A method for predicting AUV maneuvering motion based on HAM-BiTL combined network model

Abstract

The present invention is applicable to the field of autonomous underwater vehicle technology, and proposes a method for predicting AUV maneuvering motion based on the HAM‑BiTL combined network model. The method combines bidirectional temporal convolution, bidirectional long short-term memory network and hybrid attention mechanism, and the constructed model is named HAM‑BiTL combined network model. By combining the sliding window method and its improved form, the multivariate state prediction model of AUV maneuvering motion is trained and verified. Bidirectional temporal convolution ensures the causality of network prediction and can effectively extract the spatial features of multivariate posture data; the bidirectional long short-term memory network effectively extracts the sequence time features and ensures the front-to-back correlation of time series data. The hybrid attention mechanism can reasonably allocate weights in the sample data, thereby realizing multi-step prediction of multivariate posture data of AUV maneuvering motion under various working conditions. After experimental verification of three typical maneuvering motion scenarios, the HAM‑BiTL combined network model demonstrated excellent prediction performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous underwater vehicles, and in particular relates to a method for predicting AUV maneuvering motion based on a HAM-BiTL combined network model. Background Art

[0002] In recent years, the development of AUVs (autonomous underwater vehicles) has matured. Countries around the world have developed a variety of configurations and series of underwater unmanned vehicles. At the same time, advanced underwater navigation equipment now has autonomous navigation capabilities. AUVs play a vital role in underwater operations such as ocean development, underwater monitoring, and environmental exploration, many of which are pre-programmed with maneuvering. However, due to the difficulty and complexity of underwater environments, AUVs have been lost during many underwater missions.

[0003] The research scope of motion forecasting for moving objects at sea encompasses state and trajectory forecasting. Classical methods, with their simple algorithms, are often used for forecasting objects at sea. These methods are well-suited for forecasting low-dimensional data with clear distribution patterns and simple coupling relationships, as well as for forecasting time series that meet these characteristics. However, when faced with nonlinear, multidimensional, and relatively complex data, classical methods often struggle to meet forecasting requirements, necessitating the use of more complex models.

[0004] Data processing models based on artificial intelligence-related methods are more suitable for the forecasting needs of complex time series. This model is data-driven and can construct relatively accurate nonlinear relationships between the input and output data of unknown systems, and has a certain degree of robustness. Current algorithms have achieved relatively accurate real-time forecasts for moving objects at sea, but there is less research on the motion forecast of AUVs. There are some problems with existing research. Some of the network models in some studies lack verification of experimental data of multiple motion types; the predicted posture state dimension is low, but in practical applications, it is necessary to forecast multi-dimensional posture state data; in addition, many motion forecasts are only single-step forecasts, lacking the practical application value of advance forecasts.

[0005] Therefore, a combined network model based on deep learning is proposed. This model adopts a data-driven approach and can more flexibly and accurately utilize the posture state data of the previous moment when the mathematical model of the AUV is unknown, so as to realize the prediction of the posture state data of the future time during its manipulation movement. Summary of the Invention

[0006] The purpose of the embodiments of the present invention is to provide an AUV maneuvering motion prediction method based on the HAM-BiTL combined network model, aiming to solve the problems raised in the above background technology.

[0007] The embodiment of the present invention is implemented as follows: a method for predicting AUV maneuvering motion based on a HAM-BiTL combined network model, comprising the following steps:

[0008] Step 1: Model construction;

[0009] Construct bidirectional temporal convolutional networks, channel attention mechanisms, bidirectional long short-term memory networks, and multi-head self-attention mechanisms respectively;

[0010] Step 2: Data preparation and processing;

[0011] Collect typical AUV maneuvering motion simulation experimental data under different working conditions to construct a data set for training and validating the network model, and use the maximum normalization method to normalize the data;

[0012] Step 3: Specific implementation of the HAM-BiTL combined network model;

[0013] The proposed HAM-BiTL combined network model is implemented by cascading the bidirectional temporal convolutional network, channel attention mechanism, bidirectional long short-term memory network, and multi-head self-attention mechanism in step 1. The input multidimensional time series consists of samples of a sliding window with a specified sampling width. BiTCN is designed to convert multidimensional AUV motion sample blocks into one-dimensional feature vectors. The temporal spatial features extracted by BiTCN are weighted through the channel attention mechanism, and then the forward and reverse temporal features are captured from the features processed by the self-attention mechanism using BiLSTM. Finally, the multi-head self-attention mechanism adaptively assigns time-varying weights to the spatiotemporal feature data in the current window.

[0014] An embodiment of the present invention provides a method for predicting AUV maneuvering motion based on a HAM-BiTL combined network model. The dataset used in this method consists of simulation results of multi-condition maneuvering motion of the REMUS100, which is independently divided into training and test sets. Experimental verification of maneuvering motion prediction in three typical maneuvering motion scenarios—vertical trapezoidal motion, horizontal Z-shaped motion, and spiral dive motion—demonstrates that the HAM-BiTL combined network model has excellent prediction performance. Ablation experiments also verify that the bidirectional temporal convolution within the HAM-BiTL component effectively ensures causality in the prediction results while extracting effective spatial features. Furthermore, the bidirectional long-short-term memory network effectively extracts sequential temporal features, ensuring the contextual relationship of the time series data. The hybrid attention mechanism effectively allocates weights in the sample data, preserving key temporal and spatial features. Combined with the sliding window method and its improved form, a multivariate state prediction model for AUV maneuvering motion is effectively trained and verified, achieving multi-step prediction. Experimental results demonstrate that the HAM-BiTL outperforms the three compared network models, and that each of its components is indispensable. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is the BiLSTM principle diagram and LSTM basic unit structure;

[0016] Figure 2 The BiTCN schematic and the basic structure of TCN;

[0017] Figure 3 The basic principles of sliding window and multi-step forecasting;

[0018] Figure 4 This is the structural diagram of the HAM-BiTL combined neural network model;

[0019] Figure 5 Prediction results for a portion of the test set for AUV vertical trapezoidal maneuvering motion;

[0020] Figure 6 Prediction results for another part of the test set for AUV vertical trapezoidal maneuvering motion;

[0021] Figure 7 for Figure 5 The corresponding forecast error;

[0022] Figure 8 for Figure 6 The corresponding forecast error;

[0023] Figure 9 is the MAE of the vertical trapezoidal motion prediction result;

[0024] Figure 10Prediction results for a portion of the test set for AUV horizontal plane Z-type maneuvering motion;

[0025] Figure 11 Prediction results for another part of the test set for AUV horizontal plane Z-type maneuvering motion;

[0026] Figure 12 for Figure 10 The corresponding forecast error;

[0027] Figure 13 for Figure 11 The corresponding forecast error;

[0028] Figure 14 is the MAE of the horizontal Z-shaped motion prediction result;

[0029] Figure 15 is the MAE of the spiral diving motion prediction result;

[0030] Figure 16 The MAE of the prediction results of the horizontal plane Z-motion ablation experiment. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0032] The specific implementation of the present invention is described in detail below with reference to specific embodiments.

[0033] An embodiment of the present invention provides an AUV maneuvering motion prediction method based on a HAM-BiTL combined network model, comprising the following steps:

[0034] Step 1: Model construction;

[0035] Step 1.1: Improvement of Long Short-Term Memory Network;

[0036] The Long Short-Term Memory (LSTM) network has three gate structures: forget gate, input gate, and output gate.

[0037] The function of the forget gate is to precisely control the information that needs to be forgotten in the cell state, thereby ensuring that the cell state C t It is properly updated during the information flow. The forget gate receives h t-1 and x t As input parameters, the corresponding forget gate parameters are calculated through the sigmoid layer. tWhen , LSTM first generates a candidate value for update through the tanh layer At the same time, it is also necessary to obtain an input gate parameter i through the sigmoid layer. t To determine the information that needs to be updated. t and Multiply to get the updated information, and at the same time, forget the gate f t and the old cell state C t-1 Multiplying them together, thus forgetting some of the old information, and combining the two to get the updated cell state C t Finally, LSTM needs to calculate the final output information, which mainly depends on the cell state C t , but it needs to be filtered by the output gate. Specifically, the cell state C is first filtered by the tanh layer. t The value is normalized to the interval [-1,1], and then the output gate parameter o is obtained through the sigmoid layer. t , and finally o t and the normalized cell state C t Perform point multiplication to obtain the final filtered result. The detailed calculation process of the forget gate, input gate, and output gate is shown in Equation (1), Equation (2), and Equation (3).

[0038] f t =σ(W f ·[h t-1 ,x t ]+b f ) (1)

[0039]

[0040] Among them, σ represents the sigmoid function, f t represents the forget gate, W f Represents the weight matrix of the forget gate, h t-1 represents the hidden state of the previous time step, x t represents the input of the current time step, b f Represents the bias term of the forget gate, i t represents the input gate, W i represents the weight matrix of the input gate, b i represents the bias term of the input gate, represents the candidate cell state, tanh represents the hyperbolic tangent function, W c Represents the weight matrix for candidate cell state calculation, b C represents the bias term of the candidate cell state, C t Indicates the cell state, o t represents the output gate, W o represents the weight matrix of the output gate, bo Represents the bias term of the output gate, h t Represents the hidden state at the current time step.

[0041] The BiLSTM network used not only contains the basic LSTM structure, but also combines the forward and backward information flows to extract the temporal features of the sequence. In each time step, the BiLSTM network will use formulas (1)-(3) to calculate and update the forward information flow. The BiLSTM principle diagram and LSTM basic unit structure are shown in the figure. Figure 1 shown.

[0042] Step 1.2: Improvement of Convolutional Neural Network;

[0043] Compared with traditional convolutional neural networks, temporal convolutional networks (TCN) show better performance in processing time series data. TCN is a network structure specially designed for time series data processing. Its uniqueness lies in the use of convolutional layers instead of recurrent layers to process sequence dependencies. The core features of TCN include causal convolution and dilated convolution. Causal convolution ensures that when predicting the value at the current moment, only the data at the current moment and before are used, ensuring the causality of the model. Dilated convolution further expands the convolution layer, enabling the network to capture long-range sequence dependencies without increasing the number of parameters or computational complexity. In order to effectively extract implicit information in the reverse direction and obtain a larger receptive field, a bidirectional temporal convolutional network (BidirectionalTemporalConvolutionalNetwork, BiTCN) is constructed through bidirectional links TCN to capture hidden features in the forward and backward directions, so as to better obtain the long-term dependency of the sequence, effectively extract the spatial features of multivariate pose data and ensure the causality of network predictions. The structure of the BiTCN used is as follows: Figure 2 As shown, the part within the frame is the basic structure of TCN.

[0044] Step 1.3: Improvement of attention mechanism;

[0045] The attention mechanism mimics human visual attention, enabling the model to focus on the most important parts of the input data. The channel attention mechanism (CAM) is a mechanism that focuses on assigning importance to feature map channels in convolutional neural networks. Its primary purpose is to improve model performance by assigning different weights to each channel, emphasizing the channels that contribute most to the task and suppressing irrelevant or redundant channels. The hybrid attention mechanism effectively distributes weights across sample data, preserving the key spatiotemporal features of motion data sequences.

[0046] The Multi-Head Self-Attention mechanism is an extension of the Self-Attention Mechanism. It obtains the attention distribution of different subspaces of the input sequence by running multiple independent attention mechanisms in parallel, thereby more comprehensively capturing the potential correlation between multiple features in the sequence. During the operation of the Multi-Head Self-Attention mechanism, the input sequence first passes through three different linear transformation layers to obtain Query, Key, and Value respectively; Q represents the query vector related to the task, K represents the key vector, and V represents the value vector. (K, V) represents the key-value pair; the key-value pair uses vector K to calculate the attention distribution and vector V to calculate the aggregate information. The values ​​of Q, K, and V are calculated by formula (4); for each attention head, a scaled dot product attention operation is performed. The operation process is shown in formula (5):

[0047]

[0048] Among them, W Q ,W K ,W V Respectively, Q, K, and V are weight matrices obtained through training; X represents the input sequence; d k Indicates the dimension of the vector and uses To scale the weights, we can avoid large dot product results.

[0049] Step 2: Data preparation and processing;

[0050] Step 2.1: Data preparation;

[0051] Based on the mathematical model of the AUV Remus 100, simulation experiments were conducted and a data set was obtained. By subjecting the AUV to vertical trapezoidal motion, horizontal Z-shaped motion, and spiral dive experiments under different operating conditions, simulation experimental data for its typical maneuvering motions under these conditions was collected to construct a data set for subsequent network model training and verification. At the same time, different navigation speeds were preset for the three typical maneuvers mentioned above and matched them one by one with the set steering commands. Through these combinations, a variety of experimental conditions were formed, resulting in different data sets. A test set of experimental condition data independent of the training set conditions was also set up to verify the actual effectiveness of the algorithm. Table 1 shows the data set for vertical trapezoidal maneuvering motion, Table 2 shows the data set for horizontal Z-shaped maneuvering motion, and Table 3 shows the data set for spiral dive motion.

[0052] Table 1 Dataset of vertical trapezoidal manipulation motion

[0053]

[0054] Table 2 Dataset of horizontal Z-type maneuvering motion

[0055]

[0056] Table 3 Dataset of spiral rotation diving motion

[0057]

[0058]

[0059] Step 2.2: Data processing;

[0060] Since the features in the data set usually have different measurement units and dimensions, if they are used directly for model training, it may cause network deviation or distortion. Data preprocessing can normalize or standardize the data so that different features are comparable in value and reduce problems caused by different dimensions. Batch normalization can reduce the model's dependence on input data, thereby improving the model's generalization ability and reducing the risk of overfitting. By observing the characteristics and distribution of the data set, in the scenarios of horizontal plane Z-type motion experiment, vertical plane trapezoidal motion experiment and spiral diving motion experiment, many parameters are oscillating on both sides of the value of 0, so the maximum normalization method is used here, and its calculation principle is shown in the following formula (6):

[0061]

[0062] Among them, y n,t is the initial data of the tth time step in the nth feature; is the data with the largest absolute value among the nth features, It is the data normalized by the maximum value at the t-th time step in the n-th feature. Through this normalization method, the data corresponding to all time steps of each feature vector are normalized to the range of [-1, 1], ensuring the original distribution characteristics of the data while eliminating the influence of dimension.

[0063] After completing the preprocessing of the vertical trapezoidal motion, horizontal Z-shaped motion and spiral rotation diving experimental data, the data set is further sorted according to the format specifications of the training multi-step prediction network model; with the help of sliding window technology, the information input into the network model for training or verification is continuously updated according to the established step size; the key parameters of the sliding window are set, including the time step of the training input data and the input step size when the input data is updated each time. The sliding window size used in this method is the input data dimension and time step size for each dimensional variable. The basic principle of the sliding window method is as follows Figure 3 shown.

[0064] The AUV maneuvering motion prediction method adopts a multi-input and multi-output prediction model. Among them, X in formula (7) t|t-(kim-1) As the input for model training and prediction, it is a two-dimensional variable with spatial and temporal dimensions. t|t-(kim-1) is an external input variable, which is related to X t|t-(kim-1) Together as the input data for model training and prediction. t|t-(kim-1) with u t|t-(kim-1) Data is used as input and processed by the trained network model to predict the state quantity at the next moment. As shown in the formula.

[0065]

[0066] u t|t-(kim-1) =[u delta,t-(kim-1) u delta,t-(kim-2) … u delta,t-(kim-n) … u delta,t ] T (8)

[0067]

[0068] Among them, X t|t-(kim-1) Represents the input data consisting of kim steps of motion from the current time t to the historical time t-(kim-1). feature,t-(kim-n) Represents the vector of all characteristic motion states at the historical moment t-(kim-n). f,t-(kim-n) Indicates the value of the f-th characteristic motion state at the historical moment t-(kim-n). u t|t-(kim-1)Represents the vector of external instructions from the current time t to the historical time t-(kim-1), a total of kim steps. delta,t-(kim-n) Indicates the value of the external instruction at the historical time t-(kim-n). Indicates that the total kim steps of motion data X t|t-(kim-1) vector u with external instructions totaling kim steps t|t-(kim-1) Input into the forecaster and get the forecast result at time t+1

[0069] According to the steps of formula (7)-(9), the sliding window is moved downward according to the current time to realize the training of the network model. The data of the sliding window are all real historical data. According to the calculation steps of formula (7)-(9), the test set data is input into the trained model to realize the single-step forecast. Then, in order to realize the multi-step forecast, an improved sliding window data iteration method is proposed. The estimated value obtained by formula (9) is As new data, it is included in the sliding window range of the next step to replace the unknown motion state data in the future, and together with the data from the current time t to the historical time t-(kim-2), it constitutes the input data of the new forecaster. This allows a prediction for time t+2 to be achieved at the current time t. The updated data is shown in Equation (10). Following this procedure, Equations (11) and (12) are executed alternately, and with the help of the improved sliding window method, a multi-step forecast of time t+dim in the future can be made at time t. During the network training phase, the traditional sliding window method is used, while when the predictor is actually put into use, the improved sliding window method is used to achieve multi-step forecasts.

[0070]

[0071] in, is the input state vector for multi-step forecast, is the mth estimated value of the multi-step forecast, where m∈(1,2,3,...,dim).

[0072] Step 3: Specific implementation of the HAM-BiTL combined network model;

[0073] The proposed HAM-BiTL combined network model is constructed by sequentially cascading the bidirectional temporal convolutional network, channel attention mechanism, bidirectional long short-term memory network, and multi-head self-attention mechanism in step 1, as shown in Figure 2. Figure 4As shown in the figure. The input multidimensional time series consists of samples of a sliding window with a specified sampling width. BiTCN is designed to convert multidimensional AUV motion sample blocks into one-dimensional feature vectors. Since the output of the TCN network at each moment is obtained only by the convolution operation of the input at that moment and before, it ensures that it has causal constraints when processing sequences. The temporal spatial features extracted by BiTCN are assigned reasonable weights through the channel attention mechanism. Subsequently, the forward and reverse temporal features are captured from the features processed by the BiLSTM self-attention mechanism. Finally, the multi-head self-attention mechanism adaptively assigns time-varying weights to the spatiotemporal feature data in the current window.

[0074] To gain a systematic and in-depth understanding of the HAM-BiTL combined neural network training and multi-step forecasting implementation, refer to the pseudocode in Algorithm 1 (Table 5), which details the key steps and logic. The experimental environment configuration is shown in Table 4. During training, the HAM-BiTL model uses sampled data from the current moment to the past moment, totaling 20 time steps, as input and sampled data from the next moment in the future as output. Since the sampling time is 0.5 seconds, the network model is trained using sampled data from the current moment to the past moment, totaling 10 seconds, and data from the next moment in the future as output.

[0075] Table 4 Experimental environment configuration

[0076]

[0077]

[0078] Table 5 Algorithm 1

[0079]

[0080] As a preferred embodiment of the present invention, based on the experimental data characteristics, the root mean square error (RMSE), mean square error (MSE), and mean absolute error (MAE) are used as evaluation indicators for forecasting performance. The calculation of RMSE, MSE, and MAE indicators can be achieved using equations (13)-(15). Based on the data characteristics, in order to compare the performance of the traditional network and the HAM-BiTL combined neural network, an evaluation indicator for comparing network forecasting performance, namely the Accuracy Improvement Rate (AIR), is proposed here. This indicator can be expressed by equation (16).

[0081]

[0082] Among them, xi represents the true value, Represents the predicted value, and N represents the number of samples in the test set. MAE new Represents the mean absolute error of the proposed network, MAE old Represents the mean absolute error of the traditional network, and AIR represents the accuracy improvement rate of the proposed network compared with the network used for comparison. When AIR ≥ 0, it means that the accuracy of the proposed network is improved, and when AIR < 0, it means that the accuracy is reduced.

[0083] As a preferred embodiment of the present invention, the following table 1 to 2 are used:

[0084]

[0085]

[0086] The network training and verification of HAM-BiTL are carried out on the data sets corresponding to vertical trapezoidal maneuvering motion, horizontal Z-type maneuvering motion and spiral rotation diving motion in Table 3. At the same time, the superiority of the prediction performance of the HAM-BiTL combined network model is verified by comparing the three typical indicators of RMSE, MSE and MAE with the proposed AIR indicator through the experimental data of the classical method.

[0087] For the AUV vertical trapezoidal maneuvering motion prediction experiment, the horizontal rudder angle δ s , speed U, longitudinal speed u, vertical speed w, pitch rate q, yaw rate r, longitudinal displacement x, vertical displacement z, pitch angle θ and heading angle ψ are used as inputs. The rudder angle command is an external signal, which is usually calculated in advance by the AUV's motion control system. Therefore, in the training process of the prediction model, the horizontal rudder angle δ is used as the input. s The future values ​​of other state variables except for t are used as the predicted output. At the same time, the sliding window method is used to carry out network training. That is, during training, the data from the current time t to the historical time t-kim+1 is used as input to predict the output at the future time t+1 in a single step. Then, during the network model testing process, the improved sliding window method is used to update the single-step prediction results to the input data to achieve a multi-step prediction from the future time t+1 to t+dim. The specific calculation is implemented by equations (7)-(12). Here, kim=20 and dim=6 are set, that is, the motion state of the future time 3s is predicted in real time using a total of 10s from the current time to the historical time. Figure 5 (where a, b, c, and d subgraphs show the prediction results of ship speed U, longitudinal speed u, vertical speed w, and pitch rate q, respectively) and Figure 6(Among them, sub-graphs a, b, c, d, and e respectively show the prediction results of the yaw rate r, longitudinal displacement x, vertical displacement z, pitch angle θ, and heading angle ψ) are the prediction results of the test set of the AUV vertical trapezoidal maneuvering motion. The black curve is the actual motion state data at the corresponding moment, and the diamonds of different colors are the first prediction results of the corresponding color curve algorithm. Figure 5 and Figure 6 It can be seen that HAM-BiTL has excellent prediction performance. In order to observe the prediction performance of the algorithm more clearly and intuitively, Figure 7 and Figure 8 The prediction error curve is presented. The HAM-BiTL combined network model achieves significantly lower prediction errors than existing LSTM, GRU, and CNN-LSTM-Attention algorithms in predicting the multivariate pose state of an AUV. Table 6 shows the evaluation metrics for the vertical trapezoidal motion prediction results.

[0088] Table 6 Evaluation indexes of vertical trapezoidal motion prediction results

[0089]

[0090]

[0091] From the RMSE, MSE, MAE, and AIR evaluation index results in Table 6, it can be seen that the RMSE, MSE, and MAE of HAM-BiTL are all smaller than those of the other three algorithms, and the AIR index of each algorithm is greater than 0. According to formula (16), it can be concluded that the prediction accuracy of the combined network model is significantly improved compared with the traditional algorithm. Figure 9 The MAE values ​​for the prediction results of HAM-BiTL, LSTM, GRU, and CNN-LSTM-Attention for each pose state data are shown. For ease of observation, the MAE for the unified pose data has been normalized. This radar chart clearly shows that HAM-BiTL's MAE is significantly lower than that of the other three algorithms, confirming its excellent prediction performance. The MAE values ​​for each dimension were normalized when plotting this and subsequent radar charts for easier comparison.

[0092] For the horizontal Z-shaped maneuvering motion prediction of AUV, the vertical rudder angle δ at the current and historical moments is v , speed U, longitudinal speed u, vertical speed w, pitch rate q, yaw rate r, longitudinal displacement x, vertical displacement z, pitch angle θ and heading angle ψ are used as inputs. The rudder angle command is an external signal, which is usually calculated in advance by the AUV's motion control system. Therefore, in the training process of the prediction model, the vertical rudder angle δ is used as the input. vThe future values ​​of other state variables except are used as the predicted output. The rest of the calculation steps and parameters are the same as those in the previous experiment. Figure 10 (where a, b, c, and d subgraphs show the prediction results of ship speed U, longitudinal speed u, vertical speed w, and pitch rate q, respectively) and Figure 11 (Among them, sub-graphs a, b, c, d, and e respectively show the prediction results of the yaw rate r, longitudinal displacement x, vertical displacement z, pitch angle θ, and heading angle ψ) are the prediction results of the AUV horizontal plane Z-type maneuvering motion. The black curve in the figure is the actual motion state data at the corresponding moment. Figure 10 and Figure 11 It can be seen that HAM-BiTL has excellent prediction performance. In order to observe the prediction performance of the algorithm more clearly and intuitively, Figure 12 and Figure 13 The prediction error curve is given. It shows that the prediction error of the HAM-BiTL combined network model is significantly smaller than that of the LSTM, GRU and CNN-LSTM-Attention algorithms in the process of predicting the multi-dimensional posture state of the AUV.

[0093] The evaluation indicators of the horizontal plane Z-type motion prediction results are shown in Table 7 below.

[0094] Table 7 Evaluation indexes of horizontal plane Z-shaped motion prediction results

[0095]

[0096]

[0097] From the RMSE, MSE, MAE, and AIR evaluation index results in Table 7, it can be seen that the RMSE, MSE, and MAE of HAM-BiTL are all smaller than those of the other three algorithms, and the AIR index of each algorithm is greater than 0. According to formula (16), it can be concluded that the prediction accuracy of the combined network model is significantly improved compared with the traditional algorithm. Figure 14 The MAE values ​​of the prediction results of HAM-BiTL, LSTM, GRU and CNN-LSTM-Attention for each posture state data can also be clearly seen that the MAE of HAM-BiTL is significantly smaller than that of the other three algorithms, which also confirms its excellent prediction performance.

[0098] For the prediction of spiral diving motion of AUV, the vertical rudder angle δ at current and historical moments is v 、Horizontal rudder angle δ s, speed U, longitudinal speed u, lateral speed v, vertical speed w, pitch rate q, yaw rate r, longitudinal displacement x, lateral displacement y, vertical displacement z, pitch angle θ and heading angle ψ are used as inputs. Because the rudder angle command is an external signal, it is usually calculated in advance according to the motion control system of the AUV. Therefore, except for the vertical rudder angle δ v and horizontal rudder angle δ s The future values ​​of the remaining state variables as input are used as the predicted output.

[0099] Table 8 shows the evaluation indicators of the spiral diving motion prediction results.

[0100] Table 8 Evaluation indicators of spiral diving motion prediction results

[0101]

[0102]

[0103] From the RMSE, MSE, MAE, and AIR evaluation index results in Table 8, it can be seen that the RMSE, MSE, and MAE of HAM-BiTL are all smaller than those of the other three algorithms, and the AIR index of each algorithm is greater than 0. According to formula (16), it can be concluded that the prediction accuracy of the combined network model is significantly improved compared with the traditional algorithm. Figure 15 The MAE values ​​of the prediction results of HAM-BiTL, LSTM, GRU, and CNN-LSTM-Attention for each posture state data are shown in the radar chart. It can be clearly seen that HAM-BiTL is significantly smaller than the other three algorithms, which also confirms its excellent prediction performance.

[0104] To verify the effectiveness of the HAM-BiTL model and its related components in improving network forecasting, ablation experiments were conducted. The results of horizontal Z-shaped experiments using the combined BiLSTM and BiTL models in HAM-BiTL were analyzed. The experimental setup was identical to the horizontal Z-shaped experiments, and the necessary evaluation metrics of RMSE, MSE, MAE, and AIR were used to verify the proposed network combination.

[0105] from Figure 16 As can be seen from the table, the MAE of the prediction results of the HAM-BiTL model for each pose state is lower than that of the other two algorithms, and the MAE of the BiTL model prediction results is lower than the corresponding MAE of the BiLSTM model prediction results. The evaluation indicators of the prediction results of the horizontal plane Z-motion ablation experiment are shown in Table 9.

[0106] Table 9 Evaluation indicators of prediction results of horizontal plane Z-motion ablation experiment

[0107]

[0108] The specific values ​​of the four evaluation metrics (RMSE, MSE, MAE, and AIR) in Table 9 clearly show that the HAM-BiTL model outperforms both BiTL and BiLSTM in terms of forecasting performance, with the BiTL model outperforming BiLSTM. Therefore, it can be concluded that the bidirectional temporal convolution in the proposed HAM-BiTL effectively ensures causality in the forecast results while extracting effective spatial features. The hybrid attention mechanism effectively allocates weights to the sample data, preserving key temporal and spatial features in the data.

[0109] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting AUV maneuvering motion based on a HAM-BiTL combined network model, characterized in that: The following steps are involved: Step 1: Model construction; Construct bidirectional temporal convolutional networks, channel attention mechanisms, bidirectional long short-term memory networks, and multi-head self-attention mechanisms respectively; Step 2: Data preparation and processing; The data set is constructed by collecting typical AUV maneuvering motion simulation experimental data under different working conditions. The data set is used for training and verifying the network model, and the maximum normalization method is used to normalize the data. Step 3: Specific implementation of the HAM-BiTL combined network model; The proposed HAM-BiTL combined network model is implemented by sequentially cascading the bidirectional temporal convolutional network, channel attention mechanism, bidirectional long short-term memory network, and multi-head self-attention mechanism in step 1. The input multidimensional time series consists of samples from a sliding window with a specified sampling width. The BiTCN is designed to convert multidimensional AUV motion sample blocks into one-dimensional feature vectors. The temporal spatial features extracted by the BiTCN are weighted using the channel attention mechanism. Subsequently, the forward and reverse temporal features are captured from the features processed by the self-attention mechanism using the BiLSTM. Finally, the multi-head self-attention mechanism adaptively assigns time-varying weights to the spatiotemporal feature data within the current window. The step 2 comprises the following steps: Step 2.1: Data preparation; Based on the mathematical model of the AUV Remus 100, simulation experiments were conducted to obtain a data set. By subjecting the AUV to vertical trapezoidal motion, horizontal Z-shaped motion, and spiral dive experiments under different operating conditions, simulation experimental data of its typical maneuvering motions under these conditions were collected to construct a data set for subsequent network model training and verification. At the same time, different navigation speeds were preset for the three typical maneuvers and matched them with set steering commands. These combinations formed a variety of experimental conditions, thereby obtaining different data sets. A test set of experimental condition data, independent of the training set conditions, was also set up to verify the actual effectiveness of the algorithm. Step 2.2: Data processing; After obtaining a data set that meets the requirements, the experimental data under multiple working conditions are preprocessed, that is, the data are normalized or standardized. Here, the maximum normalization method is used, and its calculation principle is shown in the following formula (6): Among them, y n,t is the initial data of the tth time step in the nth feature; is the data with the largest absolute value among the nth features, It is the data normalized by the maximum value at the t-th time step in the n-th feature. Through this normalization method, the data corresponding to all time steps of each feature vector are normalized to the range of [-1, 1], ensuring the original distribution characteristics of the data while eliminating the influence of dimension. After preprocessing the data for vertical trapezoidal motion, horizontal Z-shaped motion, and spiral rotation diving experiments, the datasets were further processed according to the data format requirements for training the multi-step prediction network model. The sliding window method was used to continuously update the information input to the network model for training or verification at a predetermined step size. Key parameters of the sliding window were set, including the time step for training the input data and the input step size for each input update. The AUV maneuvering motion prediction adopts a multi-input and multi-output prediction model; among them, X in formula (7) t|t-(kim-1) As the input for model training and prediction, it is a two-dimensional variable with spatial and temporal dimensions; u in formula (8) t|t-(kim-1) is an external input variable, which is related to X t|t-(kim-1) Together as input data for model training and prediction; X within the window range at time t t|t-(kim-1) with u t|t-(kim-1) Data is used as input and processed by the trained network model to predict the state at the next moment. As shown in the formula; in t|t-(kim-1) =[in delta,t-(kim-1) in delta,t-(kim-2) ... in delta,t-(kim-n) ... in delta,t ] T (8) Among them, X t|t-(kim-1) represents the input data consisting of kim steps of motion from the current time t to the historical time t-(kim-1); x feature,t-(kim-n) represents the vector of all characteristic motion states at the historical moment t-(kim-n); x f,t-(kim-n) represents the value of the f-th characteristic motion state at the historical moment t-(kim-n); u t|t-(kim-1) represents the vector of external instructions from the current time t to the historical time t-(kim-1), a total of kim steps; u delta,t-(kim-n) Represents the value of the external instruction at the historical moment t-(kim-n); Indicates that the total kim steps of motion data X t|t-(kim-1) vector u with external instructions totaling kim steps t|t-(kim-1) Input into the forecaster and get the forecast result at time t+1 n∈(1,2,3,...,kim); According to the steps of formula (7)-(9), the sliding window is moved downward according to the current time to realize the training of the network model; then, according to the calculation steps of formula (7)-(9), the test set data is input into the trained model, and then the improved sliding window data iteration method is used to perform multi-step prediction; first, the estimated value obtained by formula (9) is As new data, it is included in the sliding window range of the next step to replace the unknown motion state data in the future, and together with the data from the current time t to the historical time t-(kim-2), it constitutes the input data of the new forecaster. Thus, the prediction of time t+2 is achieved at the current time t; the updated data is shown in formula (10), and then formulas (11) and (12) are executed alternately according to this step, and with the help of the improved sliding window method, a multi-step prediction of time t+dim in the future can be made at time t; it is particularly noted that the traditional sliding window method is used in the network training stage, while the improved sliding window method is used to achieve multi-step prediction when the predictor is actually put into use; in, is the input state vector for multi-step forecast, is the mth estimated value of the multi-step forecast, where m∈(1,2,3,...,dim).

2. The AUV maneuvering motion prediction method based on the HAM-BiTL combined network model according to claim 1 is characterized in that: The step 1 includes the following specific steps: Step 1.1: Feature extraction of bidirectional long short-term memory network; LSTM has three gate structures, namely forget gate, input gate and output gate. The calculation process of forget gate, input gate and output gate is shown in formula (1), formula (2) and formula (3); f t =σ(W f ·[h t-1 ,x t ]+b f ) (1) Among them, σ represents the sigmoid function, f t represents the forget gate, W f Represents the weight matrix of the forget gate, h t-1 represents the hidden state of the previous time step, x t represents the input of the current time step, b f Represents the bias term of the forget gate, i t represents the input gate, W i represents the weight matrix of the input gate, b i represents the bias term of the input gate, represents the candidate cell state, tanh represents the hyperbolic tangent function, W c Represents the weight matrix for candidate cell state calculation, b C represents the bias term of the candidate cell state, C t Indicates the cell state, o t represents the output gate, W o represents the weight matrix of the output gate, b o Represents the bias term of the output gate, h t represents the hidden state of the current time step; The BiLSTM network used not only contains the basic LSTM structure, but also combines the forward and backward information flows to extract the temporal features of the sequence. In each time step, the BiLSTM network will use equations (1)-(3) to calculate and update the information flow. Step 1.2: Feature extraction of convolutional neural network; BiTCN is obtained by bidirectionally linking TCN to capture the hidden features in the front-to-back direction to better obtain the long-term dependency of the sequence, thereby extracting the spatial features of multivariate pose data and ensuring the causality of network prediction; Step 1.3: Feature extraction of attention mechanism; The channel attention mechanism can effectively capture key information by dynamically weighting multivariate features, enhancing the model's ability to model long-term dependencies and nonlinear relationships, while improving robustness and prediction accuracy. The multi-head self-attention mechanism is adopted to fully capture the attention distribution of the input sequence in different subspaces by running multiple independent attention mechanisms in parallel, so as to fully explore the potential correlation between multiple features in the sequence. During the operation of the multi-head self-attention mechanism, the input sequence will first pass through three different linear transformation layers to obtain Query, Key and Value respectively. Q represents the query vector related to the task, K represents the key vector, and V represents the value vector. (K, V) represents the key-value pair. The key-value pair uses the vector K to calculate the attention distribution and the vector V to calculate the aggregate information. The values ​​of Q, K and V are calculated by formula (4). For each attention head, a scaled dot product attention operation is performed. The operation process is shown in formula (5): Among them, W Q ,W K ,W V Respectively, Q, K, and V are weight matrices obtained through training; X represents the input sequence; d k Indicates the dimension of the vector and uses to scale the weights.

3. The AUV maneuvering motion prediction method based on the HAM-BiTL combined network model according to claim 2 is characterized in that: The root mean square error (RMSE), mean square error (MSE) and mean absolute error (MAE) are used as evaluation indicators for forecast performance. The calculation of RMSE, MSE and MAE is realized by formulas (13)-(15). According to the characteristics of the data, in order to compare the performance of the traditional network and the HAM-BiTL combined neural network, an evaluation indicator for comparing the forecast performance of the network, namely the accuracy improvement rate (AIR), is proposed as an evaluation indicator. This indicator is expressed by formula (16). Among them, x i represents the true value, Represents the predicted value, N represents the number of samples in the test set; MAE new Represents the mean absolute error of the proposed network, MAE old Represents the mean absolute error of the traditional network, and AIR represents the accuracy improvement rate of the proposed network compared with the network used for comparison. When AIR ≥ 0, it means that the accuracy of the proposed network is improved, and when AIR < 0, it means that the accuracy is reduced.

Images