AUV navigation method based on Transformer model

Abstract

This invention relates to the field of AUV navigation technology, and in particular to an AUV navigation method based on the Transformer model. It includes the following steps: S1, obtaining observation data of the AUV through its own sensors; the observations at each time step constitute an observation vector; and using the time series composed of N unit time observation vectors as the input ψ of the Transformer model. input S2, the time series consisting of N observation vectors over N units of time is input into the Transformer model, and the Transformer model predicts and outputs a time series ψ over M units of time. output The time series ψ output by the Transformer model output With training label ψ label The evaluation process involves training a Transformer model; step S3 involves using the output of the trained Transformer model to predict the displacement of the AUV. This method allows the raw sensor values ​​to be used as input to the Transformer model to predict and obtain the AUV's position, effectively preventing unavoidable errors during the modeling process and improving the navigation accuracy of the AUV.

Description

Technical Field

[0001] This invention relates to the field of AUV navigation technology, and in particular to an AUV navigation method based on the Transformer model. Background Technology

[0002] In recent years, Autonomous Underwater Vehicles (AUVs) have played a crucial role in tasks such as marine surveying, marine data acquisition, seabed rescue, and underwater equipment maintenance. Accurate navigation and positioning are particularly important for AUVs to safely complete these tasks and collect effective marine data. However, unlike outdoor and aerial robots, underwater navigation and positioning is an extremely challenging task due to limitations in information transmission methods and distances. For example, the Global Positioning System (GPS) used by vehicles is unavailable in AUVs. Furthermore, the complex underwater environment also presents difficulties for navigation.

[0003] One of the most commonly used state estimation techniques in AUV navigation is the Kalman filter (KF). KF is an optimal Bayesian estimator that assumes the system is Markovian, linear, and its uncertainties follow a Gaussian distribution. In fact, the system models established for AUV navigation are often nonlinear, and KF performs well in AUV navigation. The aforementioned state estimation techniques are effective means of achieving accurate vehicle position estimation and have been widely used in the field of AUV navigation. However, these techniques introduce modeling errors and state estimation errors, which to some extent affect the final accuracy of the AUV navigation system.

[0004] In addition, the time series algorithms currently used in AUV navigation are mainly based on RNN algorithms for prediction. Their prediction models can discover non-linear, complex, and long-term dependencies in the data, such as RNN, LSTM, GRU, and the recently emerged ConvLSTM. However, these models cannot be modeled in parallel and cannot capture data dependencies over longer periods in parallel. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned defects in the existing technology and propose an AUV navigation method based on the Transformer model. This method can use the original sensor values ​​as input to the Transformer model to predict and obtain the position of the AUV, effectively preventing the unavoidable errors in the modeling process and improving the navigation accuracy of the AUV.

[0006] The technical solution of this invention is: an AUV navigation method based on the Transformer model, comprising the following steps:

[0007] S1: AUV observation data is obtained through its own sensors. The observations at each time step constitute an observation vector. The time series composed of N observation vectors per unit time is used as the input ψ of the Transformer model. input ;

[0008] S2, the time series consisting of N observation vectors over N units of time is input into the Transformer model, and the Transformer model predicts and outputs a time series ψ over M units of time. output The time series ψ output by the Transformer model output With training label ψ label Perform evaluation and train the Transformer model;

[0009] S3 uses the output value of the trained Transformer model to predict the displacement of the AUV.

[0010] In this invention, in step S1 above, the observations at each time step constitute an observation vector.

[0011] ψ i =(v xi ,v yi ,v zi ,β i ,α i ,θ i ,a xi ,a yi ,a zi ,ω xi ,ω yi ,ω zi )

[0012] In the formula: (v xi ,v yi ,v zi Let be the forward velocity, starboard velocity, and downward velocity of the AUV in the carrier coordinate system measured by DVL at time i, and (β) i ,α i ,θ i (a) represents the yaw angle, pitch angle, and roll angle of the AUV measured by the AHRS unit at time i. xi ,a yi ,a zi Let ω be the triaxial acceleration measured by AHRS at time i. xi ,ω yi ,ω zi (i) represents the angular velocity measured by the AHRS at time i.

[0013] The time series of N units of time is:

[0014] ψinput ={ψ i-N+1 ,…,ψ i-1 ,ψ i}

[0015] The training labels are the actual displacements of the AUV over the next M time units:

[0016] ψ label ={ψ label_i+1 ,ψ label_i+2 ,…,ψ label_i+M}={(δx i+1 ,δy i+1 ),(δx i+2 ,δy i+2 ),…,(δx i+M ,δy i+M )}.

[0017] The Transformer model includes an encoder and a decoder, and the time series input to the encoder is {ψ}. i-N+1 …,ψ i-1 ,ψ i};

[0018] The encoder outputs a time series of M units as input to the decoder, and the decoder outputs a time series ψ for the next M units. output :

[0019]

[0020] The encoder includes a stacked encoder input layer, an encoder position encoding layer, and four identical encoder layers, namely a first encoder layer, a second encoder layer, a third encoder layer, and a fourth encoder layer.

[0021] In the encoder input layer, the input time series ψ input By using a fully connected neural network, the input time series data is mapped to a dimension d. model The vector, the encoder position encoding layer pair of dimension d model The vector is positionally encoded using sine and cosine positional encoding, and the positional encoding formula is as follows:

[0022]

[0023]

[0024] Where pos represents the vector position in the input sequence; j represents the dimension index of the encoding, which can be incremented from 0; PE(ps,2j) represents the sine code at position pos, and PE(ps,2j+1) represents the cosine code at position pos. We can see that even positions use the sine function, and odd positions use the cosine function.

[0025] Each encoder layer includes two sub-layers: a multi-head self-attention sub-layer and a fully connected feedforward network sub-layer. Each multi-head attention sub-layer and the fully connected feedforward network sub-layer is followed by a normalization layer. The multi-head attention sub-layer, the fully connected feedforward network sub-layer, and the normalization layer following them are added in the form of residual connections.

[0026] The formula used in the multi-head attention sublayer is:

[0027]

[0028] In the formula: Q represents the query calculation vector; K represents the vector being queried and calculated; V represents the current actual feature; is the scaling factor; Attention(Q,K,V) represents the attention weights, which are output as a vector sequence by scaling the dot product; softmax() is the activation function.

[0029] The output of the normalization layer connected after the multi-head attention sublayer is used as the input of the fully connected pre-feedback network sublayer. The output of the normalization layer connected after the fully connected pre-feedback network sublayer is used as the input of the multi-head attention sublayer in the next encoder layer. In the fourth encoder layer, the output of the normalization layer connected after the fully connected pre-feedback network sublayer is used as the input of the decoder.

[0030] The decoder layer includes a decoder input layer, a decoder position encoder layer, four identical decoder layers and an output layer. The decoder layer includes a first decoder layer, a second decoder layer, a third decoder layer and a fourth decoder layer. The output of the encoder serves as the input of the first decoder layer, and the input of the decoder starts from the last data point of the encoder output.

[0031] The decoder input layer feeds into the first decoder layer through the decoder position encoder layer: a fully connected neural network maps the parameters of the decoder input layer to the decoder position encoder as d. model The position encoding vector is added element by element to the position encoding vector to encode the sequential information in the time series data. After generating the position encoding vector, it is input to the first decoder layer.

[0032] Each decoder layer includes three sub-layers: a multi-head attention sub-layer, an encoder-decoder multi-head attention sub-layer, and a fully connected feedforward network sub-layer. Each sub-layer is followed by a normalization layer, and each sub-layer and its following normalization layer are added together through residual connections.

[0033] The input of the normalized layer connected to the multi-head attention sub-layer is used as the input of the encoding-decoding multi-head attention sub-layer. The input of the normalized layer connected to the encoding-decoding multi-head attention sub-layer is used as the input of the fully connected feedforward network sub-layer. The input of the normalized layer connected to the fully connected feedforward network sub-layer is used as the input of the multi-head attention sub-layer in the next decoder layer. In the fourth decoder layer, the input of the normalized layer connected to the fully connected feedforward network sub-layer is used as the input of the output layer.

[0034] In step S2, after obtaining the displacement increment of the AUV per unit time through the Transformer model, the predicted displacement increment is... The model is trained by comparing it with the true value and using the mean squared error loss between the true and predicted values:

[0035]

[0036] The trained Transformer model outputs a temporal information dataset ψ of the AUV's displacement per unit time. output The coordinates of the AUV's motion position at time step k are obtained by displacement accumulation prediction.

[0037]

[0038]

[0039] In the formula: (S) x0 ,S y0 (S) represents the initial position of the AUV in the current mission. xk ,S yk ) represents the coordinates of the AUV in the navigation coordinate system at time k, thereby realizing the positioning and navigation of the AUV.

[0040] The beneficial effects of this invention are:

[0041] (1) AUV navigation is achieved through the Transformer model. Compared with existing algorithms, it does not require the use of a specific system model. The original sensor values ​​are used as the input of the Transformer model to predict and obtain the position of the AUV. Therefore, the unavoidable errors in the modeling process can be avoided, thereby improving the accuracy of AUV navigation.

[0042] (2) Compared with existing RNN algorithms, the computational complexity of each layer of the Transformer model is better than that of existing recurrent neural networks, thus speeding up the training process. Attached Figure Description

[0043] Figure 1 This is an example diagram of a fixed-length sliding time window method;

[0044] Figure 2 This is a technical roadmap for predicting AUV displacement using the Transformer model described in this application;

[0045] Figure 3 This is the architecture diagram of the Transformer model in this application. Detailed Implementation

[0046] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0047] Specific details are set forth in the following description to provide a full understanding of the invention. However, the invention can be practiced in many ways other than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0048] This invention discloses a Transformer-based AUV navigation method, which includes the following steps.

[0049] The first step is AUV data acquisition and preprocessing.

[0050] Treating AUV navigation as a time-series regression task, data observed by AHRS, including AUV attitude, acceleration, and angular velocity data, as well as velocity data measured by DVL, are used as input data for the Transformer framework.

[0051] The observations at each time step constitute an observation vector, which can be described as:

[0052] ψ i =(v xi ,v yi ,v zi ,β i ,α i ,θ i ,a xi ,a yi ,a zi ,ω xi ,ω yi ,ω zi )

[0053] In the formula: (vxi ,v yi ,v zi Let be the forward velocity, starboard velocity, and downward velocity of the AUV in the carrier coordinate system measured by DVL at time i. (β) i ,α i ,θ i (a) represents the yaw angle, pitch angle, and roll angle of the AUV measured by the AHRS unit at time i. xi ,a yi ,a zi Let ω be the triaxial acceleration measured by AHRS at time i. xi ,ω yi ,ω zi Let be the angular velocity measured by the AHRS at time i. In this embodiment, a unit time is defined as the time interval consistent with the GPS sampling frequency. Therefore, the time series of N unit times can be represented as:

[0054] ψ input ={ψ i-N+1 …,ψ i-1 ,ψ i}

[0055] To describe the motion of the AUV, in this embodiment, the northeastward displacement per unit time is used as the output of the deep-sea frame:

[0056]

[0057] In this embodiment, the training label is the GPS northeastward displacement per unit time:

[0058] ψ label ={ψ label_i+1 ,ψ label_i+2 ,…,ψ label_i+M}={(δx i+1 ,δy i+1 ),(δx i+2 ,δy i+2 ),…,(δx i+M ,δy i+M )}

[0059] To generate a labeled dataset, the following method is used: Figure 1 The fixed-length sliding time window method shown is used to construct ψ for model training and evaluation. input ψ outputYes. Before applying the sliding window to obtain features and labels, the data above is normalized; that is, min-max scaling is performed on all data using the maximum and minimum values ​​of the training dataset. Then, a sliding window is run on the scaled training set to obtain training samples with features and labels, namely the first N and last M observations, respectively.

[0060] The second step involves using the time-series data dataset ψ of AUV attitude and velocity collected by AHRAS and DVL. input Train the Transformer model.

[0061] In this application, the improved Transformer model is used to perform time series learning on the input data collected in the first step. The technical approach is as follows: Figure 2 As shown. Existing Transformer models are typically applied to classification problems, but the location prediction of AUVs in this application is essentially a regression problem. Therefore, in order to make the Transformer model usable for the location prediction of AUVs in this application, the structure of the Transformer model needs to be improved.

[0062] like Figure 3 As shown, the Transformer model in this application includes an encoder and a decoder. The encoder consists of an encoder input layer, an encoder position encoding layer, and four identical encoder layers stacked together. In this embodiment, the four encoder layers include a first encoder layer, a second encoder layer, a third encoder layer, and a fourth encoder layer. In the encoder input layer, the input vector ψ of the sensor data stream... input By using a fully connected neural network, the input time series data is mapped to a dimension d. model The vector is crucial for the model to employ a multi-sensor attention mechanism.

[0063] Next, for dimension d model The vector is position-encoded using sine and cosine position encoding, which encodes the time series ψ of N units of time acquired through AHRAS and DVL. input In the input encoder position encoding layer, the sine and cosine position encoding formulas for the encoder position encoding layer are as follows:

[0064]

[0065]

[0066] Where j represents the vector dimension of the time series data; pos represents the vector position in the input sequence; PE(ps,2j) represents the sine code at position pos; and PE(ps,2j+1) represents the cosine code at position pos.

[0067] By adding the input vector element-wise with positional codes, which may be cosine, sinine, etc., the sequential information in the time series data of the input vector is encoded. The encoder positional encoding layer allows the sequential information of the input vector to enter the model, and the finally generated positional encoded vector is fed into four identical encoder layers connected in sequence.

[0068] Each encoder layer in this application comprises two sub-layers: a multi-head self-attention sub-layer and a fully connected feedforward network sub-layer. Each multi-head attention sub-layer and fully connected feedforward network sub-layer is followed by a normalization layer. The multi-head attention sub-layer, fully connected feedforward network sub-layer, and their respective normalization layers are added together via residual connections. The output of the normalization layer following the multi-head attention sub-layer serves as the input to the fully connected feedforward network sub-layer. The output of the normalization layer following the fully connected feedforward network sub-layer serves as the input to the multi-head attention sub-layer in the next encoder layer. In the fourth encoder layer, the output of the normalization layer following the fully connected feedforward network sub-layer serves as the input to the decoder.

[0069] In this application, the framework of the multi-head attention sublayer is as follows: Figure 3 As shown, the formula used in the multi-head attention sub-layer is as follows:

[0070]

[0071] In the formula: Q represents the query calculation vector; K represents the vector being queried and calculated; V represents the current actual feature; and V represents the dimension. is the scaling factor; Attention(Q,K,V) represents the attention weights, which are output as a vector sequence by scaling the dot product; softmax() is the activation function.

[0072] This embodiment uses six self-attention heads, and adds the position encoding vectors in the form of residual connections and normalizes them. The global feature vector is obtained through a multi-head attention sublayer.

[0073] In this embodiment, the second sub-layer employs a fully connected feedforward network to enhance the model's nonlinear expressive power. Within this fully connected feedforward network layer, the output from the six-head attention layer is further processed by layer normalization and a multilayer perceptron. The data is then summed pixel-wise using residual connections and normalized, further enhancing the fitting ability of the fully connected feedforward network. Normalization avoids the slow convergence of the fully connected feedforward network layer caused by excessively large or small parameters, thus yielding the output of the encoder layer of a single Transformer module. The four encoder layers extract the attention information and generate a d... modelA dimensional vector is generated and fed to the decoder. In this embodiment, the d-dimensional vector generated by the fourth encoder layer in the encoder... model The dimensional vector is input into the decoder.

[0074] The decoder layer in this application includes a decoder input layer, a decoder position encoder layer, four identical decoder layers, and an output layer. The decoder layer comprises a first decoder layer, a second decoder layer, a third decoder layer, and a fourth decoder layer. The output of the fourth encoder layer serves as the input to the first decoder layer. The decoder input begins from the last data point output by the encoder.

[0075] Meanwhile, the decoder input layer feeds into the first decoder layer through the decoder position encoder layer. A fully connected neural network maps the parameters of the decoder input layer to the decoder position encoder as d. model The positional encoding vector is added element-wise to the positional encoding vector to encode the sequential information in the time series data. After generating the positional encoding vector, it is input to the first decoder layer.

[0076] In this application, each decoder layer includes three sub-layers: a multi-head attention sub-layer, an encoder-decoder multi-head attention sub-layer, and a fully connected feedforward network sub-layer. Each sub-layer is followed by a normalization layer, and each sub-layer and its subsequent normalization layer are added together via residual connections. The input of the normalization layer following the multi-head attention sub-layer serves as the input of the encoder-decoder multi-head attention sub-layer, and the input of the normalization layer following the encoder-decoder multi-head attention sub-layer serves as the input of the fully connected feedforward network sub-layer. The input of the normalization layer following the fully connected feedforward network sub-layer serves as the input of the multi-head attention sub-layer in the next decoder layer. In the fourth decoder layer, the input of the normalization layer following its fully connected feedforward network sub-layer serves as the input of the output layer.

[0077] In this embodiment, the first multi-head attention sublayer in the decoder layer is designed in a masked form to ensure that the prediction at position i can only depend on the known output at positions less than i. The vector after position i is set to -∞ to ensure that the prediction of time series data points depends only on previous data points.

[0078] The output layer of the decoder layer is a linear layer. The output of the last decoder layer is mapped to the target time series and used as the model output, thereby predicting the displacement of the AUV per unit time.

[0079] After obtaining the displacement increment of the AUV per unit time using the Transformer model, the predicted displacement increment is... The Transformer model is trained by comparing the predicted and actual values. In this application, the model is trained using the mean squared error loss between the predicted and actual values.

[0080]

[0081] In this embodiment, after inputting ten trailing time-series data points into the Transformer model, five future displacements per unit time are predicted. That is, given the time-series values ​​(ψ1,…,ψ9,ψ) input to the encoder... 10 The encoder outputs time series values ​​(ψ) 10 ,…,ψ 13 ,ψ 14 The input to the decoder is used to output the displacement (ψ) for the next five time units. 11 ,…,ψ 14 ,ψ 15 ).

[0082] Applying a lookahead mask ensures that the model only applies attention to data points preceding the target data; that is, when the target is predicted as (ψ... 11 ,ψ 12 When ), the mask ensures that attention weights are only applied within (ψ). 10 ,ψ 11 Therefore, the decoder will not leak information about ψ from the decoder input. 12 and ψ 13 The information is provided. In this embodiment, a small batch size of 64 is used for training.

[0083] In this embodiment, the Adam optimizer is used to obtain the most useful model parameters, wherein the value of controlling the first-order momentum β1 is selected as 0.9, the value of controlling the second-order momentum β2 is selected as 0.98, and the value of the fuzzy factor ε is selected as 10. -9 The learning rate lr was set to 0.001.

[0084] Dropout is employed in both sub-layers of the encoder and all three sub-layers of the decoder. A dropout rate of 0.2 is set for each sub-layer: the multi-head attention sub-layer, the fully connected feedforward network sub-layer, and the normalization layer connected to them.

[0085] The third step is the prediction and output of AUV navigation displacement.

[0086] The trained Transformer model outputs a time-series dataset ψ showing the eastward and northward displacements of the AUV. output The coordinates of the AUV's motion position at time step k are obtained by displacement accumulation prediction.

[0087]

[0088]

[0089] In the formula: (S) x0 ,S y0 (S) represents the initial position of the AUV in the current mission. xk ,S yk ) represents the coordinates of the AUV in the navigation coordinate system at time k, which is oriented northeast, thus enabling accurate prediction of the AUV's navigation displacement.

[0090] The AUV navigation method based on the Transformer model provided by this invention has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from the principles of this invention, and these improvements and modifications also fall within the protection scope of the claims of this invention. The above description of the disclosed embodiments enables those skilled in the art to implement or use this invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of this invention. Therefore, this invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An AUV navigation method based on the Transformer model, characterized in that, Includes the following steps: S1: AUV observation data is obtained through its own sensors. The observation values ​​at each time step constitute an observation vector. The time series composed of the observation vectors at N unit times is used as the input to the Transformer model. ; S2, a time series consisting of N observation vectors over N units of time is input into the Transformer model, which then predicts and outputs a time series over M units of time. The time series output by the Transformer model With training labels Perform evaluation and train the Transformer model; S3, using the output value of the trained Transformer model, predicts the displacement of the AUV; The Transformer model includes an encoder and a decoder, and the time series input to the encoder is... ; The encoder outputs a time series of M units as input to the decoder, and the decoder outputs a time series of the next M units. : ； The encoder includes a stacked encoder input layer, an encoder position encoding layer, and four identical encoder layers, namely a first encoder layer, a second encoder layer, a third encoder layer, and a fourth encoder layer. In the encoder input layer, the input time series By using a fully connected neural network, the input time series data is mapped to a dimension of [dimensionality missing]. The vector, the encoder position encoding layer pairs of dimension The vector is positionally encoded using sine and cosine positional encoding, and the positional encoding formula is as follows: ， ， Where pos represents the vector position in the input sequence; j represents the dimension index of the encoding, which increments from 0; This represents the sine code at position pos. This represents the cosine code at position pos.

2. The AUV navigation method based on the Transformer model according to claim 1, characterized in that, In step S1, the observations at each time step constitute an observation vector. ， In the formula: Let i be the forward velocity, starboard velocity, and downward velocity of the AUV in the carrier coordinate system measured by DVL at time i. The yaw angle, pitch angle, and roll angle of the AUV measured by the AHRS unit at time i are... Let be the triaxial acceleration measured by AHRS at time i. Let be the angular velocity measured by AHRS at time i; The time series of N units of time is: ， The training labels are the actual displacements of the AUV over the next M time units: 。 3. The AUV navigation method based on the Transformer model according to claim 1, characterized in that, Each encoder layer includes two sub-layers: a multi-head self-attention sub-layer and a fully connected feedforward network sub-layer. Each multi-head attention sub-layer and the fully connected feedforward network sub-layer is followed by a normalization layer. The multi-head attention sub-layer, the fully connected feedforward network sub-layer, and the normalization layer following them are added in the form of residual connections. The formula used in the multi-head attention sublayer is: ， In the formula: Q represents the query calculation vector; K represents the vector being queried and calculated; V represents the current actual feature; This is the scaling factor; The attention weights are represented by a sequence of output vectors scaled by the dot product. It is an activation function; The output of the normalization layer connected after the multi-head attention sublayer is used as the input of the fully connected pre-feedback network sublayer. The output of the normalization layer connected after the fully connected pre-feedback network sublayer is used as the input of the multi-head attention sublayer in the next encoder layer. In the fourth encoder layer, the output of the normalization layer connected after the fully connected pre-feedback network sublayer is used as the input of the decoder.

4. The AUV navigation method based on the Transformer model according to claim 1, characterized in that, The decoder includes a decoder input layer, a decoder position encoder layer, four identical decoder layers and an output layer. The decoder layers include a first decoder layer, a second decoder layer, a third decoder layer and a fourth decoder layer. The encoder output serves as the input of the first decoder layer, and the decoder input starts from the last data point of the encoder output. The decoder input layer feeds into the first decoder layer through the decoder position encoder layer: a fully connected neural network maps the parameters input to the decoder position encoder into... The position encoding vector is added element by element to the position encoding vector to encode the sequential information in the time series data. After generating the position encoding vector, it is input to the first decoder layer. Each decoder layer includes three sub-layers: a multi-head attention sub-layer, an encoder-decoder multi-head attention sub-layer, and a fully connected feedforward network sub-layer. Each sub-layer is followed by a normalization layer, and each sub-layer and its following normalization layer are added together through residual connections. The input of the normalized layer connected to the multi-head attention sublayer is used as the input of the encoding-decoding multi-head attention sublayer. The input of the normalized layer connected to the encoding-decoding multi-head attention sublayer is used as the input of the fully connected feedforward network sublayer. The input of the normalized layer connected to the fully connected feedforward network sublayer is used as the input of the multi-head attention sublayer in the next decoder layer. In the fourth decoder layer, the input of the normalized layer connected to the fully connected feedforward network sublayer is used as the input of the output layer.

5. The AUV navigation method based on the Transformer model according to claim 1, characterized in that, In step S2, after obtaining the displacement increment of the AUV per unit time through the Transformer model, the predicted displacement increment is... The model is trained by comparing it with the true value and using the mean squared error loss between the true and predicted values: 。 6. The AUV navigation method based on the Transformer model according to claim 1, characterized in that, The trained Transformer model outputs a dataset of time-series information on the displacement of AUVs per unit time. The coordinates of the AUV's motion position at time step k are obtained by displacement accumulation prediction. ， ， In the formula: Indicates the initial position of the AUV for the current mission. This represents the coordinates of the AUV in the navigation coordinate system at time k, thus enabling the AUV's positioning and navigation.

Images