix, which can be expressed as

[0028]

[0029] Since the target sample y i are independent, so the likelihood function of the training sample set can be expressed as:

[0030]

[0031] In the formula, y=[y1, y2,...,y N ] is the training target sample.

[0032] Furthermore, in step 6, for the multi-weight regression model, the maximum likelihood method will get overfitting results. In the correlation vector machine regression model, the maximum likelihood method is used under the Bayesian framework to train the model weights, so that the model has strong generalization. The prior distribution of the parameters is

[0033]

[0034] In the formula, α = [α1, α2,..., α N ] is a hyperparameter;

[0035] Based on the prior distribution and likelihood distribution, the posterior distribution of the weight can be obtained by Bayesian reasoning, and its posterior distribution is

[0036]

[0037] The posterior mean and variance are

[0038]

[0039] In the formula, A=diag(α0,α1,...,α N ).

[0040] Furthermore, in step 7: the output likelihood distribution can be obtained by performing marginal integration on the weights, that is,

[0041] P(y|α,σ 2 )=∫P(y|ω,σ 2 )P(ω|α)dω

[0042] Thus, the marginal likelihood distribution of the hyperparameters can be obtained, that is,

[0043] P(y|α,σ 2 )~N(0,C)

[0044] Where C = σ 2 I+ΦA -1 Φ T , I is the identity matrix;

[0045] According to MacKay's method, we can get

[0046] Where μ i is the i-th element of the mean vector μ, and γ is defined in the MacKay method. i =1-α i N ii , where N ii is the i-th element on the diagonal of the variance Σ;

[0047]

[0048] The learning process of the RVM is to continuously iterate and update and (σ 2 ) new , and then update the statistics μ and Σ until the set convergence conditions are met; in the actual calculation process, the posterior distribution of many weights tends to zero. In the regression model of the correlation vector machine, the non-zero weights represent the prototype samples in the data, which are called "correlation vectors"; for a new set of inputs x * , and its corresponding output y * The prediction distribution of in

[0049]

[0050] Where μ * For expectations, is the variance, It is the maximum likelihood estimate of P(ω|α), which determines the optimal value of the weight.

[0051] Furthermore, in step eight, the latitude and longitude information provided by GNSS and the depth information provided by the depth meter are introduced to constrain the objective function of the artificial bee colony algorithm. The improved artificial bee colony objective function can be expressed as

[0052]

[0053] in

[0054]

[0055] Where RVM1:fitness, RVM2:fitness and RVM3:fitness represent the artificial bee colony search objective functions in the eastward velocity, northward velocity and celestial velocity training models respectively, and v e 、v n and v u They represent the velocity of SINS output in the geographic coordinate system, v e* 、v n* and v u* where m1, m2 and m3 are the weight coefficients of the objective functions in the RVM1, RVM2 and RVM3 models, respectively. The weight coefficients are adaptively adjusted according to the proportional relationship between the velocity information provided by SINS and the equivalent velocity information output by external sensors, namely, GNSS and depth sensor.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] The present invention proposes a DVL velocity measurement model training method based on an artificial bee colony optimized relevance vector machine. This method has excellent approximation capabilities for complex nonlinear functions and aims to address the environmental adaptability and poor generalization capabilities of traditional model training methods. This method can significantly improve the positioning accuracy of underwater integrated navigation and provides an effective means for compensating for DVL velocity measurement errors in underwater integrated navigation systems under complex maneuvering conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of a DVL speed measurement model training method based on artificial bee colony optimization relevance vector machine provided by an embodiment of the present invention;

[0059] Figure 2 Schematic diagram of a DVL speed measurement model training method based on artificial bee colony optimization relevance vector machine provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0061] like Figure 1-2 As shown, a DVL speed measurement model training method based on artificial bee colony optimization relevance vector machine includes the following steps:

[0062] Step 1: Build an AUV test platform to obtain the raw data of the inertial navigation system, the three-axis velocity information output by the DVL, the longitude and latitude output by the GNSS, and the depth training samples provided by the depth meter. The raw data of the inertial navigation system includes the three-axis accelerometer and three-axis gyroscope information;

[0063] Step 2: Initially align the acquired inertial navigation system data training samples. After the initial alignment is completed, perform real-time attitude and velocity calculations to obtain the vehicle's heading, pitch, roll, heading change rate dHeading, and the three-axis vehicle system velocity information V_x, V_y, and V_z;

[0064] Step 3: Get the three-axis velocity information DVL_x, DVL_y and DVL_z of the DVL output carrier, and use DVL_x, DVL_y, DVL_z, Heading, Pitch, Roll and dHeading as the RVM model input set x, and V_x, V_y and V_z as the output set y to form the data training set Where N is the number of sampling points, which are used as training samples for the three RVM models;

[0065] Step 4: Initialize the hybrid kernel function K(x,x) in each RVM i ) in the parameters;

[0066] Step 5: Initialize hyperparameters α and noise variance σ 2 , using kernel function to map the normalized data into feature space;

[0067] Step 6: Calculate the mean μ and variance Σ of the posterior distribution;

[0068] Step 7: Update and calculate the hyperparameter α and noise variance σ 2 , remove the infinite elements in α, corresponding to ω i is 0, ω i is the i-th element of the model weight ω, and we get non-zero ω iThe corresponding vector is the relevant vector, thus obtaining a sparse model;

[0069] In step eight, the hybrid kernel function parameters in each RVM are optimized using the artificial bee colony algorithm, wherein the objective function in the artificial bee colony algorithm is adaptively adjusted using GNSS and depth meter to obtain the optimal kernel function parameters.

[0070] The specific embodiments are as follows:

[0071] Step 1: Build an AUV test platform to obtain the raw data of the inertial navigation system, the three-axis velocity information output by the DVL, the longitude and latitude output by the GNSS, and the depth training samples provided by the depth meter. The raw data of the inertial navigation system includes the three-axis accelerometer and three-axis gyroscope information.

[0072] In the step 1, under the AUV test platform, the vehicle stays near the water surface for 15 minutes for initial alignment, then dives underwater and maintains a uniform speed and straight navigation, with a navigation distance of not less than 1 km, and finally surfaces to the water surface; real-time collection of inertial navigation system and DVL, GNSS and depth meter data, wherein the inertial navigation system raw data update rate is 100Hz, and the DVL, GNSS and depth meter data update rate is 1Hz, to ensure that no data frames are lost during the data collection process.

[0073] Step 2: Perform initial alignment on the acquired inertial navigation system data training samples. After the initial alignment is completed, perform real-time attitude and velocity calculations to obtain the vehicle's heading, pitch, roll, heading change rate dHeading, and the three-axis vehicle system velocity information V_x, V_y, and V_z.

[0074] In step 2, considering the poor observability of the heading under non-maneuvering conditions, a SINS / GNSS combined initial alignment is performed near the water surface to improve the observability of the heading. In addition, position and velocity information are used as observation information for the Kalman filter to accelerate the initial alignment process.

[0075] Coning error will affect the accuracy of attitude solution in underwater application environment. In order to further improve the attitude solution accuracy of inertial navigation system, the three-sample rotating vector algorithm is applied to the attitude solution of strapdown inertial navigation.

[0076] Step 3: Get the three-axis velocity information DVL_x, DVL_y and DVL_z of the DVL output carrier, and use DVL_x, DVL_y, DVL_z, Heading, Pitch, Roll and dHeading as the RVM model input set x, and V_x, V_y and V_z as the output set y to form the data training set Where N is the number of sampling points, which are used as training samples for the three RVM models.

[0077] In step 3, the velocity information output by the inertial navigation system is expressed in the geographic coordinate system, while the velocity output by the DVL is in the carrier coordinate system. In order to unify the two coordinate systems, the velocity information output by the inertial navigation system in the carrier coordinate system is converted into the velocity in the carrier coordinate system through the attitude conversion matrix.

[0078] Step 4: Initialize the hybrid kernel function K(x,x) in each RVM i ) in the .

[0079] In the above step 4: the choice of kernel function in the RVM model directly affects the learning and generalization capabilities of the model training. Currently, the common kernel functions are mainly divided into local kernel functions and global kernel functions. Local kernel functions have strong learning and training capabilities, but weak generalization capabilities, such as the Gaussian kernel function; global kernel functions have strong generalization capabilities, but weak learning and training capabilities, such as the radial kernel function. Therefore, the advantages of the two kernel functions, local kernel functions and global kernel functions, are used to perform a weighted combination to form the following hybrid kernel function:

[0080]

[0081] In the formula, a represents the weight of the Gaussian kernel function, b and c represent the width of the corresponding kernel function, and x i Represents the i-th input value of the model sample.

[0082] Step 5: Initialize hyperparameters α and noise variance σ 2 , using kernel function to map the normalized data into feature space.

[0083] In the step 5: given a training sample set is Among them, x i ∈R d ,y i ∈R; where x i represents the input sample, y i represents the target sample, d represents the dimension of the input sample; the regression model of the target sample can be expressed as:

[0084] y i =f(x i ;ω)+ε n

[0085] Where, ε n is the additional noise and satisfies ε n ∈N(0,σ 2 ), f(x) is defined as follows:

[0086]

[0087] Where N is the number of samples, ω=[ω0,ω1,...,ω N ] represents the weight, K(x,x i ) is the kernel function, Φ(x) is an N×(N+1) order matrix, which can be expressed as

[0088]

[0089] Since the target sample y i are independent, so the likelihood function of the training sample set can be expressed as:

[0090]

[0091] In the formula, y=[y1, y2,...,y N ] is the training target sample.

[0092] Step 6: Calculate the mean μ and variance Σ of the posterior distribution.

[0093] In step 6, for the multi-weight regression model, the maximum likelihood method will get overfitting results. In the correlation vector machine regression model, the maximum likelihood method is used under the Bayesian framework to train the model weights, so that the model has strong generalization. The prior distribution of the parameters is

[0094]

[0095] In the formula, α = [α1, α2,..., α N ] is a hyperparameter;

[0096] Based on the prior distribution and likelihood distribution, the posterior distribution of the weight can be obtained by Bayesian reasoning, and its posterior distribution is

[0097]

[0098] The posterior mean and variance are

[0099]

[0100] In the formula, A=diag(α0,α1,...,α N ).

[0101] Step 7: Update and calculate the hyperparameter α and noise variance σ 2 , remove the infinite elements in α, corresponding to ω i is 0, ω i is the i-th element of the model weight ω, and we get non-zero ω i The corresponding vector is the correlation vector, thus obtaining a sparse model.

[0102] In step 7, the output likelihood distribution can be obtained by performing marginal integration on the weights, that is,

[0103] P(y|α,σ 2 )=∫P(y|ω,σ 2 )P(ω|α)dω

[0104] Thus, the marginal likelihood distribution of the hyperparameters can be obtained, that is,

[0105] P(y|α,σ 2 )~N(0,C)

[0106] Where C = σ 2 I+ΦA -1 Φ T , I is the identity matrix;

[0107] According to MacKay's method, we can get

[0108]

[0109] Where μ i is the i-th element of the mean vector μ, and γ is defined in the MacKay method. i =1-α i N ii , where N ii is the i-th element on the diagonal of the variance Σ;

[0110]

[0111] The learning process of the RVM is to continuously iterate and update and (σ 2 ) new , and then update the statistics μ and Σ until the set convergence conditions are met; in the actual calculation process, the posterior distribution of many weights tends to zero. In the regression model of the correlation vector machine, non-zero weights represent prototype samples in the data, which are called "correlation vectors";

[0112] For a new set of inputs x * , and its corresponding output y * The prediction distribution of in

[0113]

[0114] Where μ * For expectations, is the variance, It is the maximum likelihood estimate of P(ω|α), which determines the optimal value of the weight.

[0115] In step eight, the hybrid kernel function parameters in each RVM are optimized using the artificial bee colony algorithm, wherein the objective function in the artificial bee colony algorithm is adaptively adjusted using GNSS and depth meter to obtain the optimal kernel function parameters.

[0116] In the step eight, the latitude and longitude information provided by GNSS and the depth information provided by the depth meter are introduced to constrain the objective function of the artificial bee colony algorithm. The improved artificial bee colony objective function can be expressed as

[0117]

[0118] in

[0119]

[0120] Where RVM1:fitness, RVM2:fitness and RVM3:fitness represent the artificial bee colony search objective functions in the eastward velocity, northward velocity and celestial velocity training models respectively, and v e 、v n and v u They represent the velocity of SINS output in the geographic coordinate system, v e* 、v n* and v u* where m1, m2 and m3 are the weight coefficients of the objective functions in the RVM1, RVM2 and RVM3 models, respectively. The weight coefficients are adaptively adjusted according to the proportional relationship between the velocity information provided by SINS and the equivalent velocity information output by external sensors, namely, GNSS and depth sensor.

[0121] The above embodiments are implemented under the premise of the technical solution of the present invention, and detailed implementation methods and specific operating processes are given, but the protection scope of the present invention is not limited to the above embodiments. The methods used in the above embodiments are conventional methods unless otherwise specified.

Claims

1. A DVL speed measurement model training method based on artificial bee colony optimization relevance vector machine, characterized in that: The following steps are involved: Step 1: Build an AUV test platform to obtain the raw data of the inertial navigation system, the three-axis velocity information output by the DVL, the longitude and latitude output by the GNSS, and the depth training samples provided by the depth meter. The raw data of the inertial navigation system includes the three-axis accelerometer and three-axis gyroscope information; Step 2: Initially align the acquired inertial navigation system data training samples. After the initial alignment is completed, perform real-time attitude and velocity calculations to obtain the vehicle's heading, pitch, roll, heading change rate dHeading, and the three-axis vehicle system velocity information V_x, V_y, and V_z; Step 3: Get the three-axis velocity information DVL_x, DVL_y and DVL_z of the DVL output carrier, and use DVL_x, DVL_y, DVL_z, Heading, Pitch, Roll and dHeading as the RVM model input set x, and V_x, V_y and V_z as the output set y to form the data training set Where N is the number of sampling points, which are used as training samples for the three RVM models; Step 4: Initialize the hybrid kernel function K(x,x) in each RVM i ) in the parameters; Step 5: Initialize hyperparameters α and noise variance σ 2 , using kernel function to map the normalized data into feature space; Step 6: Calculate the mean μ and variance Σ of the posterior distribution; Step 7: Update and calculate the hyperparameter α and noise variance σ 2 , remove the infinite elements in α, corresponding to ω i is 0, ω i is the i-th element of the model weight ω, and we get non-zero ω i The corresponding vector is the relevant vector, thus obtaining a sparse model; In step eight, the hybrid kernel function parameters in each RVM are optimized using the artificial bee colony algorithm, wherein the objective function in the artificial bee colony algorithm is adaptively adjusted using GNSS and depth meter to obtain the optimal kernel function parameters.

2. The DVL speed measurement model training method based on artificial bee colony optimization relevance vector machine according to claim 1 is characterized in that: In the above step 1: under the AUV test platform, the vehicle is initially aligned near the water surface, then dives underwater and maintains a uniform speed and straight navigation, with a navigation distance of not less than 1 km, and finally surfaces to the water surface; real-time data from the inertial navigation system and DVL, GNSS and depth meter are collected.

3. The DVL speed measurement model training method based on artificial bee colony optimization relevance vector machine according to claim 1 is characterized in that: In step 2, considering that the observability of the heading is poor under non-maneuvering conditions, a SINS / GNSS combined initial alignment is performed near the water surface to improve the observability of the heading; in addition, position and velocity information are used as observation information for the Kalman filter to accelerate the initial alignment process; Coning error will affect the accuracy of attitude solution in underwater application environment. In order to further improve the attitude solution accuracy of inertial navigation system, the three-sample rotating vector algorithm is applied to the attitude solution of strapdown inertial navigation.

4. The DVL speed measurement model training method based on artificial bee colony optimization relevance vector machine according to claim 1, characterized in that: In step 3, the velocity information output by the inertial navigation system is expressed in the geographic coordinate system, while the velocity output by the DVL is in the carrier coordinate system. In order to unify the two coordinate systems, the velocity information output by the inertial navigation system in the carrier coordinate system is converted into the velocity in the carrier coordinate system through the attitude conversion matrix.

5. The DVL speed measurement model training method based on artificial bee colony optimization relevance vector machine according to claim 1 is characterized in that: In step 4, the kernel function selection in the RVM model directly affects the learning and generalization capabilities of the model training. By utilizing the respective advantages of the local kernel function and the global kernel function, a weighted combination is performed to form the following hybrid kernel function: In the formula, a represents the weight of the Gaussian kernel function, b and c represent the width of the corresponding kernel function, and x i Represents the i-th input value of the model sample.

6. The DVL speed measurement model training method based on artificial bee colony optimization relevance vector machine according to claim 1 is characterized in that: In the step 5: given a training sample set is Among them, x i ∈R d ,y i ∈R; where x i represents the input sample, y i represents the target sample, d represents the dimension of the input sample; the regression model of the target sample can be expressed as: y i =f(x i ;ω)+ε n Where, ε n is the additional noise and satisfies ε n ∈N(0,σ 2 ), f(x) is defined as follows: Where N is the number of samples, ω=[ω0,ω1,...,ω N ] represents the weight, K(x,x i ) is the kernel function, Φ(x) is an N×(N+1) order matrix, which can be expressed as Since the target sample y i are independent, so the likelihood function of the training sample set can be expressed as: In the formula, y=[y1, y2,...,y N ] is the training target sample.

7. The DVL speed measurement model training method based on artificial bee colony optimization relevance vector machine according to claim 6, characterized in that: In step 6, for the multi-weight regression model, the maximum likelihood method will get overfitting results. In the correlation vector machine regression model, the maximum likelihood method is used under the Bayesian framework to train the model weights, so that the model has strong generalization. The prior distribution of the parameters is In the formula, α = [α1, α2,..., α N ] is a hyperparameter; Based on the prior distribution and likelihood distribution, the posterior distribution of the weight can be obtained by Bayesian reasoning, and its posterior distribution is The posterior mean and variance are In the formula, A=diag(α0,α1,...,α N )。 8. The DVL speed measurement model training method based on artificial bee colony optimization relevance vector machine according to claim 1 is characterized in that: In step 7, the output likelihood distribution can be obtained by performing marginal integration on the weights, that is, P(y|α,σ 2 )=∫P(y|ω,σ 2 )P(ω|α)dω Thus, the marginal likelihood distribution of the hyperparameters can be obtained, that is, P(y|α,σ 2 )~N(0,C) Where C = σ 2 I+ΦA -1 Φ T , I is the identity matrix; According to MacKay's method, we can get Where μ i is the i-th element of the mean vector μ, and γ is defined in the MacKay method. i =1-α i N ii , where N ii is the i-th element on the diagonal of the variance Σ; The learning process of the RVM is to continuously iterate and update and (σ 2 ) new , and then update the statistics μ and Σ until the set convergence conditions are met; In the actual calculation process, the posterior distribution of many weights tends to zero. In the regression model of the relevance vector machine, non-zero weights represent prototype samples in the data, which are called "relevance vectors"; For a new set of inputs x * , and its corresponding output y * The prediction distribution of in Where μ * For expectations, is the variance, It is the maximum likelihood estimate of P(ω|α), which determines the optimal value of the weight.

9. The DVL speed measurement model training method based on artificial bee colony optimization relevance vector machine according to claim 1, characterized in that: In the step eight, the latitude and longitude information provided by GNSS and the depth information provided by the depth meter are introduced to constrain the objective function of the artificial bee colony algorithm. The improved artificial bee colony objective function can be expressed as in Where RVM1:fitness, RVM2:fitness and RVM3:fitness represent the artificial bee colony search objective functions in the eastward velocity, northward velocity and celestial velocity training models respectively, and v e 、v n and v u They represent the velocity of SINS output in the geographic coordinate system, v e* 、v n* and v u* where m1, m2 and m3 are the weight coefficients of the objective functions in the RVM1, RVM2 and RVM3 models, respectively. The weight coefficients are adaptively adjusted according to the proportional relationship between the velocity information provided by SINS and the equivalent velocity information output by external sensors, namely, GNSS and depth sensor.

Images