Adaptive navigation method for autonomous underwater vehicle based on combination of deep learning and physical model

Abstract

The present application relates to the technical field of navigation, in particular to an AUV adaptive navigation method based on combination of deep learning and physical model. The method comprises the following steps: in the process of modeling AUV motion system, a weighted gain bias estimation EKF algorithm is proposed, the DVL velocity bias is taken as a state variable to be estimated online, the sensor system error is compensated, and a Kalman gain update strategy weighted according to state confidence is introduced; a deep learning network of adaptive measurement noise is constructed, the network is used to dynamically predict the DVL velocity x to time-varying noise standard deviation, and y to time-varying noise standard deviation; the time-varying noise standard deviation obtained in the above step is written into the measurement noise covariance matrix at the moment, and the obtained AUV position trajectory is calculated. The AUV navigation adaptability and navigation precision can be significantly improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of navigation, in particular to an AUV adaptive navigation method based on combination of deep learning and a physical model. BACKGROUND

[0002] Although a traditional Extended Kalman Filter (EKF) has been successfully applied in AUV navigation, its performance is heavily dependent on the accuracy of the system model and the measurement model, and their matching degree with noise covariance matrices (Q, R).

[0003] In AUV navigation tasks, the process noise covariance Q and the measurement noise covariance R are usually assumed to be constants, which are set by prior calibration or experience. However, in actual operation, the environment and sensor characteristics may change, for example, the Doppler Velocity Log (DVL) carried by the AUV is affected by various marine organisms and uneven water bottom reflection during observation, and its measurement noise will change with the environment, so that the fixed covariance does not match the true statistics, thereby destroying the optimality of the filter estimation.

[0004] In addition, when there are unmodeled system biases or environmental disturbances in the sensor, the traditional EKF modeling does not consider the error bias, which leads to the accumulation and even divergence of the error of the filter. The DVL zero drift bias is a prominent problem: in long-time applications, the forward and lateral velocities measured by the DVL may have a small constant error, which is similar to sensor zero drift or calibration bias. This error will be integrated into a significant position offset in inertial calculation. The uncorrected DVL velocity bias is equivalent to encountering an unknown constant ocean current during AUV navigation, which causes a continuous deviation between the estimated trajectory and the actual trajectory. SUMMARY

[0005] The application aims to overcome the above-mentioned defects in the prior art and provides an AUV adaptive navigation method based on combination of deep learning and a physical model, which can significantly improve the navigation adaptability and navigation accuracy.

[0006] The technical scheme of the application is as follows: an AUV adaptive navigation method based on combination of deep learning and a physical model, comprising the following steps:

[0007] S1. In the process of modeling the AUV motion system, a weighted gain bias estimation EKF algorithm is proposed, the DVL velocity bias is estimated online as a state variable, the sensor system error is compensated, and a Kalman gain update strategy weighted according to the state confidence is introduced;

[0008] S2, construct a deep learning network of adaptive measurement noise, utilize network dynamic prediction of x-direction time-varying noise standard deviation of DVL speed quantity , and y-direction time-varying noise standard deviation ;

[0009] S3, write the time-varying noise standard deviation obtained in step S2 into the measurement noise covariance matrix of time , and calculate the obtained AUV position trajectory . .

[0010] In the application, step S1 comprises the following steps:

[0011] S1.1, construct an AUV augmented navigation evolution model;

[0012] S1.2, construct an AUV navigation observation model;

[0013] S1.3, state confidence weighted Kalman gain.

[0014] The specific implementation process of step S1.1 is as follows:

[0015] S1.1.1, augment the state vector of EKF, and the augmented EKF state vector is as follows:

[0016] ,

[0017] Among them, , respectively, the plane coordinate position of AUV at time , x , y , , , , respectively, the motion speed of AUV in the carrier coordinate system at time , x , y , , respectively, the motion acceleration of AUV in the carrier coordinate system at time , x , y , , , , , x , , , ybias of the DVL's x-direction velocity measurement at time

[0018] S1.1.2, construct the augmented AUV navigation system one-step evolution model;

[0019] construct the augmented AUV navigation system one-step evolution model. from time to time . Wherein, is the time step:

[0020] ,

[0021] ,

[0022] ,

[0023] ,

[0024] ,

[0025] ,

[0026] ,

[0027] ,

[0028] ,

[0029] ,

[0030] wherein, , are the x-direction and y-direction plane coordinate positions of the AUV at time ; is the heading angle of the AUV at time ; , respectively represent the x-direction and y-direction motion velocities of the AUV at time in the AUV carrier coordinate system; , respectively represent the x-direction and y-direction motion accelerations of the AUV at time in the AUV carrier coordinate system; represents the z-axis angular velocity of the AUV at time ; is the bias of the DVL's x-direction velocity measurement at time ; is the bias of the DVL's y-direction velocity measurement at time .

[0031] The specific implementation process of step S1.2 is as follows:

[0032] S1.2.1, construct the augmented AUV navigation system observation vector ;

[0033] ,

[0034] wherein, is the AUV heading angle obtained by the INS observation; and are the x-direction and y-direction motion velocities observed by the DVL respectively; and are the x-direction and y-direction motion accelerations obtained by the INS respectively; is the z-axis angular velocity measured by the INS;

[0035] S1.2.2, based on the AUV navigation system observation vector obtained in step S1.2.1 , construct the AUV navigation observation model, and the corresponding observation model is as follows:

[0036] ,

[0037] ,

[0038] ,

[0039] ,

[0040] ,

[0041] ,

[0042] wherein, is the observation model heading angle; is the observation model x-direction velocity; is the observation model y-direction velocity; is the observation model x-direction acceleration; is the observation model y-direction acceleration; is the angular velocity of the measurement model heading angle; is the measurement noise of the heading angle; is the measurement noise of the x-direction motion velocity; is the measurement noise of the y-direction motion velocity; is the measurement noise of the x-direction motion acceleration; is the measurement noise of the y-direction motion acceleration; is the measurement noise of the z-axis angular velocity;

[0043] The corresponding measurement equation corresponds to the observation matrix For a 6x10 matrix:

[0044] ,

[0045] Observation noise covariance matrix is:

[0046] ,

[0047] where, is the standard deviation of the heading angle measurement of the INS; is the standard deviation of the velocity measurement of the DVL; represents the standard deviation of the acceleration measurement noise of the INS; represents the standard deviation of the angular velocity measurement of the INS.

[0048] The specific implementation process of step S1.3 is as follows:

[0049] In the filter update phase, the confidence weight matrix W of the state is introduced by weighting the Kalman gain, and the state confidence weight matrix is as follows, where n is the state dimension,

[0050] ,

[0051] Each diagonal element represents the confidence coefficient of the corresponding state; when filtering is updated, the standard Kalman gain K is multiplied by the W weight to obtain a new Kalman gain , which is used to perform subsequent state update:

[0052] .

[0053] Step S2 includes the following steps:

[0054] S2.1, time update;

[0055] The state estimate value of the previous time step is used to predict the state of the current time step:

[0056] ,

[0057] where, is the predicted prior system state;

[0058] The covariance of the prediction state estimation error is:

[0059] ,

[0060] where, is the predicted covariance; is the state transition matrix; the process noise covariance matrix at time ;

[0061] the state transition matrix is:

[0062] ;

[0063] S2.2, construct a deep learning network model of adaptive measurement noise, use the model to output the time-varying noise standard deviation of the DVL speed quantity.

[0064] The specific implementation process of step S2.2 is as follows:

[0065] S2.2.1, offline training data set acquisition;

[0066] S2.2.2, feature construction and standardization for noise causes;

[0067] After global standardization of the feature values of the training set, the model is trained as the input of the model, and the standardization formula is:

[0068] ,

[0069] wherein, is the multi-source sensor feature, and the multi-source sensor feature of the model input includes: constant term, x-axis speed, y-axis speed, z-axis speed detected by DVL, bottom height, bottom state, pitch angle, roll angle, x-axis acceleration, y-axis acceleration, z-axis acceleration, and z-axis angular velocity detected by IMU; is the mean value of ;

[0070] S2.2.3, input the standardized feature vector obtained in step S2.2.2 into the deep learning network model of adaptive measurement noise, dynamically predict the time-varying noise standard deviation of the DVL speed quantity , ;

[0071] The deep learning network model of adaptive measurement noise includes a channel interaction branch and a local time sequence branch arranged in parallel, and the standardized feature vector is input into the channel interaction branch and the local time sequence branch respectively;

[0072] The output of the channel interaction branch is connected with a squeeze-excitation attention module, and the output of the local time sequence branch is connected with a normalization and regularization module;

[0073] The output of the squeeze-excitation attention module and the output of the normalization and regularization module are respectively connected with a feature fusion module;

[0074] The feature fusion module, the gating network, the expert network, the weighted output module and the positive value constraint module are connected in series.

[0075] The data processing process of the adaptive measurement noise deep learning network model is as follows:

[0076] In the channel interaction branch, the input is first linearly upgraded to K dimensions, where is the hidden layer dimension, and the standardized feature vector is first mapped to a high-dimensional space through a fully connected layer:

[0077] ,

[0078] wherein, is the output vector of the first hidden layer of the deep learning network; is the weight matrix corresponding to the fully connected layer, , denotes the real number set; is the bias vector corresponding to the fully connected layer, ; and

[0079] The GELU activation function provides a smooth nonlinear transformation, and its expression is:

[0080] ,

[0081] wherein, is the cumulative distribution function of the standard normal distribution;

[0082] Subsequently, B residual MLP blocks are stacked, each residual MLP block comprising a LayerNorm layer, a first Dense layer, a GELU layer, a Dropout layer, a second Dense layer and a residual layer connected in series, and the data processing process of each residual MLP block is as follows:

[0083] ,

[0084] wherein, is the result of the input feature in the th residual block after layer normalization processing; is the original feature data input into the th residual block;

[0085] ,

[0086] wherein, is the output result of the first Dense layer;

[0087] ,

[0088] where, is the output vector of the Dropout layer;

[0089] ,

[0090] where, is the linear output result of the second Dense layer;

[0091] ,

[0092] where, is the new feature representation with residual information obtained after the residual sum and GELU activation of the th residual block;

[0093] The SE attention mechanism is introduced after the residual MLP block, which includes two fully connected layers, and the data processing process is as follows:

[0094] ,

[0095] where, is a D-dimensional vector, which is a channel-level global descriptor obtained after the squeezing operation on the feature map ; ; is the original feature map input to the SE attention module, which comes from the output of the residual MLP block;

[0096] ,

[0097] where, is the output weight vector; is the weight of the second fully connected layer, ; is the weight of the second fully connected layer, ; is the bias of the first fully connected layer; is the bias of the second fully connected layer; is the Sigmoid function;

[0098] ,

[0099] where, is the output weight vector re-calibrated output feature map; represents channel-wise multiplication;

[0100] The local timing branch includes two causally connected convolutional layers in series, which will standardize the feature vector Two layers of one-dimensional convolution are stacked in the input, each layer of convolution uses causal padding:

[0101] ,

[0102] For a sequence of length , causal padding fills zeros at the beginning of the sequence, the output channel number of the first layer of convolution and the second layer of convolution are both , and each layer of convolution is followed by a GELU activation function:

[0103] ,

[0104] wherein is the deep feature representation processed by two layers of one-dimensional convolution Conv1D and GELU activation function;

[0105] The normalization and regularization module includes a LayerNorm layer and a Dropout layer,

[0106] ,

[0107] wherein is the output feature processed by normalization and Dropout regularization;

[0108] The feature fusion module realizes the information fusion of the channel interaction branch and the local time sequence branch, splices the output of the channel interaction branch and the output of the local time sequence branch in the feature dimension, and fuses the information through a fusion layer:

[0109] ,

[0110] wherein is the fused feature vector; is the weight matrix of the fusion layer, ; is the bias matrix of the fusion layer, ;

[0111] The fused feature is input into two layers of gate networks to generate the weights of expert networks:

[0112] ,

[0113] wherein is the hidden layer output of the gate network; is the weight matrix of the first layer of gate network, ; ​is the bias matrix of the first layer gated network;

[0114] ,

[0115] where, is the weight matrix of the second layer gated network, ; is the bias matrix of the second layer gated network; is the expert weight vector with dimension , ;

[0116] The network model includes E lightweight expert networks connected in parallel, each of which is a two-layer MLP, and each expert network independently learns the mapping from the fused features to the noise standard deviation:

[0117] ,

[0118] where, is the hidden layer output of the th expert network; is the weight matrix of the first layer MLP of the th expert network; is the bias matrix of the first layer MLP of the th expert network;

[0119] ,

[0120] where, is the output of the th expert network, ; is the bias matrix of the second layer MLP of the th expert network;

[0121] The final noise standard deviation raw output is the weighted sum of the outputs of each expert network:

[0122] ,

[0123] To ensure that the output noise standard deviation is non-negative, a Softplus function is used for transformation:

[0124] ,

[0125] where, is the final time-varying noise standard deviation predicted by the deep learning network model with adaptive measurement noise, which has two dimensions, including ;

[0126] The final Applying hard constraints:

[0127] ,

[0128] wherein, is a minimum threshold value below which the predicted time-varying noise standard deviation is not allowed; is a maximum threshold value above which the predicted time-varying noise standard deviation is not allowed.

[0129] Write the time-varying noise standard deviation obtained in step S2 into the measurement noise covariance matrix of the AUV navigation system at time :

[0130] ,

[0131] wherein, is the measurement noise of the heading angle; is the measurement noise of the x-direction acceleration; is the measurement noise of the y-direction acceleration; is the measurement noise of the z-axis angular velocity;

[0132] Perform measurement update calculation using the sensor data and the measurement noise output by the deep learning network model of the adaptive measurement noise: calculate the innovation between the observation prediction and the actual observation:

[0133] ,

[0134] wherein, is the innovation vector, is the observation prediction value, is the AUV navigation system observation vector at time ;

[0135] Calculate the innovation vector covariance :

[0136] ,

[0137] wherein, denotes the observation matrix at time ; is the transpose matrix of the observation matrix at time .

[0138] Determine the Kalman gain according to the value of the innovation vector covariance:

[0139] ,

[0140] Perform weighted calculation on using the confidence weight to obtain the weighted Kalman gain ;

[0141] the navigation system state estimation value at the final time is:

[0142] .

[0143] the system state covariance estimation value at the final time is:

[0144] ,

[0145] wherein, is a unit diagonal matrix;

[0146] the AUV position trajectory obtained through iterative calculation of each time step .

[0147] The beneficial effects of the present application are:

[0148] (1) The present application proposes a weighted gain bias estimation EKF algorithm, which estimates the DVL velocity bias as a state variable online, compensates for the sensor system error, and introduces a Kalman gain update strategy weighted by state confidence to avoid excessive correction of modeling bias.

[0149] (2) A convolution-mixed expert network is proposed, which uses a deep learning network to regress the noise standard deviation in real time and write it into the measurement covariance R, realizes channel-level self-adaptation according to time and working conditions, and cooperates with Total Variation positive robust updating, and is used in the state estimation process of the WB-EKF algorithm, thereby significantly improving the navigation adaptability and navigation accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0150] Figure 1 is a flowchart of the method described in the present application;

[0151] Figure 2 is a structural diagram of a deep learning network model for adaptive measurement noise;

[0152] Figure 3 is a comparison chart of the AUV motion trajectory predicted by the method described in the present application and the AUV motion trajectory predicted by the existing EKF method. DETAILED DESCRIPTION

[0153] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0154] ​​In the following description specific details are set forth in order to provide a thorough understanding of the application. However, the application can be practiced without the specific details. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to obscure the application. Accordingly, the particular implementation described is meant to be illustrative only and not limiting as to the scope of the application.

[0155] The flow chart of the AUV adaptive navigation method based on deep learning and physical model combination described in the present application is shown in Figure 1 The process of the method is described as follows.

[0156] First, in the process of modeling the AUV motion system, a weighted Kalman-gain bias-estimating EKF algorithm (WB-EKF for short) is proposed, which estimates the DVL velocity bias as a state variable online, compensates for the sensor system error, and introduces a Kalman gain update strategy weighted by state confidence to avoid excessive correction of modeling bias. The specific implementation process is described as follows.

[0157] First, an augmented navigation evolution model of AUV is constructed, and the specific implementation process is described as follows.

[0158] (1) Augment the state vector of EKF.

[0159] The AUV is equipped with an inertial navigation system (INS for short) to obtain the azimuth angle, angular velocity and acceleration of the AUV, and the velocity of the AUV is obtained through the DVL. The navigation system state vector covers the navigation core physical quantities of the AUV including position, velocity, heading, acceleration and angular velocity, provides complete state dimension support for subsequent motion prediction, ensures that the model can truly map the underwater motion law of the AUV, avoids prediction deviation caused by state loss, and is a basic prerequisite for realizing accurate navigation.

[0160] The navigation system state vector at time is . Wherein, and represent the position state information of the AUV at time , and represent the velocity state information of the AUV at time , represents the heading state information of the AUV at time , and represent the acceleration state information of the AUV at time , is the angular velocity state information of the AUV at time the heading angle velocity state information of the AUV at time

[0161] Augment the navigation system state vector at time , the augmented EKF state vector is shown as follows:

[0162] ,

[0163] wherein, , are the planar coordinate positions of the AUV in the east and north directions at time x y is the heading angle of the AUV at time , respectively represent the motion velocities of the AUV in the east and north directions at time x y , respectively represent the motion accelerations of the AUV in the east and north directions at time x y represents the z-axis angular velocity of the AUV at time is the bias of the DVL velocity measurement in the east direction at time x is the bias of the DVL velocity measurement in the north direction at time y Augment , to the AUV state vector as state information, which enables the filter algorithm to learn and compensate the DVL zero drift bias or scale error during execution, so as to improve the estimation performance of the filter algorithm.

[0164] In view of the systematic deviation problems such as zero drift and scale error of the DVL in long-term underwater work, the online estimation and compensation of the deviation are realized through state augmentation, without the need for additional sensor calibration operation, which reduces the engineering maintenance cost, avoids the navigation precision decay caused by deviation accumulation, and significantly improves the reliability of long-range navigation of the AUV.

[0165] (2) Construct an augmented one-step evolution model of the AUV navigation system.

[0166] Based on the kinematic equation, an augmented one-step state evolution model of the AUV navigation system at time from time is constructed. Wherein,​​​​​​​​​​​​​​ For time step:

[0167] ,

[0168] ,

[0169] ,

[0170] ,

[0171] ,

[0172] ,

[0173] ,

[0174] ,

[0175] ,

[0176] ,

[0177] in, , AUV time Planar coordinate positions in the x and y directions; For a moment The heading angle; , These represent the times in the AUV carrier coordinate system at time [time]. The speed of motion in the x and y directions; , These represent the times in the AUV carrier coordinate system at time [time]. The accelerations in the x and y directions; Indicates the AUV at time angular velocity along the z-axis; For a moment The bias of the x-axis velocity measurement in DVL; For a moment The bias of the y-axis velocity measurement in DVL.

[0178] In the above model, position , The heading is obtained by integrating the AUV's own velocity after rotating it around the heading; Obtained by integrating angular velocity; velocity , It is obtained by integrating its own acceleration; and acceleration , angular velocity and DVL bias , is modeled as a first-order Gaussian-Markov process, which is considered as a constant in short time scale and is allowed to wander in long time scale. This design reflects our assumption that the acceleration and yaw rate of AUV are approximately constant in small time steps, but will slowly drift in longer time; similarly, the DVL bias is considered as a constant bias but is allowed to have small random variations to simulate the slow change of sensor zero drift over time.

[0179] Combining the actual motion characteristics of AUV such as short-time acceleration / angular velocity approximately constant, long-time slow drift, and sensor bias characteristics such as DVL zero drift slow change, the augmented navigation evolution model is optimized to make the model more suitable for the actual engineering scene, avoid the prediction error caused by the deviation of ideal model from actual motion, and improve the accuracy of state prediction.

[0180] Secondly, the AUV navigation observation model is constructed. The specific implementation process includes the following steps.

[0181] (I) Construct the augmented AUV navigation system observation vector .

[0182] ,

[0183] wherein, is the AUV heading angle obtained by the INS observation; and are the x-direction and y-direction motion velocities obtained by the DVL observation; and are the x-direction and y-direction motion accelerations obtained by the INS; is the z-axis angular velocity measured by the INS.

[0184] The augmented AUV observation vector is consistent with that before augmentation, and the main sensors involved in AUV navigation include the heading angle measurement, acceleration measurement, angular velocity measurement provided by the INS, and velocity measurement provided by the DVL.

[0185] Keeping the observation vector consistent with the original sensor output, no modification is needed for the sensor hardware or data acquisition process, which reduces the engineering deployment difficulty of the algorithm, and at the same time ensures that the observation data can directly match the augmented state model, avoiding additional errors introduced by data conversion.

[0186] (II) Based on the AUV navigation system observation vector obtained in step S1.2.1, the AUV navigation system observation model is constructed.

[0187] The corresponding observation model can be expressed as follows:

[0188] ,

[0189] ,

[0190] ,

[0191] ,

[0192] ,

[0193] ,

[0194] where, is the observed model heading angle; is the observed model x-velocity; is the observed model y-velocity; is the observed model x-acceleration; is the observed model y-acceleration; is the angular velocity of the measured model heading angle; is the measurement noise of the heading angle; is the measurement noise of the x-velocity; is the measurement noise of the y-velocity; is the measurement noise of the forward acceleration; is the measurement noise of the lateral acceleration; is the measurement noise of the z-axis angular velocity. The corresponding measurement equation corresponds to the observation matrix is a 6x10 matrix.

[0195] ,

[0196] In the observation matrix, each row corresponds to the linear relationship between the state and the observation in one of the above observation equations. For example, the second row indicates that the forward DVL reading is equal to the sum of the state and . As can be seen, by adding and to the state vector, the observation model can depict the systematic difference between the DVL measurement and the true state when there is a zero bias, so that the filter can estimate and compensate for the bias online, thereby achieving explicit modeling of the DVL zero drift error.

[0197] It should be noted that this model does not separately model the constant drift of the accelerometer and the gyroscope by expanding the dimension, but is implicitly included in , , ​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​The same idea can be extended to estimate the INS gyro bias and accelerometer bias, and different state augmentation schemes can be chosen according to the actual engineering requirements.

[0198] Observation noise covariance matrix As shown below, it is a diagonal matrix, and each element is the measurement noise variance of each sensor.

[0199] ,

[0200] Wherein, is the standard deviation of the INS heading angle measurement; is the standard deviation of the DVL velocity measurement; represents the INS acceleration measurement noise standard deviation; represents the INS heading angle velocity measurement standard deviation.

[0201] By explicitly associating the DVL measurement, the true velocity and the bias through the observation matrix, the problem of observation and state mismatch caused by ignoring the bias in the traditional model is solved, so that the filter can estimate the state and bias at the same time; The diagonal noise covariance matrix conforms to the characteristics of independent sensor noise in engineering, simplifies the calculation while ensuring the reasonableness of noise modeling, and provides reliable noise statistical information for filter update.

[0202] Third, state confidence weighted Kalman gain.

[0203] In the filter update phase, the state confidence weight matrix W is introduced to weight the Kalman gain to improve the filter performance.

[0204] State confidence weight matrix As shown below, where n is the state dimension,

[0205] ,

[0206] The state dimension of the model constructed in this application is 10, and each diagonal element represents the confidence coefficient of the corresponding state. When the filter is updated, the standard Kalman gain K is multiplied by W to obtain a new Kalman gain , which is used to perform subsequent state update.

[0207] .

[0208] For the state , the first row Kalman gain will be reduced, which is equivalent to reducing the amount of information obtained from the current observation, so as to make it more dependent on model prediction; and The smaller, the more "blunt" the response to observation residuals. In particular, if a state represents complete distrust in the observation's correction to the state, in which case the filter will not update the state but only propagate according to the model prediction. By choosing the values of each state's appropriately, one can to some extent suppress the estimation divergence of low-confidence states caused by observation errors. Note that the setting of the matrix should be balanced with the specific application and tuning experience: Too small may slow down the filter convergence or cause bias accumulation, and too large will not have an inhibitory effect.

[0209] For the scenario of the decline in the reliability of some sensors in the underwater complex environment, such as DVL lock loss and INS disturbance by turbulence, the weight matrix is used to adjust the observation confidence of each state flexibly, avoid the pollution of low-confidence observations to state estimation, and prevent filter divergence. At the same time, the correction effect of high-confidence observations is retained, the effective observation information is maximized on the premise of ensuring navigation stability, and the robustness of the algorithm in complex marine environments is improved.

[0210] Second, build a deep learning network for adaptive measurement noise, use the deep learning network to regress the parameter value of the measurement noise covariance matrix in real time, realize channel-level adaptation according to time and working conditions, cooperate with Total Variation (abbreviated as TV) positive robust update, and be used in the state estimation process of the WB-EKF algorithm, thereby significantly improving the navigation accuracy of the AUV.

[0211] The WB-EKF proposed in the first step is used as the main navigation architecture, which includes two main parts, the time update process and the measurement update process. The specific implementation process of this step is described as follows.

[0212] First, time update.

[0213] In the time update process, according to the augmented navigation evolution model of the AUV constructed in the first step, the WB-EKF uses the state estimation value of the previous time step to predict the state of the current time step:

[0214] ,

[0215] where is the predicted prior system state.

[0216] At the same time, the EKF predicts the covariance of the state estimation error according to the system noise model:

[0217] ,

[0218] where is the predicted covariance; is the state transition matrix; is the time of process noise covariance matrix, which quantifies the uncertainty of the system model itself.

[0219] The system noise model of EKF mainly contains two parts: process noise and observation noise. Among them, the process noise is used to describe the uncertainty of the system model; the observation noise is used to describe the uncertainty of the sensor measurement.

[0220] The state transition matrix is:

[0221] ,

[0222] The dynamic correlation between states is quantified by the state transition matrix, the current state prediction based on historical states is realized, and the navigation continuity problem when the underwater sensor data transmission is delayed or temporarily interrupted is solved; At the same time, the prediction state error covariance provides the basis for the calculation of the Kalman gain of the subsequent measurement update, ensures the scientificity of the filter update, and avoids blindly fusing observation data.

[0223] Second, an adaptive measurement noise deep learning network model is constructed, and the time-varying noise standard deviation of the DVL velocity quantity is output by using the model.

[0224] (1) Off-line training data set collection.

[0225] In order to realize the training of the adaptive noise estimation model, it is necessary to construct an AUV navigation data set covering multiple scenarios. The specific collection process is that the AUV executes a variety of preset trajectory tasks in the marine environment, and synchronously collects the following data through the integrated sensor system:

[0226] 1. Attitude information: the pitch angle, roll angle and heading angle of the AUV are obtained through the inertial measurement unit (abbreviated as IMU);

[0227] 2. Motion information: the Doppler velocity log (abbreviated as DVL) provides the three-axis velocity in the AUV carrier coordinate system and the height data relative to the bottom;

[0228] 3. Kinetic information: the three-axis linear acceleration measured by the IMU, with the unit of m / s²; the z-axis angular velocity, with the unit of rad / s;

[0229] 4. Position reference: the GPS receiver provides the two-dimensional position in the UTM coordinate system east and north as the trajectory true value.

[0230] The collected data is preprocessed. To ensure time alignment, all sensor data is interpolated and synchronized based on the IMU timestamp to ensure consistency. To avoid the influence of invalid values on model training, linear interpolation is used to fill in missing or abnormal data such as DVL lock loss and GPS signal interruption, and a valid mask is marked in the data. The data set is divided, and the collected data is divided into training set, validation set and test set according to 7:2:1, ensuring diversity.

[0231] The multi-scene data set covers different marine environments such as calm sea area, turbulent area, complex terrain, and motion trajectory, ensuring the generalization ability of the model; time synchronization solves the time sequence misalignment problem caused by the sampling rate difference of multiple sensors; abnormal data processing avoids invalid information interference with model training; 7:2:1 division conforms to machine learning engineering practice, ensuring sufficient model training and effective verification, providing high-quality model support for online adaptive noise estimation.

[0232] (II) Problem modeling and objective function.

[0233] We want to learn a time-varying measurement noise standard deviation vector that only depends on past and current information, and write it into the measurement noise covariance matrix at time :

[0234] ,

[0235] ,

[0236] where is the x-direction velocity time-varying noise standard deviation, which describes the accuracy of the DVL measured x-direction velocity; is the y-direction velocity time-varying noise standard deviation, which describes the accuracy of the DVL measured y-direction velocity; is the multi-source sensor feature, the multi-source sensor feature of the model input includes: constant term, DVL detected x-axis velocity, DVL detected y-axis velocity, DVL detected z-axis velocity, DVL detected height, DVL detected bottom state, pitch angle, roll angle, IMU detected x-axis acceleration, IMU detected y-axis acceleration, IMU detected z-axis acceleration, z-axis angular velocity; is a learnable parameter; represents a deep learning network model of adaptive measurement noise; is the measurement noise of the heading angle; is the measurement noise of the x-direction acceleration; is the measurement noise of the y-direction acceleration; is the measurement noise of the z-axis angular velocity, that is, the measurement noise of the heading angular velocity.

[0237] Network output uses activation function To ensure a positive standard deviation, the EKF forward pass uses the Cholesky solution and Joseph form to guarantee numerical robustness.

[0238] ,

[0239] The AUV position trajectory calculated by WB-EKF is denoted as... The objective function is shown in the following equation:

[0240] ,

[0241] Where M is the effective quantity of GPS information; the TV item is... Weighting can prevent over-smoothing; These are the time smoothing regularization coefficients. The parameter is the regularization coefficient.

[0242] The objective function described above primarily aims to optimize the RMSE between the WB-EKF predicted trajectory and the GPS 2D trajectory, incorporating time smoothing regularization (TV regularization) and parametric regularization (L2 regularization). The time smoothing regularization term penalizes abrupt changes in noise estimates, calculated according to the time step. Weighting is used to avoid over-smoothing. In this embodiment, Pick The parameter regularization term controls model overfitting. In this embodiment, Pick .

[0243] The primary objective, RMSE, is directly linked to the core navigation performance metric, positional accuracy, ensuring that the model's learning direction aligns with engineering requirements. TV regularization prevents filter oscillations caused by abrupt changes in noise estimation, aligning with the slowly varying noise characteristics of the marine environment. L2 regularization prevents overfitting, ensuring stable operation even in unfamiliar scenarios. Softplus activation and Cholesky solutions ensure numerical stability, preventing algorithm crashes due to numerical anomalies during deployment.

[0244] (III) Characterization and standardization based on the causes of noise.

[0245] The feature values ​​of the training set are globally standardized before being used as input to train the model. The standardization formula is as follows:

[0246] ,

[0247] in, for The mean.

[0248] The standardized statistics are saved together with the model, and are strictly reused during reasoning to avoid distribution drift between training and testing.

[0249] The feature vectors comprehensively cover key factors affecting DVL noise, such as echo quality reflecting the state of the bottom and attitude angle reflecting the geometric relationship, enabling the model to actively predict noise changes rather than passively respond to residuals; global standardization eliminates dimensional differences in features such as speed m / s and angle rad, avoiding convergence difficulties caused by imbalanced feature scales during model training, and reuse of standardized statistics prevents distribution drift, ensuring consistency between online reasoning and offline training, and improving model prediction accuracy.

[0250] (Four) input the standardized feature vectors obtained in the above steps into the deep learning network model for adaptive measurement noise to dynamically predict the x-direction time-varying noise standard deviation of the DVL velocity quantity and the y-direction time-varying noise standard deviation .

[0251] The deep learning network model for adaptive measurement noise adopts a dual-branch hybrid expert architecture, aiming to capture both cross-channel nonlinear interactions and local temporal dependency patterns of sensor features, and strictly follows the principle of causality in overall design to ensure that only historical and current information is relied on during online deployment.

[0252] The deep learning network model for adaptive measurement noise is shown in Figure 2 . The specific construction process is as follows. The network model includes a channel interaction branch and a local time sequence branch arranged in parallel, and the standardized feature vectors are input into the channel interaction branch and the local time sequence branch, respectively. The output of the channel interaction branch is connected with a squeeze-and-excitation attention module, and the output of the local time sequence branch is connected with a normalization and regularization module. The output of the squeeze-and-excitation attention module and the output of the normalization and regularization module are connected with a feature fusion module, respectively. The feature fusion module, the gating network, the expert network, the weighted output module, and the positive constraint module are connected in series.

[0253] 1. Channel interaction branch.

[0254] The core goal of the channel interaction branch is to learn the complex nonlinear combination relationship between different sensor features, such as the synergistic effect of DVL velocity and IMU attitude angle on noise.

[0255] In the channel interaction branch, the input is first linearly upgraded to K dimensions, where is the hidden layer dimension.

[0256] The standardized feature vectors are first mapped to a high-dimensional space through a fully connected layer to enhance the model's expression ability:

[0257] ,

[0258] wherein, is the output vector of the first hidden layer of the deep learning network; is the weight matrix corresponding to the fully connected layer, , denotes the set of real numbers; is the bias vector corresponding to the fully connected layer, .

[0259] The GELU activation function provides a smooth nonlinear transformation, whose expression is:

[0260] ,

[0261] wherein, is the cumulative distribution function of the standard normal distribution.

[0262] Subsequently, B residual MLP blocks are stacked, each of which contains a LayerNorm layer, a first Dense layer, a GELU layer, a Dropout layer, a second Dense layer, and a residual layer connected in sequence, which are used to learn the nonlinear combination across channels. In this embodiment, B = 3, and the feature is processed through identical residual MLP blocks.

[0263] The data processing process of each residual MLP block is as follows:

[0264] ,

[0265] wherein, is the result of the input feature in the th residual block after layer normalization processing; is the original feature data input into the th residual block.

[0266] ,

[0267] wherein, is the output result of the first Dense layer.

[0268] ,

[0269] wherein, is the output vector of the Dropout layer.

[0270] ,

[0271] wherein, is the linear output result of the second Dense layer.

[0272] ,

[0273] wherein, is the new feature representation with residual information obtained after the residual sum and GELU activation of the first residual block.

[0274] The residual MLP block can effectively alleviate the gradient vanishing problem in deep networks; at the same time, the residual connection allows the gradient to be directly backpropagated, ensuring the effective training of deep networks.

[0275] The output of the channel interaction branch is connected with the squeeze-and-excitation attention module. After the residual MLP block, the SE attention mechanism is introduced to mimic the channel attention mechanism, allowing the network to automatically learn the importance weight of each feature channel. The SE attention mechanism includes two fully connected layers. The data processing process is described as follows.

[0276] ,

[0277] wherein, is a D-dimensional vector, which is a channel-level global descriptor obtained after the squeeze operation on the feature map , ; is the original feature map input to the SE attention module, which comes from the output of the residual MLP block.

[0278] ,

[0279] wherein, is the output weight vector; is the weight of the second fully connected layer, ; is the weight of the second fully connected layer, ; is the bias of the first fully connected layer; is the bias of the second fully connected layer; is the Sigmoid function.

[0280] ,

[0281] wherein, is the output weight vector the output feature map after recalibration; represents the channel-wise multiplication.

[0282] The squeeze-and-excitation attention module can effectively improve the contribution of noise-sensitive feature channels and suppress unimportant channels.​

[0283] The residual MLP block can solve the deep network gradient vanishing problem, ensuring that the model can learn complex nonlinear feature correlations such as the synergistic effect of DVL speed and IMU attitude. The SE attention mechanism automatically focuses on key feature channels sensitive to noise, suppresses redundant information, and improves the efficiency of model feature extraction. The GELU activation function provides a smooth nonlinear transformation, enhancing the model's expression ability, and the overall design ensures that the model can accurately capture the complex correlations between multi-sensor features, providing high-quality feature support for noise prediction.

[0284] 2. Local time sequence branch.

[0285] The local time sequence branch aims to capture patterns with local time correlation such as short-term degradation of DVL signals such as beam instantaneous loss of lock. The branch includes two causally connected convolutional layers.

[0286] The normalized feature vector is input into a stacked two-layer one-dimensional convolution. To ensure causality, i.e., the current output only depends on the current and past inputs, each layer of convolution uses causal padding:

[0287] ,

[0288] For a sequence of length , causal padding will fill zeros at the beginning of the sequence, ensuring that the output sequence has the same length as the input sequence and that the calculation at each time point does not depend on future information. The output channel number of the first layer of convolution is , and the output channel number of the second layer of convolution is also . Each layer of convolution is followed by a GELU activation function:

[0289] ,

[0290] where is the deep feature representation after two layers of one-dimensional convolution Conv1D and GELU activation function processing.

[0291] Causal convolution ensures that it does not rely on future data when deployed online, meeting the real-time requirements of navigation systems. The 5th order kernel size can effectively capture the local time sequence patterns of short-term signal degradation of DVL, such as beam instantaneous loss of lock, solving the prediction lag problem when noise suddenly changes.

[0292] 3. Normalization and regularization module.

[0293] The normalization and regularization module includes a LayerNorm layer and a Dropout layer to stabilize training and prevent overfitting.

[0294] ,

[0295] wherein, is the output feature after normalization and dropout regularization.

[0296] The LayerNorm layer and the Dropout layer can improve the training stability and generalization ability of the model, avoid short-term noise prediction failure caused by overfitting, and ensure that the model can respond to the instantaneous changes of the sensor signal in time.

[0297] 4. Feature fusion module.

[0298] Through this module, information fusion of the channel interaction branch and the local time sequence branch is realized.

[0299] In this module, the output of the channel interaction branch and the output of the local time sequence branch are spliced in the feature dimension, and information fusion is performed through a fusion layer:

[0300] ,

[0301] wherein, is the fused feature vector; is the weight matrix of the fusion layer, ; is the bias matrix of the fusion layer.

[0302] 5. Gating network.

[0303] The fused feature is input into two layers of gating network to generate the weight of expert network:

[0304] ,

[0305] wherein, is the hidden layer output of the gating network; is the weight matrix of the first layer of gating network, ; is the bias matrix of the first layer of gating network.

[0306] ,

[0307] wherein, is the weight matrix of the second layer of gating network, ; is the bias matrix of the second layer of gating network; is the expert weight vector, with a dimension of ,​ .

[0308] Gated networks learn how to dynamically combine the opinions of different experts based on the current input.

[0309] 6. Expert network.

[0310] The network model comprises E parallel-connected lightweight expert networks, each of which is a two-layer MLP. Each expert network independently learns a mapping from fused features to the noise standard deviation. In this embodiment, In other words, the network system contains three lightweight expert networks.

[0311] ,

[0312] in, For the first Hidden layer output of an expert network; For the first The weight matrix of the first layer MLP of an expert network; For the first The bias matrix of the first layer MLP of an expert network.

[0313] ,

[0314] in, For the first The output of an expert network, ; For the first The bias matrix of the second layer MLP of an expert network.

[0315] 7. Weighted output module.

[0316] Final noise standard deviation raw output It is a weighted sum of the outputs from each expert network:

[0317] ,

[0318] The adaptive measurement noise deep learning network model proposed in this application allows the network to learn more complex functions, and different expert networks can focus on different regions of the input space.

[0319] Dual-branch feature fusion integrates cross-channel nonlinear correlations and local temporal patterns, providing comprehensive feature support for noise prediction; the gated MoE architecture allows different experts to focus on different scenarios such as calm waters, turbulent areas, and DVL lockout recovery periods, with dynamic weighting to achieve scenario adaptation and improve the model's generalization ability in complex and variable marine environments; the lightweight expert network can balance model performance and computational load, ensuring that the real-time requirements of AUV navigation systems can be met when deployed online.

[0320] 8. Positive constraint module.

[0321] Network raw output It can be any real number. To ensure that the noise standard deviation of the output is non-negative, the Softplus function is used for transformation:

[0322] ,

[0323] where, is the final time-varying noise standard deviation predicted by the deep learning network model with adaptive measurement noise, which has two dimensions, respectively, , replacing the fixed measurement noise in the existing EFK.

[0324] The Softplus function is a smooth version of ReLU, which can ensure that the output is always positive. Adding is to prevent numerical underflow and ensure stability.

[0325] In engineering implementation, to prevent extreme values such as too large or too small from polluting the measurement noise covariance matrix of the WB-EKF, a hard constraint is imposed on the final :

[0326] ,

[0327] where, is the minimum threshold value, and the predicted noise standard deviation is not allowed to be lower than this value; is the maximum threshold value, and the predicted noise standard deviation is not allowed to exceed this value.

[0328] The processing steps implemented by this module are key measures to ensure the numerical robustness of the filter.

[0329] The activation function ensures that the noise standard deviation is positive, consistent with the physical meaning, and avoids the failure of the filtering model due to negative noise values; the offset prevents numerical underflow and ensures calculation stability; the numerical guardrail limits extreme values, avoiding excessive / large noise estimates polluting the covariance matrix, preventing filter divergence or precision decay, which is a key guarantee for algorithm engineering landing, ensuring long-term stable operation of the navigation system.

[0330] (5) Online testing.

[0331] The AUV performs tasks on the sea, and real-time collects navigation data of attitude sensors, GPS, DVL and the like. The deep learning network model of adaptive measurement noise obtained through offline training is loaded, and the collected AUV attitude, AUV acceleration and DVL state data in a unit time are input into the above network model, so as to calculate the measurement noise values of the predicted DVL x-direction velocity and y-direction velocity at each time step. The corresponding measurement noise values output by the network model are written into the measurement noise covariance matrix for subsequent filtering update process.

[0332] The online adaptive estimation of measurement noise is realized, without manual intervention to adjust noise parameters, and the dynamic changes of the ocean environment such as entering the turbulent area from the calm sea area and the change of the DVL bottom state are adapted. The covariance matrix is updated in real time, so that the filtering algorithm can dynamically match the current sensor noise characteristics, avoid the decline of navigation precision caused by fixed noise parameters, and significantly improve the navigation adaptability and reliability of the AUV in complex and changeable ocean environment.

[0333] Third, measurement update.

[0334] After predicting the velocity measurement noise of the DVL, the WB-EKF performs measurement update calculation by using sensor data and the measurement noise output by the adaptive measurement noise deep learning network model. The innovation between the observation prediction and the actual observation is calculated:

[0335] ,

[0336] wherein, is the innovation vector, is the observation prediction value, is the AUV navigation system observation vector at time .

[0337] At the same time, the innovation vector covariance is calculated:

[0338] ,

[0339] wherein, represents the observation matrix at time ; is the transpose matrix of the observation matrix at time .

[0340] According to the value of the innovation vector covariance, the Kalman gain can be determined:

[0341] ,

[0342] The confidence weight is used to The weighted calculation is performed to obtain a weighted Kalman gain .

[0343] The navigation system state estimation value at the final time is:

[0344] .

[0345] The system state covariance estimation value at the final time is:

[0346] ,

[0347] wherein, is a unit diagonal matrix.

[0348] After that, through the iterative calculation of each time step, the application can realize accurate and adaptive AUV navigation.

[0349] Combined with the adaptive prediction noise covariance and state confidence weighted gain, the optimal fusion of observation data and predicted state is realized, and the state estimation error is dynamically corrected; each step iteration forms a closed loop optimization, ensures that the navigation error is not accumulated, meets the engineering requirements of AUV long-range and high-precision navigation, such as the position accuracy requirements of underwater exploration, target tracking, pipeline detection and other tasks; the weighted gain further improves the estimation robustness, avoids the navigation fluctuation caused by abnormal observation, and ensures the stability of the navigation system.

[0350] As shown in Figure 3 , by comparing the AUV motion trajectory predicted by the WB-EKF method proposed in the application with the AUV motion trajectory predicted by the existing EKF method, it can be clearly obtained that the AUV motion trajectory predicted by the application is closer to the real trajectory of the AUV observed by GPS.

[0351] ​​The above describes in detail the AUV adaptive navigation method provided by the present application based on the combination of deep learning and a physical model. The principles and implementation manners of the present application are described by applying specific examples, and the above description of the examples is only used to help understand the method of the present application and its core idea. It should be pointed out that, for those skilled in the art, some improvements and modifications can be made to the present application without departing from the principles of the present application, and these improvements and modifications also fall within the protection scope of the claims of the present application. The above description of the disclosed examples enables those skilled in the art to implement or use the present application. Various modifications of the examples will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other examples without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the examples shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An AUV adaptive navigation method based on a combination of deep learning and physical models, characterized in that, Includes the following steps: S1. In the process of modeling the motion system of AUV, the weighted gain deviation estimation EKF algorithm is proposed, which uses the DVL velocity bias as the state variable for online estimation, to compensate for the sensor system error, and introduces a Kalman gain update strategy weighted by state confidence. S2. Construct a deep learning network that adapts to measurement noise, and use the network to dynamically predict the x-axis time-varying noise standard deviation of the DVL velocity. and the standard deviation of time-varying noise in the y-direction. ; S3. Write the time-varying noise standard deviation obtained in step S2 into the time interval. Measurement noise covariance matrix And calculate the AUV position trajectory ; Step S1 includes the following steps: S1.1 Constructing an AUV augmented navigation evolution model; S1.2, Construct an AUV navigation observation model; S1.3, State confidence weighted Kalman gain; The specific implementation process of step S2 is as follows: S2.1, Time Update; The state estimate from the previous time step is used to predict the state at the current time step: ， in, The predicted prior system state; Covariance of the prediction state estimation error: ， in, To predict covariance; This is the state transition matrix; For a moment The process noise covariance matrix; State transition matrix for: ； S2.2 Construct a deep learning network model for adaptive measurement noise, and use this model to output the time-varying noise standard deviation of the DVL velocity quantity; The specific implementation process of step S2.2 is as follows: S2.2.1 Offline training dataset collection; S2.2.2, Feature construction and standardization based on noise causes; The feature values ​​of the training set are globally standardized before being used as input to train the model. The standardization formula is as follows: ， in, For multi-source sensor features, the multi-source sensor features input to the model include: constant terms, x-axis velocity, y-axis velocity, z-axis velocity, ground clearance, ground clearance, pitch angle, roll angle detected by DVL, and x-axis acceleration, y-axis acceleration, z-axis acceleration, and z-axis angular velocity detected by IMU; for The mean; S2.2.3, The standardized feature vector obtained in step S2.2.

2. Input a deep learning network model that adapts to measurement noise, and dynamically predict the time-varying noise standard deviation of DVL velocity quantities. , ; The deep learning network model for adaptive measurement noise includes a channel interaction branch and a local temporal branch arranged in parallel. The standardized feature vectors are input into the channel interaction branch and the local temporal branch, respectively. The output of the channel interaction branch is connected to the squeezing-excitation attention module, and the output of the local temporal branch is connected to the normalization and regularization module; The outputs of the squeeze-excitement attention module and the normalization and regularization modules are connected to the feature fusion module, respectively. The feature fusion module, gating network, expert network, weighted output module, and positive value constraint module are connected in series.

2. The AUV adaptive navigation method based on a combination of deep learning and physical models according to claim 1, characterized in that, The specific implementation process of step S1.1 is as follows: S1.1.

1. Augment the EKF state vector. The augmented EKF state vector is shown below: ， in, , AUV time Planar coordinate positions in the x and y directions; For a moment The heading angle; , These represent the times in the AUV carrier coordinate system at time [time]. The speed of motion in the x and y directions; , These represent the times in the AUV carrier coordinate system at time [time]. The accelerations in the x and y directions; Indicates the AUV at time z-axis angular velocity; For a moment The bias of the x-axis velocity measurement in DVL; For a moment The bias of the y-axis velocity measurement in DVL; S1.1.2 Construct a one-step evolution model of the augmented AUV navigation system; Constructing the augmented AUV navigation system From moment to moment A one-step state evolution model; where, For time step: ， ， ， ， ， ， ， ， ， ， in, , AUV time Planar coordinate positions in the x and y directions; For a moment The heading angle; , These represent the times in the AUV carrier coordinate system at time [time]. The speed of motion in the x and y directions; , These represent the times in the AUV carrier coordinate system at time [time]. The accelerations in the x and y directions; Indicates the AUV at time z-axis angular velocity; For a moment The bias of the x-axis velocity measurement in DVL; For a moment The bias of the y-axis velocity measurement in DVL.

3. The AUV adaptive navigation method based on a combination of deep learning and physical models according to claim 1, characterized in that, The specific implementation process of step S1.2 is as follows: S1.2.1 Constructing the augmented AUV navigation system observation vector ; ， in, The heading angle of the AUV obtained from INS observations; and These are the x- and y-direction velocities obtained from DVL observations, respectively. and These are the x- and y-axis accelerations obtained from the INS, respectively. The z-axis angular velocity is measured by INS; S1.2.2, Based on the AUV navigation system observation vector obtained in step S1.2.1 An AUV navigation observation model is constructed, and the corresponding observation model is shown below: ， ， ， ， ， ， in, To observe the heading angle of the model; For the x-axis velocity of the observation model; To observe the y-axis velocity of the model; The x-axis acceleration of the observation model; For the y-axis acceleration of the observation model; To measure the angular velocity of the model's heading angle; The measurement noise is for the heading angle; The measurement noise is for the velocity in the x-direction. The noise in the measurement of the y-axis velocity; The measurement noise for the x-axis acceleration; The measurement noise for the y-axis acceleration; The measurement noise for the z-axis angular velocity; The observation matrix corresponding to the measurement equation A 6×10 matrix: ， Observation noise covariance matrix for: ， in, The standard deviation of the heading angle of the INS is measured; The standard deviation of the speed measurement in DVL; This represents the standard deviation of the INS acceleration measurement noise. This represents the standard deviation of the INS heading angular velocity measurement.

4. The AUV adaptive navigation method based on a combination of deep learning and physical models according to claim 1, characterized in that, The specific implementation process of step S1.3 is as follows: During the filtering update phase, the Kalman gain is weighted by a state-specific confidence weight matrix W. As shown in the following formula, where n is the state dimension, ， Each diagonal element This represents the confidence coefficient for the corresponding state; during filter updates, the standard Kalman gain K is weighted and multiplied by W to obtain the new Kalman gain. Used to perform subsequent state updates: 。 5. The AUV adaptive navigation method based on a combination of deep learning and physical models according to claim 1, characterized in that, The data processing procedure for the deep learning network model that adapts to measurement noise is as follows: In the channel interaction branch, input First, it undergoes linear dimensionality increase to K dimensions, where The standardized feature vector represents the hidden layer dimension. First, a fully connected layer is used to map to a high-dimensional space: ， in, This is the output vector of the first hidden layer of the deep learning network; This is the weight matrix corresponding to the fully connected layer. , Represents the set of real numbers; This is the bias vector corresponding to the fully connected layer. ; The GELU activation function provides a smooth nonlinear transformation, and its expression is: ， in, The cumulative distribution function of the standard normal distribution; Subsequently, B residual MLP blocks are stacked. Each residual MLP block contains a LayerNorm layer, a first Dense layer, a GELU layer, a Dropout layer, a second Dense layer, and a residual layer connected in sequence. The data processing procedure for each residual MLP block is as follows: ， in, For the first The result of the input features in each residual block after layer normalization; For input to the first The original feature data in each residual block; ， in, This is the output of the first Dense layer; ， in, This is the output vector of the Dropout layer; ， in, This is the linear output of the second Dense layer; ， in, For the first A new feature representation with residual information is obtained by summing the residuals and activating the GELU of each residual block; Following the residual MLP block, an SE attention mechanism is introduced. The SE attention mechanism consists of two fully connected layers, and its data processing procedure is as follows: ， in, It is a D-dimensional vector, which is a representation of the feature map. The channel-level global descriptor obtained after the compression operation. ; It is the original feature map input to the SE attention module, which comes from the output of the residual MLP block; ， in, To output the weight vector; The weights of the second fully connected layer, ; The weights of the first fully connected layer. ; The bias for the first fully connected layer; For the bias of the second fully connected layer; For the Sigmoid function; ， in, For the output weight vector The output feature map after recalibration; This represents channel-by-channel multiplication; The local temporal branch consists of two cascaded causal convolutional layers that convert the standardized feature vectors into a single pass. The input consists of two stacked one-dimensional convolutional layers, each with causal padding: , For length of For sequences, causal padding will pad the beginning of the sequence. The number of output channels for both the first and second convolutional layers is zero. Each convolutional layer is followed by the GELU activation function: ， in, This is the deep feature representation after processing with two layers of one-dimensional convolution Conv1D and GELU activation function; The normalization and regularization module includes LayerNorm and Dropout layers. ， in, for Output features after normalization and Dropout regularization; The feature fusion module integrates information from the channel interaction branch and the local timing branch, and converts the output of the channel interaction branch into a single data structure. and the output of local timing branches The data is concatenated along the feature dimension and then fused using a fusion layer. ， in, The fused feature vector; This is the weight matrix of the fusion layer. ; The bias matrix of the fusion layer. ; Features after fusion Input a two-layer gating network and generate Weights of each expert network: ， in, This is the output of the hidden layer of the gated network; This is the weight matrix of the first-layer gated network. ; This is the bias matrix of the first-layer gated network; ， in, This is the weight matrix of the second-layer gated network. ; This is the bias matrix for the second-layer gated network; The expert weight vector has dimensions of . , ; The network model consists of E parallel-connected lightweight expert networks, each of which is a two-layer MLP. Each expert network independently learns a mapping from fused features to the noise standard deviation. ， in, For the first Hidden layer output of an expert network; For the first The weight matrix of the first layer MLP of an expert network; For the first The bias matrix of the first layer MLP of an expert network; ， in, For the first The output of an expert network, ; For the first The bias matrix of a second-layer MLP in an expert network; Final noise standard deviation raw output It is a weighted sum of the outputs from each expert network: ， To ensure that the output noise standard deviation is non-negative, the Softplus function is used for transformation: ， in, The final time-varying noise standard deviation predicted by a deep learning network model that adaptively measures noise has two dimensions, including ; For the final Apply hard constraints: ， in, This is the minimum threshold value; the predicted time-varying noise standard deviation must not be lower than this value. This is the maximum threshold value; the predicted time-varying noise standard deviation is not allowed to exceed this value.

6. The AUV adaptive navigation method based on a combination of deep learning and physical models according to claim 1, characterized in that, Write the time-varying noise standard deviation obtained in step S2 into the time. Measurement noise covariance matrix : ， in, The measurement noise is for the heading angle; The measurement noise for acceleration in the x-axis; The measurement noise is for the y-axis acceleration. The measurement noise for the z-axis angular velocity; Measurement update calculations are performed using sensor data and measurement noise output from a deep learning network model that adapts to measurement noise: This calculates the information between the predicted and actual observations. , in, For the innovation vector, For the observed predicted value, For a moment AUV navigation system observation vector; Calculate the covariance of the innovation vector : ， in, Indicates time The observation matrix; For a moment The transpose of the observation matrix; Determine the Kalman gain based on the value of the innovation vector covariance. : ， Using confidence weights The weighted Kalman gain is obtained by performing a weighted calculation. ; Final moments Navigation system state estimate for: ， Final moments System state covariance estimate for: , in, It is a unit diagonal matrix; The AUV position trajectory is obtained through iterative calculation at each time step. .

Images