CN118500392A - Underwater robot DVL speed measurement error correction method based on improved ELM - Google Patents

Abstract

The invention discloses an underwater robot Doppler log (DVL) speed measurement error correction method based on an improved Extreme Learning Machine (ELM). The purpose is to improve the accuracy and reliability of the SINS/DVL navigation system of the underwater robot. The invention applies the improved ELM model to a DVL speed error correction method, and the whole process comprises the following steps: the method comprises the steps of collecting output DVL measurement data of DVL equipment of underwater robot motion, preprocessing, generating a training set and a testing set of a DVL speed correction model, constructing a DVL speed prediction model of an improved ELM mixing method, training an ELM model under the condition that GPS signals are effective, and outputting the DVL speed prediction. The invention can solve the problem of the decline of the underwater navigation precision of the underwater robot caused by inconsistent installation angles of the inertial measurement unit and the DVL in the underwater robot navigation system.

Description

Error correction method of underwater robot DVL speed measurement based on improved ELM

Technical Field

The invention belongs to the technical field of underwater robot navigation and positioning, and in particular relates to an underwater robot DVL speed measurement error correction method based on improved ELM.

Background Art

Marine resources are abundant and there is great potential for their development. However, considering the particularity and complexity of the marine environment, its exploration and development face a series of challenges. Traditional manual exploration and development of marine resources is costly and dangerous. Therefore, people have developed autonomous underwater robots, which can help us realize the development of marine resources through the control of underwater robots. Because they can work in deep-sea environments and can conduct long-term detection and monitoring, they can not only significantly reduce costs, but also reduce the risk of personnel being exposed to dangerous environments.

The navigation and positioning of underwater robots is crucial to their safety and mission execution. Inertial navigation and DVL (Doppler Velocity Log) are commonly used underwater navigation and positioning technologies, but their accuracy is affected by the installation error angle. These errors may come from inaccurate equipment installation or changes in the underwater environment, such as underwater terrain, water flow, etc. In view of the impact of environmental changes on the installation angle estimation technology, it is also necessary to consider increasing the adaptability and robustness to environmental changes when designing the navigation system. This may involve the study of dynamic calibration methods, as well as real-time processing and adjustment of sensor data to adapt to different underwater environmental conditions.

Summary of the invention

In view of the above-mentioned problems, the present invention proposes an underwater robot DVL speed measurement error correction method based on improved ELM. First, the DVL speed information collected in the underwater robot navigation system is detected and eliminated. Here, the present invention proposes an outlier detection and elimination method based on least squares trend term modeling and Chauvigne criterion; secondly, the speed information is used to generate a training set of the DVL speed correction model. At the same time, the present invention constructs a DVL speed prediction model based on the improved ELM based on a hybrid weighted activation function and trains the model when the GPS signal is valid; finally, a DVL speed correction model is built based on the above method, and the accurate speed information output function is completed.

The above purpose is achieved through the following technical solutions:

The underwater robot DVL speed measurement error correction method based on improved ELM includes the following steps:

Step 1, collecting the output DVL measurement data of the DVL device of the underwater robot movement for preprocessing;

Step 2: The DVL measurement data preprocessed in step 1 is combined into a sample set , synchronously collect the robot's three-dimensional velocity information in the motion carrier coordinate system when the GPS signal is valid to form a sample set , the sample set and sample set The data are divided into time segments and combined into a training set for the DVL velocity prediction model training phase and a test set for the model testing phase;

Step 3: Construct a DVL velocity prediction model based on the improved ELM hybrid method;

Step 4: training the DVL speed prediction model of the improved ELM hybrid method constructed in step 3 when the GPS signal is valid;

Step 5: synchronously collect the speed information of INS, DVL and GPS, input the test set in step 2 into the DVL speed prediction model of the improved ELM hybrid method trained in step 4, and the model outputs the DVL speed information after error compensation, which is input into the INS and DVL combined navigation system. At the same time, the positioning error of the model is compared with the positioning error of the INS and GPS combined navigation system and the positioning error of the INS and DVL combined navigation system under the original DVL data to verify the accuracy of the ELM model.

Furthermore, the data is preprocessed in step 1, specifically including:

First, based on the least squares trend term modeling and the Chauvigne criterion, the sampling frequency of the output of the DVL device is The discrete DVL quantity data sequence ,in , U is the data length, outliers are removed, and the K-order fitting polynomial is constructed according to the trend characteristics of DVL measurement data :

,

in, are the coefficients of the fitted polynomial;

Introducing the residual sum of squares function :

,

And the residual sum of squares function Take the minimum value and adjust the coefficients of the fitted polynomial Taking the partial derivative to zero, we get

,

Calculate and expand the reorganization to get:

,

in, They are variables used to index different orders of polynomials;

Using matrix solution method to get the fitting polynomial coefficients And the trend term fitting polynomial is obtained, so the detrended term of the DVL measurement data is for:

,

The residual sequence of DVL measurement data for:

,

in, is the mean value of the detrended term of the DVL measurement data, and the calculation formula is:

,

Define the absolute value of the residual sequence satisfy The point data is suspicious data, that is, the wild value output by DVL, where is the Chauvigne criterion coefficient, and its coefficient fitting formula is: ; is the standard deviation of the DVL data after detrending, and its calculation formula is: ;

If it is determined to be an outlier, the data of the point will be removed and replaced by the mean value of the formula for detrending the DVL measurement data. At this point, the outliers in the DVL data output have been eliminated, and the DVL measurement data preprocessing is completed.

Furthermore, the training set of the DVL speed prediction model in step 2 includes:

2a. Step 1: The preprocessed DVL measurement data form a sample set ,in They are the velocity information of DVL in the x, y, and z directions in the carrier coordinate system;

2b. Synchronously collect the robot's three-dimensional velocity information in the carrier coordinate system when the GPS signal is valid to form a sample set ,in They are the speed information of the robot in the x, y, and z directions in the carrier coordinate system;

2c. , The data of these two sample sets are combined into the training set in the DVL velocity prediction model training phase and the test set in the model testing phase.

Furthermore, step 3 of constructing the DVL velocity prediction model of the improved ELM hybrid method includes the following sub-steps:

3a. Construct a single hidden layer feedforward algorithm model, including input layer, hidden layer and output layer; the input layer contains 3 data channels, the hidden layer contains 12 data channels, and the output layer contains 3 data channels;

3b. Mix weighted activation functions and determine the optimal ELM activation function in subsequent experiments :

,

in and are the hidden layer parameter weights and bias vectors respectively, and V is the input;

3c. Construct the output of the ELM model: ,

in, For the ideal output The corresponding actual output; is the weight matrix between the hidden layer and the output layer, ; represents the weight vector between the i-th unit and the hidden layer, ; The superscript T indicates the transpose of the matrix, express and The inner product of The regularization parameter balances the trade-off between the original objective function and the norm penalty term; is the weight of the input matrix, is the norm of the weights of the input matrix; is the bias vector, , L is the number of hidden layers, m is the dimension of the output vector;

3d. Simplify the output in step 3c to ,

in, is the output matrix of the model, equivalent to , ; is the output weight matrix of the hidden layer, , , is the input vector;

3e. Calculate the hidden layer output weight matrix and the weights between the hidden layer and the output layer , that is, the solution The minimum norm least squares solution of : , obtained by orthogonalization method, when the matrix When is a non-singular value, .

Furthermore, in step 4, the ELM model is trained when the GPS signal is valid. The specific method is:

4a. Input the training set in step 2 into the DVL velocity prediction model of the improved ELM hybrid method constructed in step 3 for training;

4b. Randomly set the weights of the input matrix according to a continuous probability distribution With the bias vector ;

4c. Calculate the hidden layer output weight matrix and the weights between the hidden layer and the output layer , that is, the solution The minimum norm least squares solution of : , obtained by orthogonalization method, when the matrix When is a non-singular value, ;

4d. Select the root mean square error RMSE as the criterion for whether the training is finished. , For ideal output;

4e. If the model error does not meet the RMSE requirement, determine whether the training samples are sufficient. If the training samples are insufficient, return to step 1; otherwise, increase the number of neurons in the hidden layer and return to step 3 to continue training the model.

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

(1) Compared with the prior art, the present invention proposes an outlier detection and elimination method based on least squares trend term modeling and Chauvigne criterion, which further improves the accuracy and reliability of DVL speed information;

(2) Compared with the prior art, the present invention adopts an extreme learning machine (ELM) to improve the accuracy of DVL speed information. At the same time, the present invention improves the existing ELM model based on a hybrid weighted activation function, further improving the efficiency and accuracy of DVL speed correction.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a method for correcting speed error of an underwater robot DVL based on an improved ELM of the present invention;

FIG2 is a speed prediction curve diagram of the DVL in three directions using the method of the present invention;

FIG. 3 is a DVL speed error curve diagram after correction by the method of the present invention.

DETAILED DESCRIPTION

As shown in FIG1 , the underwater robot DVL speed measurement error correction method based on improved ELM of the present invention comprises the following steps:

Step 1: Data preprocessing:

In order to prevent the abnormal value caused by the abnormal situation of DVL when the underwater robot moves, which may affect the effect of the overall model, the abnormal points in the output DVL measurement data of the DVL device that collects the movement of the underwater robot are eliminated and replaced before model construction and training to ensure the accuracy of the model.

The output DVL measurement data of the DVL device of the underwater robot is collected and preprocessed. The sampling frequency is set based on the least squares trend term modeling and the Chauvigne criterion. Discrete DVL data sequence Perform outlier elimination (where U is the data length). First, construct its K-order fitting polynomial based on the trend characteristics of DVL data :

,

in, are the coefficients of the fitted polynomial;

Introducing the residual sum of squares function :

,

And the residual sum of squares function Take the minimum value and adjust the coefficients of the fitted polynomial Taking the partial derivative to zero, we get

,

Calculate and expand the reorganization to get:

,

in, They are variables used to index different orders of polynomials;

Using matrix solution method to get the fitting polynomial coefficients And the trend term fitting polynomial is obtained, so the detrended term of the DVL measurement data is for:

,

The residual sequence of DVL measurement data for:

,

in, is the mean value of the detrended term of the DVL measurement data, and the calculation formula is:

,

Define the absolute value of the residual sequence satisfy The point data is suspicious data, that is, the wild value output by DVL, where is the Chauvigne criterion coefficient, and its coefficient fitting formula is: ; is the standard deviation of the DVL data after detrending, and its calculation formula is: ;

If it is determined to be an outlier, the data of the point will be removed and replaced by the mean value of the formula for detrending the DVL measurement data. At this point, the wild values in the DVL data output have been eliminated, that is, the DVL measurement data preprocessing is completed, preparing for the subsequent construction of the DVL speed training model.

Step 2: Generate the training set for the DVL speed prediction model training phase and the test set for the model testing phase:

2a. Step 1: The preprocessed DVL measurement data form a sample set ,in They are the velocity information of DVL in the x, y, and z directions in the carrier coordinate system;

2b. Synchronously collect the robot's three-dimensional velocity information in the carrier coordinate system when the GPS signal is valid to form a sample set ,in They are the speed information of the robot in the x, y, and z directions in the carrier coordinate system;

2c. , The data of these two sample sets are combined into the training set in the DVL velocity prediction model training phase and the test set in the model testing phase.

Step 3: Construct a DVL velocity prediction model based on the improved ELM hybrid method:

3a. Construct a single hidden layer feedforward algorithm model, including input layer, hidden layer and output layer; the input layer contains 3 data channels, the hidden layer contains 12 data channels, and the output layer contains 3 data channels;

3b. Mix weighted activation functions and determine the optimal ELM activation function in subsequent experiments :

,

in and are the hidden layer parameter weights and bias vectors respectively, and V is the input;

3c. Construct the output of the ELM model: ,

in, For the ideal output The corresponding actual output; is the weight matrix between the hidden layer and the output layer, ; represents the weight vector between the i-th unit and the hidden layer, ; The superscript T indicates the transpose of the matrix, express and The inner product of The regularization parameter balances the trade-off between the original objective function and the norm penalty term; is the weight of the input matrix, is the norm of the weights of the input matrix; is the bias vector, , L is the number of hidden layers, m is the dimension of the output vector;

3d. Simplify the output in step 3c to ,

in, is the output matrix of the model, equivalent to , ; is the output weight matrix of the hidden layer, , , is the input vector;

3e. Calculate the hidden layer output weight matrix and the weights between the hidden layer and the output layer , that is, the solution The minimum norm least squares solution of : , obtained by orthogonalization method, when the matrix When is a non-singular value, .

Step 4: Train the ELM model when the GPS signal is valid:

4a. Input the training set in step 2 into the DVL velocity prediction model of the improved ELM hybrid method constructed in step 3 for training;

4b. Randomly set the weights of the input matrix according to a continuous probability distribution With the bias vector ;

4c. Calculate the hidden layer output weight matrix and the weights between the hidden layer and the output layer , that is, the solution The minimum norm least squares solution of : , obtained by orthogonalization method, when the matrix When is a non-singular value, ;

4d. Select the root mean square error RMSE as the criterion for whether the training is finished. , For ideal output;

4e. If the model error does not meet the RMSE requirement, determine whether the training samples are sufficient. If the training samples are insufficient, return to step 1; otherwise, increase the number of neurons in the hidden layer and return to step 3 to continue training the model.

Step 5: DVL speed error model test:

The speed information of INS, DVL and GPS is collected synchronously, and the test set in step 2 is input into the trained ELM model. The model outputs the DVL speed information after error compensation, and inputs it into the INS and DVL combined navigation system. At the same time, its positioning error is compared with the positioning error of the INS and GPS combined navigation system and the positioning error of the INS and DVL combined navigation system under the original DVL data to verify the accuracy of the ELM model.

Experimental verification:

In order to verify the effectiveness of the method of the present invention, an experimental verification was designed. The experimental equipment includes: GPS equipment, DVL equipment, navigation computer, IMU equipment, etc. The navigation computer is responsible for collecting GPS, DVL and IMU information; the algorithm proposed in the present invention runs in the navigation computer. Figure 2 shows the speed prediction curve of the three directions of the DVL using the method of the present invention; Figure 3 shows the DVL speed error curve after correction using the method of the present invention. It can be seen from Figure 2 that the present invention can predict its output when the DVL signal fails, and it can be seen from Figure 3 that its predicted speed accuracy is relatively high, and it can correct the DVL speed error well.

Claims (5)

1. The method for correcting the DVL speed measurement error of the underwater robot based on the improved ELM is characterized by comprising the following steps:

step 1, acquiring output DVL measurement data of DVL equipment of underwater robot motion for preprocessing;

step 2, the preprocessed DVL measurement data in the step 1 is formed into a sample set Three-dimensional speed information of synchronous acquisition robot under GPS signal effective condition motion carrier coordinate system constitutes sample setSample setAnd a sample setRespectively combining the data of the DVL speed prediction model training stage and the data of the model testing stage according to time segments;

Step 3, constructing a DVL speed prediction model for improving the ELM mixing method;

Step4, training the DVL speed prediction model of the improved ELM mixing method constructed in the step 3 under the condition that GPS signals are effective;

and 5, synchronously acquiring speed information of inertial navigation, DVL and GPS, inputting the test set in the step 2 into the DVL speed prediction model of the improved ELM mixing method trained in the step 4, outputting error-compensated DVL speed information by the model, inputting the error-compensated DVL speed information into an inertial navigation and DVL combined navigation system, and simultaneously comparing the positioning error of the model with the positioning error of the inertial navigation and GPS combined navigation system and the positioning error of the inertial navigation and DVL combined navigation system under original DVL data, and verifying the accuracy of the ELM model.

2. The method for correcting the speed measurement error of the underwater robot DVL based on the improved ELM according to claim 1, wherein the preprocessing of the data in step 1 specifically includes:

the sampling frequency of the output of DVL equipment is firstly based on least square method trend item modeling and the Showy-Fresnel rule Is a discrete DVL volume data sequenceWhereinU is the data length, outlier rejection processing is carried out, and K-order fitting polynomials are constructed according to trend characteristics of DVL measurement data：

，

Wherein,Coefficients for a fitting polynomial;

Introducing a residual sum-of-squares function ：

，

And sum of squares function of residual errorsTake minima and pass through the coefficients of the fitting polynomialObtaining the offset guide to obtain zero

，

Calculating and expanding recombination to obtain:

，

wherein, Variables used to index the different orders of the polynomial, respectively;

obtaining fitting polynomial coefficients by matrix solution And a trend term fitting polynomial is obtained, so that the trend term of the DVL measurement data is removedThe method comprises the following steps:

，

residual sequence of DVL measurement data The method comprises the following steps:

，

wherein, Removing a mean value of a calculation formula of a trend term for DVL measurement data, wherein the calculation formula is as follows:

，

Defining absolute values of residual sequences Satisfy the following requirementsIs suspicious data, i.e. wild value of DVL output, whereinThe coefficient fitting formula is as follows:； the standard deviation of the sequence after removing the trend term for DVL data is calculated as follows: ；

If the point data is determined to be the wild value, the point data is subjected to elimination processing and replaced by the mean value of the calculation formula of the DVL measurement data trending term And eliminating the outlier in the DVL data output so far, namely finishing the preprocessing of the DVL measurement data.

3. The method for correcting speed measurement errors of an underwater robot DVL based on improved ELM as claimed in claim 2, wherein the training set of the DVL speed prediction model in step 2 includes:

2a. Step 1 the preprocessed DVL measurement data form a sample set WhereinRespectively the velocity information of the DVL in the x, y and z directions under the carrier coordinate system;

2b, synchronously acquiring three-dimensional speed information of the robot under the carrier coordinate system under the condition that GPS signals are effective to form a sample set WhereinThe speed information of the robot in the x, y and z directions under the carrier coordinate system is respectively obtained;

2c, will 、The data of the two sample sets are combined into a training set of a DVL speed prediction model training stage and a test set of a model test stage.

4. The method for correcting the DVL speed measurement error of an underwater robot based on the improved ELM according to claim 1, wherein the constructing the DVL speed prediction model of the improved ELM mixing method in step 3 includes the following sub-steps:

3a, constructing a single hidden layer feedforward algorithm model which comprises an input layer, a hidden layer and an output layer 3 part; wherein the input layer comprises 3 data channels, the hidden layer comprises 12 data channels, and the output layer comprises 3 data channels;

mixing weighted activation functions and determining an optimal ELM activation function in later experiments ：

，

Wherein the method comprises the steps ofAndRespectively an implicit layer parameter weight and a bias vector, wherein V is input;

3c, constructing an ELM model output: ，

wherein, To and ideal outputCorresponding actual output; To conceal the weight matrix between the layers and the output layer, ；Representing the weight vector of the i-th cell and the hidden layer relative to each other,; The superscript T denotes the transpose of the matrix,Representation ofAnd (3) withIs an inner product of (2); Balancing the trade-off relation between the original objective function and the norm penalty term for the regularization parameter; Is the weight of the input matrix and, Is the norm of the weights of the input matrix; as a result of the offset vector, L is the number of hidden layers, m is the dimension of the output vector;

3d. Simplifying the output in step 3c to ，

Wherein,Is the output matrix of the model, equivalent，；For the output weight matrix of the hidden layer,，，Is an input vector;

Calculating the hidden layer output weight matrix And weights between hidden layer and output layerI.e. solvingLeast-squares solution of (2)：Obtained by orthogonalization method, when matrixWhen the value is non-singular, then。

5. The method for correcting the speed measurement error of the underwater robot DVL based on the improved ELM according to claim 4, wherein the training ELM model in the case where the GPS signal is valid in step 4 is specifically as follows:

4a, inputting the training set in the step 2 into a DVL speed prediction model of the improved ELM mixing method constructed in the step 3 for training;

4b, randomly setting the weight of the input matrix according to the continuous probability distribution And offset vector；

4C, calculating the hidden layer output weight matrixAnd weights between hidden layer and output layerI.e. solvingLeast-squares solution of (2)：Obtained by orthogonalization method, when matrixWhen the value is non-singular, then；

4D, selecting the root mean square error RMSE as a judgment standard of whether training is finished,，Is an ideal output;

and 4e, judging whether the training samples are enough if the model errors do not meet the RMSE requirement, returning to the step 1 if the training samples are insufficient, otherwise, increasing the number of neurons in the hidden layer, returning to the step 3, and then training the model.