A method for underwater vehicle motion prediction based on deep learning error correction

Abstract

The present invention relates to the field of electronic digital data processing technology, and in particular to a method for underwater vehicle motion prediction based on deep learning error correction, comprising the following steps: designing the input and output of an underwater vehicle motion prediction model based on deep learning error correction; motion deduction prediction, obtaining a motion deduction sequence, and calculating the error sequence between the motion deduction sequence and the real historical motion sequence; adaptive variational mode decomposition, obtaining multiple intrinsic mode function component sequences; constructing a framework of a bidirectional long short-term memory network prediction model combined with a control command attention mechanism; training a prediction model; performing underwater vehicle motion prediction based on deep learning error correction; and improving the underwater vehicle motion prediction performance after optimizing the initial parameters. The method provided by the present invention improves the performance of motion prediction by adaptive variational mode decomposition, introducing a control command attention mechanism, and designing a deep learning prediction network model.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic digital data processing, and in particular to a method for predicting underwater vehicle motion based on deep learning error correction. Background Art

[0002] In underwater vehicle applications, motion forecasting enables the control system to give control instructions that meet actual applications in advance through short-term or very short-term forecasts of motion posture, providing better control input to the real-time control system. At the same time, it can also obtain current navigation status information and avoid risks in a timely manner to ensure navigation safety.

[0003] The development direction of motion prediction has mainly gone through frequency domain analysis and time domain analysis. Recurrent neural network is a time recursive neural network that processes information of sequence structure. Among them, long short-term memory neural network has long-term memory ability and can effectively process time series data. However, considering that underwater vehicles are disturbed by the ocean environment during navigation, there are many random characteristics in the motion prediction error, and the long short-term memory neural network has an unsatisfactory fitting effect on fluctuating data. In addition, it is not easy to obtain navigation and control data of underwater vehicles under various working conditions in the early stage, which is not conducive to the construction of motion prediction models based entirely on deep learning, resulting in relatively low accuracy of underwater vehicle motion prediction. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for underwater vehicle motion prediction based on deep learning error correction. In view of the problems that underwater data has large noise, many random features, and the accuracy of motion prediction is affected by changes in control commands, the accuracy of motion prediction is improved by adaptive variational mode decomposition, introduction of control command attention mechanism, design of deep learning prediction network model, initial parameter optimization, etc.

[0005] The present invention is achieved through the following technical solutions:

[0006] A method for predicting underwater vehicle motion based on deep learning error correction comprises the following steps:

[0007] S1: According to the motion characteristics of underwater vehicles, considering the impact of control commands on the future motion of underwater vehicles, the input and output of the underwater vehicle motion prediction model based on deep learning error correction is designed using the underwater vehicle historical data;

[0008] S2: Perform motion prediction based on the input of the underwater vehicle motion prediction model based on deep learning error correction, obtain the motion prediction sequence, and calculate the error sequence between the motion prediction sequence and the real historical motion sequence;

[0009] S3: Adaptively perform variational mode decomposition on the error sequence between the motion deduction sequence and the real historical motion sequence to obtain multiple eigenmode function component sequences;

[0010] S4: After processing multiple eigenmode function component sequences with forward LSTM neural network and reverse LSTM neural network, a bidirectional LSTM network prediction model combined with the control instruction attention mechanism is constructed;

[0011] S5: Based on multiple eigenmode function component sequences, a bidirectional long short-term memory network prediction model combined with a control instruction attention mechanism is trained to obtain a bidirectional long short-term memory network prediction model combined with a control instruction attention mechanism for each eigenmode function component;

[0012] S6: Combine the motion deduction sequence and each eigenmode function component with the bidirectional long short-term memory network prediction model combined with the control command attention mechanism to perform underwater vehicle motion prediction based on deep learning error correction, and form an underwater vehicle motion prediction model based on deep learning error correction;

[0013] S7: After optimizing the initial parameters of the bidirectional long short-term memory network prediction model combined with the control command attention mechanism for each eigenmode function component through the improved particle swarm optimization algorithm, repeat steps S5 and S6 to realize the motion prediction of the underwater vehicle.

[0014] Preferably, the input of the motion prediction model in step S1 is a sequence including the longitude, latitude, depth, heading, pitch, roll, lateral speed, longitudinal speed, vertical speed, heading angular velocity, pitch angular velocity, roll angular velocity and stern rudder command rudder angle of the underwater vehicle within a set time window, and the output of the motion prediction model is a sequence including the longitude, latitude, depth, heading, pitch and roll of the underwater vehicle within the time window from the current moment to the prediction moment.

[0015] Preferably, in step S2, motion prediction is performed based on the hydrodynamic model of the underwater vehicle. After the duration of the motion prediction is determined, the trajectory and attitude motion prediction based on the hydrodynamic model of the underwater vehicle is realized through the underwater vehicle open-loop control algorithm based on the Kalman filter algorithm to obtain the motion prediction sequence.

[0016] Preferably, when the hydrodynamic model of the underwater vehicle is missing in step S2, the duration of the motion forecast is first determined, and then the rotation matrix of the underwater vehicle from the current moment to the forecast moment is obtained according to the angular velocity, and the translation matrix of the underwater vehicle from the current moment to the forecast moment is obtained according to the velocity, and the trajectory and attitude motion forecast information of the underwater vehicle at the forecast moment are obtained by calculating the integral positioning position of rotation and translation to form a motion deduction sequence.

[0017] Furthermore, in step S3, the following method is used to perform adaptive variational mode decomposition on the error sequence between the motion deduction sequence and the real historical motion sequence:

[0018] S31: According to formula (1), the error sequence between the motion deduction sequence and the real historical motion sequence is decomposed into multiple eigenmode function components:

[0019] (1);

[0020] in: Indicates from Time has come The error sequence between the moment-by-moment motion sequence and the real historical motion sequence, Indicates from Time has come Moment decomposition eigenmode function components, Indicates from Time has come The residual sequence after moment decomposition, Indicates the forecast duration. represents the serial number of the decomposed eigenmode function component, Represents the total number of decomposed eigenmode function components;

[0021] S32: Calculate the Spearman correlation coefficient according to formula (2):

[0022] (2);

[0023] in: represents the Spearman correlation coefficient, Indicates from Time has come The error sequence between the moment motion deduction sequence and the real historical motion sequence The error of a moment, represents the mean of the error sequence between the motion deduction sequence and the real historical motion sequence, Indicates from Time has come After the moment of decomposition The residual sequence at each moment is represents the mean of the residual sequence after decomposition, Indicates the length of the sequence;

[0024] S33: Compare the absolute value of the Spearman correlation coefficient with the set threshold value. If the absolute value of the Spearman correlation coefficient is greater than or equal to the set threshold value, continue the decomposition. If the absolute value of the Spearman correlation coefficient is less than the set threshold value, stop the decomposition.

[0025] Further, in step S4, the forward long short-term memory neural network processing and the reverse long short-term memory neural network processing of the plurality of eigenmode function component sequences include the following steps:

[0026] S41: normalizing each eigenmode function component in each eigenmode function component sequence according to formula (3), and normalizing the four stern rudder command rudder angle characteristics in the input of the underwater vehicle motion prediction model:

[0027] (3);

[0028] in: represents the normalized sequence of the intrinsic mode function component sequence, represents the sequence of eigenmode function components, represents the minimum value in the sequence of components of the eigenmode function, represents the maximum value in the sequence of components of the eigenmode function;

[0029] S42: History Time has come The normalized sequence of the eigenmode function component sequence at time Time has come The input of the underwater vehicle motion prediction model at time t is the normalized sequence of the command rudder angle characteristics of the four stern rudders. Time has come The normalized sequence of the eigenmode function components at each moment is used as the output for forward long short-term memory neural network processing;

[0030] S43: The future Time has come The normalized sequence of the eigenmode function component sequence at the moment is taken as input, and the historical Time has come The sequence of normalized components of the eigenmode function at time Time has come The normalized sequence of the command rudder angles of the four stern rudders of the underwater vehicle at each moment is used as the output for inverse long short-term memory neural network processing.

[0031] Furthermore, in step S4, a bidirectional long short-term memory network prediction model combining the control instruction attention mechanism is constructed according to formula (4):

[0032] (4);

[0033] in: represents the query vector, Represents a reverse long short-term memory neural network The output of the moment, represents the parameter matrix of the query vector, represents the key vector, Indicates that each eigenmode function component predicts the model in the future through a bidirectional long short-term memory network Time has come Hidden layer state information within the time period, represents the parameter matrix of the key vector, represents a value vector, Parameter matrix representing the value vector.

[0034] Further, in step S5, the bidirectional long short-term memory network prediction model combined with the control instruction attention mechanism is trained according to the following method:

[0035] S51: After initializing the initial parameters of the bidirectional long short-term memory network prediction model combined with the control instruction attention mechanism, the bidirectional long short-term memory network prediction model combined with the control instruction attention mechanism is trained based on the eigenmode function component sequence to obtain the prediction model of the eigenmode function component:

[0036] S52: Based on the forecast output information sequence and future Time has come The bi-norm design of the sequence of the eigenmode function components at each moment is combined with the loss function of the bidirectional long short-term memory network prediction model of the control instruction attention mechanism, and the prediction model of each eigenmode function component is trained based on the loss function;

[0037] S53: After the training is completed, the prediction model of each eigenmode function component is output and saved.

[0038] Further, in step S6, the following method is used to form an underwater vehicle motion prediction model based on deep learning error correction:

[0039] S61: forecasting each eigenmode function component in the future time based on the forecast model of the eigenmode function component, obtaining a forecast sequence of each eigenmode function component in the future time, and reconstructing the forecast sequences of all eigenmode function components in the future time into a reconstructed prediction error sequence;

[0040] S62: sum the reconstructed prediction error sequence and the motion deduction sequence to obtain the underwater vehicle future time forecast sequence, perform underwater vehicle motion forecast based on deep learning error correction, and form an underwater vehicle motion forecast model based on deep learning error correction.

[0041] Further, in step S7, the initial parameters of the bidirectional long short-term memory network prediction model combined with the control instruction attention mechanism of each eigenmode function component are optimized by using the improved particle swarm optimization algorithm, including the following steps:

[0042] S71: representing the initial parameters to be optimized of the bidirectional long short-term memory network prediction model of each eigenmode function component combined with the control instruction attention mechanism as different dimensions of particles;

[0043] S72: Initialize relevant parameters of the improved particle swarm optimization algorithm;

[0044] S73: using the prediction error of the underwater vehicle motion prediction model based on deep learning error correction as the fitness function of the particle swarm optimization algorithm, and performing optimization iterations;

[0045] S74: Update the iterative particle update speed of the improved particle swarm optimization algorithm according to formula (5):

[0046] (5);

[0047] in: Indicates Particle No. Iteration particle update speed, Indicates Particle No. Iteration particle update speed, represents the inertia weight, represents the first learning factor, Represents a random number in the range of 0 to 1 generated by a random function. Indicates Particle No. The optimal value of the iteration, Indicates Particle No. The position of the particle in the iteration, represents the second learning factor, represents the global optimal value;

[0048] S75: Update the inertia weight of the improved particle swarm optimization algorithm according to formula (6):

[0049] (6);

[0050] in: represents the maximum value of the inertia weight, represents the minimum value of the inertia weight, Represents the iteration radian factor.

[0051] Beneficial effects of the invention:

[0052] The present invention provides a method for underwater vehicle motion prediction based on deep learning error correction. In the absence of real historical data of underwater vehicles, motion deduction is first performed. As the real historical data increases, the error between real motion information and motion deduction motion information is predicted based on deep learning, and the motion prediction accuracy is gradually improved. The present invention aims to improve the performance of motion prediction by adaptive variational mode decomposition, introducing a control command attention mechanism, designing a deep learning prediction network model, and optimizing initial parameters, etc., in order to solve the problem that underwater data of underwater vehicles has large noise, many random features, and the control command changes affect the motion prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0054] A method for underwater vehicle motion prediction based on deep learning error correction, the flow chart is as follows Figure 1 As shown, the specific steps include:

[0055] S1: Based on the motion characteristics of underwater vehicles, considering the impact of control commands on the future motion of underwater vehicles, the input and output of the underwater vehicle motion prediction model based on deep learning error correction is designed using the underwater vehicle historical data;

[0056] When designing the input and output of the underwater vehicle motion prediction model based on deep learning error correction, it is first necessary to collect historical data such as the underwater vehicle's position, speed, attitude, and control instructions.

[0057] The historical data sequences of motion information such as longitude, latitude, depth, heading, roll, pitch, lateral speed, longitudinal speed, vertical speed, roll angular velocity, pitch angular velocity, heading angular velocity and the historical data sequences of command rudder angle of the underwater vehicle under different working conditions are collected at a fixed period. The input of the motion prediction model is a sequence including the longitude, latitude, depth, heading, pitch, roll, lateral speed, longitudinal speed, vertical speed, heading angular velocity, pitch angular velocity, roll angular velocity and stern rudder command rudder angle of the underwater vehicle within a set time window. The output of the underwater vehicle motion prediction model based on deep learning error correction is a sequence including the longitude, latitude, depth, heading, pitch and roll of the underwater vehicle within the time window from the current moment to the prediction moment.

[0058] S2: Perform motion prediction based on the input of the underwater vehicle motion prediction model based on deep learning error correction, obtain the motion prediction sequence, and calculate the error sequence between the motion prediction sequence and the real historical motion sequence;

[0059] Considering the high cost of obtaining navigation and control data of underwater vehicles under various working conditions, it is difficult to support the training of motion prediction network models based on deep learning when the amount of data is small; at the same time, the hydrodynamic motion model of underwater vehicles is difficult to obtain accurately, and it is easily disturbed by the environment, making the model inaccurate. Therefore, in order to realize the motion prediction of underwater vehicles, motion deduction is used as the basis of motion prediction, and the error estimation and correction of motion deduction sequence and real historical motion sequence are used as the core of underwater vehicle motion prediction. A motion prediction model based on deep learning error correction is designed.

[0060] When performing motion simulation and prediction based on the hydrodynamic model of the underwater vehicle, after determining the duration of the motion prediction, the Kalman filter algorithm can be used as a basis to implement the trajectory and attitude motion prediction based on the hydrodynamic model of the underwater vehicle through the underwater vehicle open-loop control algorithm to obtain the motion simulation sequence.

[0061] When historical data of underwater vehicles is lacking, motion prediction is achieved through velocity-integrated motion deduction based on motion information or motion deduction based on hydrodynamic models. As the data increases, a deep learning prediction model is designed based on the error sequence between the real historical motion sequence and the motion deduction sequence. The accuracy of underwater vehicle motion prediction can be gradually improved through error estimation and error correction.

[0062] When the hydrodynamic model of the underwater vehicle is missing, the duration of the motion forecast is first determined, and then the rotation matrix of the underwater vehicle from the current moment to the forecast moment is obtained based on the angular velocity, and the translation matrix of the underwater vehicle from the current moment to the forecast moment is obtained based on the speed. The trajectory and attitude motion forecast information of the underwater vehicle at the forecast moment is obtained through rotation and translation integral positioning calculation to form a motion deduction sequence.

[0063] S3: Adaptively perform variational mode decomposition on the error sequence between the motion deduction sequence and the real historical motion sequence to obtain multiple eigenmode function component sequences;

[0064] Considering that underwater vehicles are disturbed by the ocean environment when navigating, the error has more random characteristics, and the navigation sensor accuracy of underwater vehicles is greatly reduced compared to surface vehicles. When predicting and estimating the error sequence, the influence of its random characteristics will affect the accuracy of motion prediction.

[0065] In order to reduce the impact of underwater environment and sensor errors on motion prediction, the error data can be decomposed by signal processing to weaken its nonlinearity and non-stationarity. In the signal processing algorithm, adaptive variational mode decomposition can decompose the time series into multiple modal components, which has the advantages of keeping the modal components independent of each other, not obvious modal aliasing, high computational efficiency and stability. Therefore, the adaptive variational mode decomposition of the error sequence can reduce the impact of underwater environment and sensor errors on motion prediction, thereby improving the prediction accuracy.

[0066] Specifically, the following method can be used to perform adaptive variational mode decomposition on the error sequence between the motion deduction sequence and the real historical motion sequence:

[0067] S31: According to formula (1), the error sequence between the motion deduction sequence and the real historical motion sequence is decomposed into multiple eigenmode function components:

[0068] (1);

[0069] in: Indicates from Time has come The error sequence between the moment-by-moment motion sequence and the real historical motion sequence, Indicates from Time has come Moment decomposition eigenmode function components, Indicates from Time has come The residual sequence after moment decomposition, Indicates the forecast duration. represents the serial number of the decomposed eigenmode function component, Represents the total number of decomposed eigenmode function components;

[0070] S32: Calculate the Spearman correlation coefficient according to formula (2):

[0071] (2);

[0072] in: represents the Spearman correlation coefficient, Indicates from Time has come The error sequence between the moment motion deduction sequence and the real historical motion sequence The error of a moment, represents the mean of the error sequence between the motion deduction sequence and the real historical motion sequence, Indicates from Time has come After the moment of decomposition The residual sequence at each moment is represents the mean of the residual sequence after decomposition, Indicates the length of the sequence;

[0073] S33: Compare the absolute value of the Spearman correlation coefficient with a set threshold. If the absolute value of the Spearman correlation coefficient is greater than or equal to the set threshold, continue decomposition. If the absolute value of the Spearman correlation coefficient is less than the set threshold, stop decomposition. The set threshold here can be a value close to zero.

[0074] S4: After processing multiple eigenmode function component sequences with forward LSTM neural network and reverse LSTM neural network, a bidirectional LSTM network prediction model combined with the control instruction attention mechanism is constructed;

[0075] Analyzing the nonlinear time series data characteristics of the complex and changeable underwater environment, noise interference, and time-varying control commands, the underwater vehicle's historical position, attitude, speed and other motion data sequences can reflect the future position, attitude, speed and other motion data change trends, and the future motion data can also reversely analyze the control characteristics of the underwater vehicle by the actuators such as the control commands. The navigation and control data sequence of the underwater vehicle has a certain contextual bidirectional relationship. Therefore, the forward long short-term memory network and the reverse long short-term memory network are used to process the sequence data in the bidirectional long short-term memory network prediction model, which can simultaneously capture the context information in the time series data, thereby improving the performance of the time series data prediction model.

[0076] Specifically, performing forward long short-term memory neural network processing and reverse long short-term memory neural network processing on multiple eigenmode function component sequences includes the following steps:

[0077] S41: normalizing each eigenmode function component in each eigenmode function component sequence according to formula (3), and normalizing the four stern rudder command rudder angle characteristics in the input of the underwater vehicle motion prediction model:

[0078] (3);

[0079] in: represents the normalized sequence of the intrinsic mode function component sequence, represents the sequence of eigenmode function components, represents the minimum value in the sequence of components of the eigenmode function, represents the maximum value in the sequence of components of the eigenmode function;

[0080] S42: History Time has come The normalized sequence of the eigenmode function component sequence at time Time has come The input of the underwater vehicle motion prediction model at time t is the normalized sequence of the command rudder angle characteristics of the four stern rudders. Time has come The normalized sequence of the eigenmode function components at each moment is used as the output for forward long short-term memory neural network processing;

[0081] S43: The future Time has come The normalized sequence of the eigenmode function component sequence at the moment is taken as input, and the historical Time has come The sequence of normalized components of the eigenmode function at time Time has come The normalized sequence of the command rudder angles of the four stern rudders of the underwater vehicle at each moment is used as the output for inverse long short-term memory neural network processing.

[0082] By using forward long short-term memory network and reverse long short-term memory network to process sequence data, the contextual information of the eigenmode function component sequence and the control command sequence can be more comprehensively captured, that is, the contextual influence of the control command on motion prediction can be captured, thereby improving the motion prediction performance of underwater vehicles.

[0083] Concatenating the output information of the forward LSTM network and the reverse LSTM network constitutes the output of the bidirectional LSTM network prediction model.

[0084] As the control command changes, the motion prediction information of the underwater vehicle will also change. However, due to the inertial characteristics of the underwater vehicle, it is difficult for the underwater vehicle to respond to the command immediately when the control command changes. In addition, the control command at different times has different effects on future motion information, and the current control information has the greatest impact on future motion information. Considering that the bidirectional long short-term memory network prediction model does not have the ability to directly focus on specific parts, this leads to a lag in the overall motion prediction when the control command changes. Therefore, it is necessary to introduce an attention mechanism and use the focusing characteristics of the multi-head attention mechanism to increase the attention weight of the impact of the control command change on the future motion data, thereby improving the motion prediction accuracy of the underwater vehicle when the control command changes.

[0085] The essence of the attention mechanism is to determine the importance of certain parts of the input features to the task by generating a weight coefficient to perform a weighted summation of the input. In trajectory data processing, the attention mechanism maps trajectory features into three different subspaces: query vector, key vector, and value vector. Through the query and scoring mechanism, control instructions at different times are assigned different weights, which accurately reflect the degree of relevance of the motion prediction sequence. When the degree of relevance is higher, the corresponding motion prediction feature is more significant, so the weight obtained in the attention mechanism is also greater.

[0086] In the reverse long short-term memory network, The hidden layer state of the output is updated at all times, which contains the contextual relationship between the steering angle of the control command and the historical motion trajectory and motion posture in the entire prediction input sequence. In order to improve the generalization of the prediction network when the amount of data is insufficient, the importance of the control command to the overall motion information is identified. The attention relationship between the hidden layer state of the reverse long short-term memory network and the predicted output sequence is calculated, and the attention weight is calculated. The query vector is obtained through linear transformation, and the hidden layer state of the reverse long short-term memory network is used as the key vector and value vector through different linear transformations.

[0087] Specifically, a bidirectional long short-term memory network prediction model combined with the control instruction attention mechanism can be constructed according to formula (4):

[0088] (4);

[0089] in: represents the query vector, Represents a reverse long short-term memory neural network The output of the moment, represents the parameter matrix of the query vector, represents the key vector, Indicates that each eigenmode function component predicts the model in the future through a bidirectional long short-term memory network Time has come Hidden layer state information within the time period, represents the parameter matrix of the key vector, represents a value vector, Parameter matrix representing the value vector.

[0090] here , represents the number of layers of the bidirectional long short-term memory network prediction model, represents the number of hidden layer neuron outputs in the bidirectional long short-term memory network prediction model, represents the dimension of the attention mechanism model, Represents the set of real numbers.

[0091] The output predicted by the attention mechanism is as follows:

[0092]

[0093] in: express , , The weighted sum of the values, is the activation function, Indicates the normalized attention weight matrix. is the matrix transpose, Represents the key vector The dimension of can make the gradient update more stable during training.

[0094] S5: Based on multiple eigenmode function component sequences, a bidirectional long short-term memory network prediction model combined with a control instruction attention mechanism is trained to obtain a bidirectional long short-term memory network prediction model combined with a control instruction attention mechanism for each eigenmode function component;

[0095] Specifically, the bidirectional long short-term memory network prediction model combined with the control instruction attention mechanism can be trained according to the following method:

[0096] S51: After initializing the initial parameters of the bidirectional long short-term memory network prediction model combined with the control instruction attention mechanism, the bidirectional long short-term memory network prediction model combined with the control instruction attention mechanism is trained based on the eigenmode function component sequence to obtain the prediction model of the eigenmode function component:

[0097] The initial parameters here include the time window size, the number of hidden layer units, the initial learning rate, and the regularization parameter.

[0098] S52: Based on the forecast output information sequence and future Time has come The bi-norm design of the sequence of the eigenmode function components at each moment is combined with the loss function of the bidirectional long short-term memory network prediction model of the control instruction attention mechanism, and the prediction model of each eigenmode function component is trained based on the loss function;

[0099] In order to avoid overfitting or underfitting of the prediction model during training, the Dropout regularization method can be used, that is, some neurons are randomly ignored during network training, so that some features of the input data are randomly deleted, so that the neurons in the network will not be overly dependent on the features of a certain input data.

[0100] S53: After the training is completed, the prediction model of each eigenmode function component is output and saved.

[0101] S6: Combine the motion deduction sequence and each eigenmode function component with the bidirectional long short-term memory network prediction model combined with the control command attention mechanism to perform underwater vehicle motion prediction based on deep learning error correction, and form an underwater vehicle motion prediction model based on deep learning error correction;

[0102] Specifically, the following method can be used to form an underwater vehicle motion prediction model based on deep learning error correction:

[0103] S61: forecasting each eigenmode function component in the future time based on the forecast model of the eigenmode function component, obtaining a forecast sequence of each eigenmode function component in the future time, and reconstructing the forecast sequences of all the eigenmode function components in the future time into a reconstructed prediction error sequence;

[0104] The specific reconstruction formula is shown in formula (1).

[0105] S62: sum the reconstructed prediction error sequence and the motion deduction sequence to obtain the underwater vehicle future time forecast sequence, perform underwater vehicle motion forecast based on deep learning error correction, and form an underwater vehicle motion forecast model based on deep learning error correction.

[0106] Furthermore, the motion estimation of multiple reference points of the underwater vehicle contour can be realized by combining the underwater vehicle attitude angle sequence and the space structure of the vehicle.

[0107] The future time prediction sequence of the underwater vehicle is a prediction sequence including longitude, latitude, depth, roll angle, pitch angle and heading angle.

[0108] Taking into account that the motion prediction of underwater vehicles cannot only consider the center of gravity of the underwater vehicle, multiple reference points should be selected from the contours in the four directions of up, down, left and right, and the spatial transformation relationship between the reference points and the center of gravity position should be constructed in the carrier coordinate system. After obtaining the motion prediction of the center of gravity position, combined with the attitude angle sequence and the spatial structure of the vehicle, the three-dimensional trajectory prediction information of multiple reference points of the underwater vehicle contour can be obtained, thereby providing information support for situation assessment and navigation safety.

[0109] S7: After optimizing the initial parameters of the bidirectional long short-term memory network prediction model combined with the control command attention mechanism for each eigenmode function component through the improved particle swarm optimization algorithm, repeat steps S5 and S6 to realize the underwater vehicle motion prediction.

[0110] Since the selection of initial parameters of the bidirectional long short-term memory network prediction model of each eigenmode function component combined with the control command attention mechanism has an important influence on the training process and the final prediction effect, different initial parameters have different motion prediction accuracies. In order to reduce the influence of the initial parameters on the motion prediction and improve the accuracy of the underwater vehicle motion prediction as much as possible, the initial parameters of the bidirectional long short-term memory network prediction model of each eigenmode function component combined with the control command attention mechanism are updated and optimized by improving the particle swarm optimization algorithm, the training process of step S5 is repeated, the parameters of the bidirectional long short-term memory network prediction model of each eigenmode function component combined with the control command attention mechanism after training are updated, and the motion prediction process of step S6 is repeated, thereby realizing the accurate prediction of the underwater vehicle motion.

[0111] Specifically, the initial parameters of the bidirectional long short-term memory network prediction model combined with the control instruction attention mechanism of each intrinsic mode function component are optimized by improving the particle swarm optimization algorithm, including the following steps:

[0112] S71: representing the initial parameters to be optimized of the bidirectional long short-term memory network prediction model of each eigenmode function component combined with the control instruction attention mechanism as different dimensions of particles;

[0113] Specifically, it can include four dimensions, namely, time window size, number of hidden layer units, initial learning rate, and regularization parameter.

[0114] S72: Initialize relevant parameters of the improved particle swarm optimization algorithm;

[0115] The relevant parameters of the improved particle swarm optimization algorithm include the population size, the initial position and speed of the particles, the maximum number of algorithm iterations, the first learning factor, the second learning factor, the initial value of the inertia weight, the maximum value of the inertia weight, the minimum value of the inertia weight and the mutation probability of the particles.

[0116] S73: using the prediction error of the underwater vehicle motion prediction model based on deep learning error correction as the fitness function of the particle swarm optimization algorithm, and performing optimization iterations;

[0117] The fitness function of the particle swarm optimization algorithm is as follows:

[0118]

[0119] in: Indicates The fitness function of a particle is Indicates The root mean square error of each particle is Indicates The mean absolute percentage error of each particle;

[0120] In order to reduce the position coordinate system origin in trajectory prediction The influence of position vector can be used to replace the position scalar information for calculation. The calculation formula is as follows:

[0121] ;

[0122] in: represents the vector value of the actual three-dimensional trajectory at each forecast time, A vector value representing the three-dimensional trajectory prediction at each forecast time.

[0123] S74: Update the iterative particle update speed of the improved particle swarm optimization algorithm according to formula (5):

[0124] (5);

[0125] in: Indicates Particle No. Iteration particle update speed, Indicates Particle No. Iteration particle update speed, represents the inertia weight, represents the first learning factor, Represents a random number in the range of 0 to 1 generated by a random function. Indicates Particle No. The optimal value of the iteration, Indicates Particle No. The position of the particle in the iteration, represents the second learning factor, represents the global optimal value;

[0126] S75: Update the inertia weight of the improved particle swarm optimization algorithm according to formula (6):

[0127] (6);

[0128] in: represents the maximum value of the inertia weight, represents the minimum value of the inertia weight, represents the iterative arc factor, ,in: Indicates the current iteration number, Indicates the maximum number of iterations.

[0129] Generally, the probability of particle mutation is 0.001-0.1, and the position of the particle after particle mutation is as follows:

[0130] ,in: represents the position of the particle after the particle mutation, represents a random distribution with mean 0 and variance 1.

[0131] By introducing a random sine into the inertia weight, the inertia weight decreases very quickly, and the later stage is manifested as a local search, which can speed up the convergence of the algorithm.

[0132] In summary, the present invention provides a method for underwater vehicle motion prediction based on deep learning error correction. In order to solve the problems that the underwater data of underwater vehicles has large noise, many random features, and the change of control commands affects the accuracy of motion prediction, the performance of motion prediction is improved through adaptive variational mode decomposition, introduction of control command attention mechanism, design of deep learning prediction network model, initial parameter optimization, etc.

[0133] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for underwater vehicle motion prediction based on deep learning error correction, characterized in that: The steps include: S1: According to the motion characteristics of underwater vehicles, considering the impact of control commands on the future motion of underwater vehicles, the input and output of the underwater vehicle motion prediction model based on deep learning error correction is designed using the underwater vehicle historical data; S2: Perform motion prediction based on the input of the underwater vehicle motion prediction model based on deep learning error correction, obtain the motion prediction sequence, and calculate the error sequence between the motion prediction sequence and the real historical motion sequence; S3: Adaptively perform variational mode decomposition on the error sequence between the motion deduction sequence and the real historical motion sequence to obtain multiple eigenmode function component sequences; S4: After processing multiple eigenmode function component sequences with forward LSTM neural network and reverse LSTM neural network, a bidirectional LSTM network prediction model combined with the control instruction attention mechanism is constructed; S5: Based on multiple eigenmode function component sequences, a bidirectional long short-term memory network prediction model combined with a control instruction attention mechanism is trained to obtain a bidirectional long short-term memory network prediction model combined with a control instruction attention mechanism for each eigenmode function component; S6: Combine the motion deduction sequence and each eigenmode function component with the bidirectional long short-term memory network prediction model combined with the control command attention mechanism to perform underwater vehicle motion prediction based on deep learning error correction, and form an underwater vehicle motion prediction model based on deep learning error correction; S7: After optimizing the initial parameters of the bidirectional long short-term memory network prediction model combined with the control command attention mechanism for each eigenmode function component through the improved particle swarm optimization algorithm, repeat steps S5 and S6 to realize the motion prediction of the underwater vehicle.

2. The method for predicting underwater vehicle motion based on deep learning error correction according to claim 1, characterized in that: The input of the motion prediction model in step S1 is a sequence including the longitude, latitude, depth, heading, pitch, roll, lateral speed, longitudinal speed, vertical speed, heading angular velocity, pitch angular velocity, roll angular velocity and stern rudder command rudder angle of the underwater vehicle within a set time window, and the output of the motion prediction model is a sequence including the longitude, latitude, depth, heading, pitch and roll of the underwater vehicle within the time window from the current moment to the prediction moment.

3. The method for predicting underwater vehicle motion based on deep learning error correction according to claim 1, characterized in that: In step S2, motion prediction is performed based on the hydrodynamic model of the underwater vehicle. After the duration of the motion prediction is determined, the trajectory and attitude motion prediction based on the hydrodynamic model of the underwater vehicle is realized through the underwater vehicle open-loop control algorithm based on the Kalman filter algorithm to obtain the motion prediction sequence.

4. The method for predicting underwater vehicle motion based on deep learning error correction according to claim 1, characterized in that: When the hydrodynamic model of the underwater vehicle is missing in step S2, first determine the duration of the motion forecast, then obtain the rotation matrix of the underwater vehicle from the current moment to the forecast moment based on the angular velocity, and obtain the translation matrix of the underwater vehicle from the current moment to the forecast moment based on the velocity, and obtain the trajectory and attitude motion forecast information of the underwater vehicle at the forecast moment through rotation and translation integral positioning position calculation to form a motion deduction sequence.

5. The method for predicting underwater vehicle motion based on deep learning error correction according to claim 1, characterized in that: In step S3, the following method is used to perform adaptive variational mode decomposition on the error sequence between the motion deduction sequence and the real historical motion sequence: S31: According to formula (1), the error sequence between the motion deduction sequence and the real historical motion sequence is decomposed into multiple eigenmode function components: （1）； in: Indicates from Time has come The error sequence between the moment-by-moment motion sequence and the real historical motion sequence, Indicates from Time has come Moment decomposition eigenmode function components, Indicates from Time has come The residual sequence after moment decomposition, Indicates the forecast duration. represents the serial number of the decomposed eigenmode function component, Represents the total number of decomposed eigenmode function components; S32: Calculate the Spearman correlation coefficient according to formula (2): （2）； in: represents the Spearman correlation coefficient, Indicates from Time has come The error sequence between the moment motion deduction sequence and the real historical motion sequence The error of a moment, represents the mean of the error sequence between the motion deduction sequence and the real historical motion sequence, Indicates from Time has come After the moment of decomposition The residual sequence at each moment is represents the mean of the residual sequence after decomposition, Indicates the length of the sequence; S33: Compare the absolute value of the Spearman correlation coefficient with the set threshold value. If the absolute value of the Spearman correlation coefficient is greater than or equal to the set threshold value, continue the decomposition. If the absolute value of the Spearman correlation coefficient is less than the set threshold value, stop the decomposition.

6. The underwater vehicle motion prediction method based on deep learning error correction according to claim 5, characterized in that: In step S4, the forward long short-term memory neural network processing and the reverse long short-term memory neural network processing of the plurality of eigenmode function component sequences include the following steps: S41: normalizing each eigenmode function component in each eigenmode function component sequence according to formula (3), and normalizing the four stern rudder command rudder angle characteristics in the input of the underwater vehicle motion prediction model: （3）； in: represents the normalized sequence of the intrinsic mode function component sequence, represents the sequence of eigenmode function components, represents the minimum value in the sequence of components of the eigenmode function, represents the maximum value in the sequence of components of the eigenmode function; S42: History Time has come The normalized sequence of the eigenmode function component sequence at time Time has come The input of the underwater vehicle motion prediction model at time t is the normalized sequence of the command rudder angle characteristics of the four stern rudders. Time has come The normalized sequence of the eigenmode function components at each moment is used as the output for forward long short-term memory neural network processing; S43: The future Time has come The normalized sequence of the eigenmode function component sequence at the moment is taken as input, and the historical Time has come The sequence of normalized components of the eigenmode function at time Time has come The normalized sequence of the command rudder angles of the four stern rudders of the underwater vehicle at each moment is used as the output for inverse long short-term memory neural network processing.

7. The method for predicting underwater vehicle motion based on deep learning error correction according to claim 6, characterized in that: In step S4, a bidirectional long short-term memory network prediction model combined with the control instruction attention mechanism is constructed according to formula (4): （4）； in: represents the query vector, Represents a reverse long short-term memory neural network Output at the moment, represents the parameter matrix of the query vector, represents the key vector, Indicates that each eigenmode function component predicts the model in the future through a bidirectional long short-term memory network Time has come Hidden layer state information within the time period, represents the parameter matrix of the key vector, represents a value vector, Parameter matrix representing the value vector.

8. The method for predicting underwater vehicle motion based on deep learning error correction according to claim 7, characterized in that: In step S5, the bidirectional long short-term memory network prediction model combined with the control instruction attention mechanism is trained according to the following method: S51: After initializing the initial parameters of the bidirectional long short-term memory network prediction model combined with the control instruction attention mechanism, the bidirectional long short-term memory network prediction model combined with the control instruction attention mechanism is trained based on the eigenmode function component sequence to obtain the prediction model of the eigenmode function component: S52: Based on the forecast output information sequence and future Time has come The bi-norm design of the sequence of the eigenmode function components at each moment is combined with the loss function of the bidirectional long short-term memory network prediction model of the control instruction attention mechanism, and the prediction model of each eigenmode function component is trained based on the loss function; S53: After the training is completed, the prediction model of each eigenmode function component is output and saved.

9. The method for predicting underwater vehicle motion based on deep learning error correction according to claim 8, characterized in that: In step S6, the following method is used to form an underwater vehicle motion prediction model based on deep learning error correction: S61: forecasting each eigenmode function component in the future time based on the forecast model of the eigenmode function component, obtaining a forecast sequence of each eigenmode function component in the future time, and reconstructing the forecast sequences of all eigenmode function components in the future time into a reconstructed prediction error sequence; S62: sum the reconstructed prediction error sequence and the motion deduction sequence to obtain the underwater vehicle future time forecast sequence, perform underwater vehicle motion forecast based on deep learning error correction, and form an underwater vehicle motion forecast model based on deep learning error correction.

10. The underwater vehicle motion prediction method based on deep learning error correction according to claim 8, characterized in that: In step S7, the initial parameters of the bidirectional long short-term memory network prediction model combined with the control instruction attention mechanism of each eigenmode function component are optimized by using the improved particle swarm optimization algorithm, which includes the following steps: S71: representing the initial parameters to be optimized of the bidirectional long short-term memory network prediction model of each eigenmode function component combined with the control instruction attention mechanism as different dimensions of particles; S72: Initialize relevant parameters of the improved particle swarm optimization algorithm; S73: using the prediction error of the underwater vehicle motion prediction model based on deep learning error correction as the fitness function of the particle swarm optimization algorithm, and performing optimization iterations; S74: Update the iterative particle update speed of the improved particle swarm optimization algorithm according to formula (5): （5）； in: Indicates Particle No. Iteration particle update speed, Indicates Particle No. Iteration particle update speed, represents the inertia weight, represents the first learning factor, Represents a random number in the range of 0 to 1 generated by a random function. Indicates Particle No. The optimal value of the iteration, Indicates Particle No. The position of the particle in the iteration, represents the second learning factor, represents the global optimal value; S75: Update the inertia weight of the improved particle swarm optimization algorithm according to formula (6): （6）； in: represents the maximum value of the inertia weight, represents the minimum value of the inertia weight, Represents the iteration radian factor.

Images