Method and device for predicting the drift trajectory of crashed AUVs
By constructing initial target and source domain datasets and utilizing Siamese neural networks and CNN+LSTM+Attention models, the problem of low accuracy in predicting the drift trajectory after an AUV crash was solved, achieving rapid and accurate prediction in complex marine environments.
Patent Information
- Application Number
- CN202311306783.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-10
- Publication Date
- 2025-10-31
- Estimated Expiration
- 2043-10-10
AI Technical Summary
Existing technologies struggle to quickly and accurately predict the drift trajectory of AUVs after an accident, especially in complex marine environments. Due to limited data and variable environmental conditions, model predictions often deviate from the correct trajectory, and errors accumulate over time.
A transfer learning-based approach is adopted. By constructing an initial target domain and source domain dataset, feature similarity analysis is performed using a Siamese neural network. A trajectory prediction model is constructed by combining CNN, LSTM and attention modules. The model is then fine-tuned using the optimal target domain dataset to generate the final trajectory prediction model.
It enables accurate and rapid prediction of the drift trajectory of an AUV in complex marine environments such as ocean currents and waves after an AUV accident, overcoming the problem of low accuracy caused by insufficient data.
Smart Images
Figure CN117150318B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of AUV navigation, and particularly relates to a method and device for predicting the drift trajectory of a crashed AUV based on transfer learning. Background Technology
[0002] Due to their small size, low noise, and high maneuverability, AUVs play a vital and irreplaceable role in marine development and coastal defense. However, human understanding of the complex marine environment and marine equipment is not yet fully mature, leading to situations such as AUVs losing control or communication during missions. Therefore, it is necessary to predict the drift trajectory of wrecked AUVs in real time, quickly, and accurately, to conduct timely salvage and recovery, and to investigate the causes of failure.
[0003] After losing control at great depths, an AUV slowly rises under the influence of remaining buoyancy, while simultaneously drifting horizontally due to ocean factors such as currents and density currents. This drift exhibits strong temporal correlation, meaning the predicted position of the AUV at the current moment depends on the combined effects of its position at the previous moment and the current ocean environment, classifying it as a multivariate time series prediction problem. However, unlike conventional time series prediction, AUVs spend most of their drift in deep water, and accurate communication based on methods like satellite positioning is impossible underwater, making it impossible to obtain their location information. Therefore, the available labeled dataset is very limited, making it difficult to train a sufficiently accurate time series prediction model. Secondly, ocean factors such as wind, waves, and currents that affect the drift trajectory of AUVs change constantly over time. Especially after the AUV floats to the thermocline, a small volume of it will be exposed above the water surface under the influence of residual buoyancy and continue to drift. The drift direction is mainly affected by ocean surface winds, currents, waves, as well as the AUV's structural shape and attitude in the water. At the same time, ocean currents and waves are very complex and require calculation and prediction based on historical and current ocean information data. This causes the model's prediction results to deviate from the correct trajectory at different ocean strata, and the error will continue to accumulate over time, leading to a decrease in the model's accuracy. Therefore, there is currently no better prediction method for the AUV trajectory prediction problem with characteristic drift. Summary of the Invention
[0004] The purpose of this invention is to propose a method and apparatus for predicting the drift trajectory of crashed AUVs based on transfer learning, which can solve the problems of insufficient drift trajectory data and complex and variable prediction environment in the prediction process, thereby further improving the prediction accuracy.
[0005] To achieve the above objectives, the present invention provides a method for predicting the drift trajectory of a crashed AUV, comprising:
[0006] Collect location data of the crashed AUV and first marine meteorological data to construct an initial target domain dataset. The first marine meteorological data is marine meteorological data that has a direct impact on the drift motion of the crashed AUV.
[0007] Collect ocean drift object location data and second ocean meteorological data to construct an initial source domain dataset. The second ocean meteorological data is ocean meteorological data that directly affects the drift motion of ocean drift objects.
[0008] Feature similarity analysis is performed on the initial target domain dataset and the initial source domain dataset. Sequence subsets with similarity exceeding a similarity threshold are removed, and sequence subsets with similarity not exceeding the similarity threshold are retained to generate the optimal target domain dataset and the optimal source domain dataset.
[0009] The optimal source domain dataset is used to train a base prediction model to generate an initial trajectory prediction model;
[0010] The initial trajectory prediction model is fine-tuned using the optimal target domain dataset to obtain the target trajectory prediction model, which outputs the final predicted drift trajectory of the crashed AUV.
[0011] Optionally, a Siamese neural network is constructed, which comprises two one-dimensional convolutional neural networks with identical structures, and the two networks share parameters;
[0012] The initial target domain dataset and the initial source domain dataset are respectively passed through a convolutional neural network to perform similarity analysis of feature distribution, with the L2 norm used as the similarity measure function.
[0013] Optionally, the base prediction model includes:
[0014] The CNN feature extraction model takes the optimal source domain dataset as input data and inputs it into the CNN feature extraction model to obtain the spatial features of the trajectory sequence.
[0015] The LSTM prediction model is connected to the output of the CNN feature extraction model. The spatial features of the trajectory sequence are used as input data and input into the LSTM prediction model to obtain the temporal features of the trajectory sequence.
[0016] The attention module, connected to the output of the LSTM prediction model, takes the features of the trajectory sequence in the time dimension as input data, and adjusts the attention coefficient through the attention module to obtain the initial predicted drift trajectory of the crashed AUV.
[0017] Optionally, for the trained LSTM prediction model, each time a feature column of the trajectory sequence in the time dimension is selected, randomly shuffled, and input into the model for prediction to obtain the corresponding value score.
[0018] Record each feature column that changes and its corresponding feature value score, and then normalize it.
[0019] Optionally, the output of the hidden layer of the LSTM prediction model is input into the attention module to obtain the weight coefficients through additive attention;
[0020] The output of each hidden layer of the LSTM prediction model is evaluated for value, and a value score is obtained.
[0021] Optionally, fine-tuning the initial trajectory prediction model using the optimal target domain dataset includes:
[0022] Freeze all layers in the initial trajectory prediction model except for the last fully connected layer.
[0023] Start from the shallowest layer and gradually thaw the current layer and the layer above it.
[0024] The test set of the optimal target domain is used for verification. If the effect of the current layer is not improved, the weight of the current layer is not retained. If the effect is improved after loading another layer, the current layer and the previous layer are retained, and the weight parameters are updated.
[0025] The target trajectory prediction model is obtained by retaining the model structure and hyperparameters that minimize the prediction error.
[0026] Optionally, the location data of the crashed AUV is a sequence of drift trajectory points returned by the crashed AUV's sensor at that time;
[0027] The ocean drift object location data is a sequence of historical drift trajectory points returned by the wrecked vessel through the AIS positioning system and sensors.
[0028] The first set of marine meteorological data includes at least: real ocean currents and wave data near the date of the AUV's accident.
[0029] The second marine meteorological data includes at least: real ocean currents and wave data near the date of the accident.
[0030] Optionally, obtain ocean current and wave data near the event dates over the years, as well as reanalysis data of ocean currents and waves.
[0031] The obtained wave and current data near the event date in previous years are compared with the reanalysis data of current and wave, and the difference is used as the influencing factor of other currents and waves near the event date.
[0032] By incorporating other current and wave influencing factors into the current and wave data constructed around the date of the incident, the true current and wave data around the date of the incident are obtained.
[0033] Optionally, the method further includes:
[0034] Based on the measured drift trajectory of the AUV and the final predicted drift trajectory of the crashed AUV, the target trajectory prediction model is evaluated using mean square error, root mean square error, and / or mean absolute error as evaluation indicators.
[0035] In another aspect, the present invention provides a device for predicting the drift trajectory of a crashed AUV, which, by adopting the above-described method for predicting the drift trajectory of a crashed AUV, includes at least:
[0036] The dataset construction module is used to collect location data of the wrecked AUV and first marine meteorological data to construct an initial target domain dataset. The first marine meteorological data is marine meteorological data that directly affects the drift motion of the wrecked AUV; and
[0037] Data on the location of ocean drifting objects and second-generation ocean meteorological data were collected to construct an initial source domain dataset. The second-generation ocean meteorological data refers to ocean meteorological data that directly influences the drift motion of ocean drifting objects.
[0038] Feature similarity analysis is performed on the initial target domain dataset and the initial source domain dataset. Sequence subsets with similarity exceeding a similarity threshold are removed, and sequence subsets with similarity not exceeding the similarity threshold are retained to generate the optimal target domain dataset and the optimal source domain dataset.
[0039] The initial model building module is used to train the base prediction model on the optimal source domain dataset to generate the initial trajectory prediction model;
[0040] The model optimization module is used to fine-tune the initial trajectory prediction model using the optimal target domain dataset to obtain the target trajectory prediction model, which outputs the final predicted drift trajectory of the crashed AUV.
[0041] In another aspect, the present invention provides a readable storage medium storing a program or instructions that, when executed by a processor, implement the steps of the above-described method for predicting the drift trajectory of a crashed AUV, and achieve the same technical effect.
[0042] As can be seen from the above solutions, the advantages of the present invention are:
[0043] The present invention provides a method for predicting the drift trajectory of a lost AUV. This method constructs an initial target domain dataset by collecting location data of the lost AUV and first marine meteorological data, and constructs an initial source domain dataset by collecting location data of ocean drifting objects and second marine meteorological data. A Siamese neural network is then used to perform feature similarity analysis on the initial target domain dataset and the initial source domain dataset to generate an optimal target domain dataset and an optimal source domain dataset. The optimal source domain dataset is then used to train a base prediction model to generate an initial trajectory prediction model. Finally, the optimal target domain dataset is used to fine-tune the initial trajectory prediction model to obtain a target trajectory prediction model. This method, based on transfer learning, constructs a source domain dataset using the drift trajectory of a wrecked vessel and a target domain dataset using the drift trajectory of a wrecked AUV. It then constructs a trajectory prediction model based on CNN+LSTM+Attention and fine-tunes the model using the drift trajectory of the wrecked AUV. This enables the prediction of the drift trajectory of an AUV after a wreck, under the influence of complex marine environments such as currents and waves. Furthermore, it overcomes the problem of low accuracy caused by limited AUV trajectory data. By training the prediction model with a large amount of source data and adjusting the weight parameters based on transfer learning, it achieves more accurate and faster predictions. Attached Figure Description
[0044] Figure 1 The overall flowchart of the method for predicting the drift trajectory of the crashed AUV is shown;
[0045] Figure 2 A schematic diagram of the drift trajectory prediction method for the crashed AUV is shown.
[0046] Figure 3 A structural diagram of sequence similarity analysis based on Siamese neural networks is shown;
[0047] Figure 4 A flowchart of sequence feature extraction based on CNN is shown;
[0048] Figure 5 The diagram shows the trajectory prediction structure based on the LSTM prediction model.
[0049] Figure 6 The prediction process based on the CNN+LSTM+Attention combined model is shown;
[0050] Figure 7 The diagram shows the architecture of a crashed AUV drift trajectory prediction device;
[0051] in:
[0052] 300-Crash AUV Drift Trajectory Prediction Device
[0053] 301 - Dataset Construction Module;
[0054] 302 - Initial Model Building Module;
[0055] 303 - Model Optimization Module;
[0056] S1-S5: Steps. Detailed Implementation
[0057] To make the above features and effects of the present invention clearer and easier to understand, specific embodiments are described below, and detailed descriptions are provided in conjunction with the accompanying drawings.
[0058] like Figures 1-6 As shown, in which, Figure 1 The overall flowchart of the method for predicting the drift trajectory of the crashed AUV is shown. Figure 2 The schematic diagram of the method for predicting the drift trajectory of a crashed AUV is shown. Figure 3 The diagram shows the structure of sequence similarity analysis based on Siamese neural networks. Figure 4 A flowchart of sequence feature extraction based on CNN is shown; Figure 5 The diagram shows the trajectory prediction structure based on the LSTM prediction model. Figure 6 The prediction process based on the CNN+LSTM+Attention combined model is shown.
[0059] A method for predicting the drift trajectory of a crashed AUV, comprising at least:
[0060] S1. Collect the location data of the wrecked AUV and the first marine meteorological data to construct the initial target domain dataset. The first marine meteorological data is the marine meteorological data that has a direct impact on the drift motion of the wrecked AUV.
[0061] In this embodiment, an initial target domain dataset is constructed by collecting the location data of the crashed AUV and the first marine meteorological data, wherein the first marine meteorological data is marine meteorological data that has a direct impact on the drift motion of the crashed AUV.
[0062] Specifically, the location data of the crashed AUV can be obtained by using the sequence of drift trajectory points returned by the sensors of the crashed AUV at the time of the accident.
[0063] Furthermore, the first marine meteorological data refers to marine meteorological data that directly influences the drift motion of the wrecked AUV, and includes at least real ocean current and wave data near the date of the wreck. In practice, this can be achieved by acquiring and analyzing ocean current and wave reanalysis data based on satellite data near the date of the wreck, as well as real-time marine meteorological data obtained by the National Oceanic and Meteorological Administration and the National Marine Science Data Center. Specifically, this can be done by collecting ocean current and wave data near the wreck dates of wrecks over the years, along with ocean current and wave reanalysis data. The acquired ocean current and wave data near the wreck dates of wrecks over the years are then compared with the ocean current and wave reanalysis data, and the difference is used as other influencing factors of ocean currents and waves near the wreck date. Then, these other influencing factors are added to the ocean current and wave data constructed near the wreck date to obtain the real ocean current and wave data near the wreck date. Then, outliers were removed from the obtained location data of the crashed AUV and the first marine meteorological data to construct the initial target domain dataset X. t , as the target domain data.
[0064] S2. Collect ocean drift object location data and second ocean meteorological data to construct an initial source domain dataset. The second ocean meteorological data is ocean meteorological data that directly affects the drift motion of ocean drift objects.
[0065] In this embodiment, for source domain data, an initial source domain dataset is constructed by collecting ocean drift object location data and second ocean meteorological data, wherein the second ocean meteorological data is ocean meteorological data that has a direct impact on the drift motion of ocean drift objects.
[0066] Specifically, the location data of ocean drifting objects can be obtained by collecting the historical drift trajectory point sequences returned by the wrecked vessel through the AIS positioning system and sensors.
[0067] Furthermore, the second marine meteorological data, which directly impacts the drift motion of ocean drift objects, includes at least real ocean current and wave data near the date of the shipwreck. In practice, this can be achieved by acquiring reanalysis data of ocean currents and waves based on satellite data near the date of the shipwreck, as well as analyzing real-time marine meteorological data obtained from the National Oceanic and Meteorological Administration and the National Marine Science Data Center. Specifically, this can be done by collecting ocean current and wave data near the dates of shipwrecks over the years, along with reanalysis data of ocean currents and waves. The acquired ocean current and wave data near the dates of shipwrecks over the years are then compared with the reanalysis data, and the differences are used as influencing factors of other ocean currents and waves near the date of the incident. Then, these other influencing factors are added to the ocean current and wave data constructed near the date of the shipwreck to obtain the real ocean current and wave data near the date of the incident. Finally, outliers are removed from the ocean drift object location data and the second marine meteorological data to construct the initial source domain dataset X. s , as source domain data.
[0068] S3. Perform feature similarity analysis on the initial target domain dataset and the initial source domain dataset, remove the sequence subsets whose similarity exceeds a similarity threshold, retain the sequence subsets whose similarity does not exceed the similarity threshold, and generate the optimal target domain dataset and the optimal source domain dataset.
[0069] In this embodiment, after obtaining the initial target domain dataset and the initial source domain dataset through steps S1 and S2, a Siamese neural network is further used to perform similarity analysis on the feature distributions of the initial target domain dataset and the initial source domain dataset. The structure of the Siamese neural network is as follows: Figure 3 As shown, this Siamese neural network comprises two identical one-dimensional convolutional neural networks that share parameters. The initial target domain dataset and the initial source domain dataset are each passed through a convolutional neural network. High-dimensional spatial features are extracted, and the similarity of the sequence data is calculated using a similarity metric function to achieve similarity analysis of the feature distributions. Specifically, the L2 norm is used as the similarity metric function, i.e.:
[0070] E w =||G W (X t )-G W (X s )||2
[0071] Among them, G W (X t G represents the features extracted from the initial target domain dataset. W (X s ) represents the features extracted from the initial source domain dataset.
[0072] Since higher similarity between the target domain dataset and the source domain dataset leads to lower prediction performance for new task models, this embodiment performs feature distribution similarity analysis on the initial target domain dataset and the initial source domain dataset. Subsets with excessively high similarity (exceeding a similarity threshold) are removed, while subsets with low similarity (not exceeding the threshold) are retained, thus generating the optimal target domain dataset. and the optimal source domain dataset To improve the predictive performance of the model.
[0073] S4. Train the base prediction model using the optimal source domain dataset to generate the initial trajectory prediction model.
[0074] In this embodiment, after filtering the initial target domain dataset and the initial source domain dataset through feature similarity analysis in step S3, the optimal target domain dataset and the optimal source domain dataset are obtained. The optimal source domain dataset and the optimal target domain dataset are then used to construct and train the model. Specifically, firstly, the optimal source domain dataset is used to train the base prediction model, generating an initial trajectory prediction model. Then, the optimal target domain dataset is used to fine-tune the constructed initial trajectory prediction model, thereby obtaining the target trajectory prediction model. The output of the target trajectory prediction model is the final predicted drift trajectory of the crashed AUV.
[0075] In the specific implementation, for the base prediction model, this embodiment adopts a combined model based on CNN+LSTM+Attention, such as... Figure 2 , Figures 4-6As shown in the diagram, the base prediction model comprises three parts: a CNN feature extraction model, an LSTM prediction model, and an attention module. The CNN feature extraction model is a one-dimensional convolutional neural network that slides one-dimensional convolutional kernels of varying sizes across the input matrix to acquire spatial features of the trajectory sequence. The LSTM prediction model undergoes ordered and repeated training with four gating settings and a fixed sliding time window to acquire temporal features of the trajectory sequence. The attention module adjusts the attention coefficients to optimize the importance of each historical moment in the prediction result, thereby improving the model's overall memory capacity. Specifically, for the CNN feature extraction model, the optimal source domain dataset is used as input data to obtain the spatial features of the trajectory sequence. For the LSTM prediction model, it is connected to the output of the CNN feature extraction model, taking the spatial features of the trajectory sequence as input data and outputting the temporal features of the trajectory sequence. For the attention module, it is connected to the output of the LSTM prediction model, taking the temporal features of the trajectory sequence as input data, and adjusting the attention coefficients to obtain the initial predicted drift trajectory of the crashed AUV.
[0076] First, by inputting the selection matrix w s The optimal source domain dataset of and Multiplication yields the input sequence x of the basis prediction model. t x t Includes density flow (m / s), current flow (m / s), and target size (m). 3 Temperature (°C), Salinity (‰), Density (g / cm³) 3 ), gravitational acceleration (m / s²) 2 The input sequence x is then used to measure the diving depth (m) and fluid resistance (N). t The input is fed into a CNN feature extraction model to extract features from the sequence data. Multiple one-dimensional convolutional kernels of varying sizes slide across the input matrix to obtain the feature matrix sequence y. s This refers to the spatial characteristics of the trajectory sequence. In the specific implementation, multiple 1D convolutional layers are set, each using a one-dimensional convolutional kernel of different size that slides with a stride of 8. Pooling layers and Dropout layers are also set to reduce the complexity of the data and prevent overfitting.
[0077] Then, the features y of the trajectory sequence in the spatial dimension sThe data is input into an LSTM prediction model, which uses a sliding time window to predict the trajectory information at the current time step by step, using feature data from the previous three time points (time_step = 8). The spatial features y of the trajectory sequence are then analyzed. s Ten batches of data were input into the LSTM model for training and parameter updates. The hidden layers used three-layer LSTM units, with each layer containing 128, 64, and 32 hidden neurons. To ensure the model's predictive ability on samples outside the training set and maintain high accuracy on test sets at different depths in the deep sea, L1 regularization was used to balance the training objective with minimizing parameter values that contribute less to features. The Adam optimizer was employed for simple and efficient computation. Weights and bias parameters were fine-tuned during model training, and the final feature y of the trajectory sequence in the time dimension was obtained using the test set. t .
[0078] Furthermore, to better assign attention weight parameters to each historical moment, this embodiment uses the output h(h1,h2,h3...h) of the LSTM hidden layer. t The input is fed into the Attention module of the network, which evaluates the value of the output of each hidden layer of the LSTM prediction model, obtaining a value score. This value score is then used to derive the additive attention weight coefficients. Specifically, the softmax function is used to evaluate the value of the output of each hidden layer of the LSTM, and the calculation formula is as follows:
[0079] s core (h t ) = tanh(W s h t +b s )
[0080] a t =softmax(s core (h t ))
[0081] Among them, s core (h t ) represents the value score of each input entering the Attention layer, a t It is h t The attention weight coefficient, W, represents the importance of a corresponding historical moment. s b s Represents a matrix.
[0082] Then, the data is transformed into a one-dimensional result through a fully connected layer for output, yielding the initial predicted drift trajectory y of the crashed AUV. T =σ(W a A+ba ), where A is the output of the attention module, W a b a Represents a matrix.
[0083] Furthermore, in this implementation, to ensure rapid convergence of the proactive neural network training and obtain the global optimal solution, the value score of each feature is calculated based on the principle of permutation feature importance and then normalized as a parameter for model initialization. For the trained LSTM prediction model, the time dimension feature y of the trajectory sequence output by the CNN feature extraction model is selected each time. t A feature column is randomly shuffled and input into the model for prediction to obtain corresponding value scores. Each value score represents the feature's importance; a higher score indicates greater importance to the model. Each changed feature column and its corresponding value score are recorded and normalized. The value score is expressed as:
[0084]
[0085] Where s represents the value score of each feature, that is, the value score of the k-th feature column y. k The loss value L(y) of the random shuffling model prediction t ,y k And using the following normalization formula, the value score of each feature is mapped to the range (0,1), which serves as the model w. x Initialization parameters.
[0086]
[0087] S5. Fine-tune the initial trajectory prediction model using the optimal target domain dataset to obtain the target trajectory prediction model. The target trajectory prediction model outputs the final predicted drift trajectory of the crashed AUV.
[0088] In this embodiment, after constructing the initial trajectory prediction model through step S4, to improve the model's accuracy, the initial trajectory prediction model is further fine-tuned using the optimal target domain dataset to obtain the final target trajectory prediction model. Specifically, all layers in the initial trajectory prediction model except the last fully connected layer are frozen. Then, starting from the shallowest layers, the current layer and the previous layer are gradually unfrozen, and the model is verified using the optimal target domain test set. If the current layer's performance does not improve, its weights are not retained; if the performance improves after loading another layer, the current layer and the previous layer are retained, and the weight parameters are updated. Finally, the model structure and hyperparameters with the smallest prediction error are retained, thus obtaining the target trajectory prediction model. Using this target trajectory prediction model, the final predicted drift trajectory of the crashed AUV can be output.
[0089] Furthermore, in this embodiment, in order to determine the prediction accuracy of the target trajectory prediction model, the target trajectory prediction model is further evaluated based on the fitting between the measured drift trajectory of the AUV and the final predicted drift trajectory of the crashed AUV, using mean square error, root mean square error, and / or mean absolute error as evaluation indicators.
[0090] Furthermore, it should be noted that both the source domain data and the target domain data of this invention can be obtained by combining the trajectory prediction results of sunken ships or wrecked AUVs in different sea areas to obtain a new dataset, thereby improving the generalization ability of the training model.
[0091] In summary, the method for predicting the drift trajectory of a lost AUV provided by this invention constructs an initial target domain dataset by collecting location data of the lost AUV and first marine meteorological data, and constructs an initial source domain dataset by collecting location data of ocean drifting objects and second marine meteorological data. A Siamese neural network is then used to perform feature similarity analysis on the initial target domain dataset and the initial source domain dataset to generate an optimal target domain dataset and an optimal source domain dataset. The optimal source domain dataset is then used to train a base prediction model to generate an initial trajectory prediction model. Finally, the optimal target domain dataset is used to fine-tune the initial trajectory prediction model to obtain a target trajectory prediction model. This method, based on transfer learning, constructs a source domain dataset using the drift trajectory of a wrecked vessel and a target domain dataset using the drift trajectory of a wrecked AUV. It then constructs a trajectory prediction model based on CNN+LSTM+Attention and fine-tunes the model using the drift trajectory of the wrecked AUV. This enables the prediction of the drift trajectory of an AUV after a wreck, under the influence of complex marine environments such as currents and waves. Furthermore, it overcomes the problem of low accuracy caused by limited AUV trajectory data. By training the prediction model with a large amount of source data and adjusting the weight parameters based on transfer learning, it achieves more accurate and faster predictions.
[0092] Furthermore, the above embodiments of the present invention can be applied to terminal devices for the function of predicting the drift trajectory of crashed AUVs. These terminal devices may include personal terminals and host computer terminals, etc., and the embodiments of the present invention do not limit this.
[0093] Reference Figure 7 , Figure 7 A crashed AUV drift trajectory prediction device 300 is shown, which can achieve prediction by means of, for example Figure 1 , Figure 2 The method for predicting the drift trajectory of an AUV involved in an accident, as shown in the present invention, and the AUV drift trajectory prediction device provided in this embodiment of the invention, can realize each process of the above-mentioned method for predicting the drift trajectory of an AUV involved in an accident.
[0094] A crash AUV drift trajectory prediction device 300, employing a crash AUV drift trajectory prediction method provided in the above embodiments, includes at least:
[0095] Dataset construction module 301 is used to collect location data of the wrecked AUV and first marine meteorological data to construct an initial target domain dataset. The first marine meteorological data refers to marine meteorological data that directly affects the drift motion of the wrecked AUV; and
[0096] Data on the location of ocean drifting objects and secondary marine meteorological data were collected to construct an initial source domain dataset. The secondary marine meteorological data refers to marine meteorological data that directly influences the drift motion of ocean drifting objects.
[0097] Feature similarity analysis is performed on the initial target domain dataset and the initial source domain dataset. Subsets of sequences with similarity exceeding a certain similarity threshold are removed, while subsets of sequences with similarity not exceeding the threshold are retained, thus generating the optimal target domain dataset and the optimal source domain dataset.
[0098] The initial model building module 302 is used to train the base prediction model on the optimal source domain dataset to generate the initial trajectory prediction model;
[0099] The model optimization module 303 is used to fine-tune the initial trajectory prediction model using the optimal target domain dataset to obtain the target trajectory prediction model. The target trajectory prediction model outputs the final predicted drift trajectory of the crashed AUV.
[0100] Furthermore, it should be understood that the above-described division of functional modules is merely an example in the crash AUV drift trajectory prediction device 300 according to the embodiments of this application. In actual applications, the above functions can be assigned to different functional modules as needed. That is, the crash AUV drift trajectory prediction device 300 can be divided into functional modules different from those illustrated above to complete all or part of the functions described above.
[0101] Furthermore, this application also provides an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps of the above-described method for predicting the drift trajectory of a crashed AUV and achieve the same technical effect.
[0102] Furthermore, this application embodiment also provides a readable storage medium storing a program or instructions that, when executed by a processor, implements the steps of the above-described method for predicting the drift trajectory of a crashed AUV, and achieves the same technical effect.
[0103] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be applied, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.
[0104] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.
[0105] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.
Claims
1. A method for predicting the drift trajectory of a crashed AUV, characterized in that, Include: Collect location data of the crashed AUV and first marine meteorological data to construct an initial target domain dataset. The first marine meteorological data is marine meteorological data that has a direct impact on the drift motion of the crashed AUV. Collect ocean drift object location data and second ocean meteorological data to construct an initial source domain dataset. The second ocean meteorological data is ocean meteorological data that has a direct impact on the drift motion of ocean drift objects. Feature similarity analysis is performed on the initial target domain dataset and the initial source domain dataset. Sequence subsets with similarity exceeding a similarity threshold are removed, and sequence subsets with similarity not exceeding the similarity threshold are retained to generate the optimal target domain dataset and the optimal source domain dataset. The optimal source domain dataset is used to train a base prediction model to generate an initial trajectory prediction model; The initial trajectory prediction model is fine-tuned using the optimal target domain dataset to obtain the target trajectory prediction model, which outputs the final predicted drift trajectory of the crashed AUV.
2. The method according to claim 1, characterized in that, Construct a Siamese neural network, which comprises two one-dimensional convolutional neural networks with identical structures and shares parameters between the two networks; The initial target domain dataset and the initial source domain dataset are respectively passed through a convolutional neural network to perform similarity analysis of feature distribution, with the L2 norm used as the similarity measure function.
3. The method according to claim 1, characterized in that, The base prediction model includes: The CNN feature extraction model takes the optimal source domain dataset as input data and inputs it into the CNN feature extraction model to obtain the spatial features of the trajectory sequence. The LSTM prediction model is connected to the output of the CNN feature extraction model. The spatial features of the trajectory sequence are used as input data and input into the LSTM prediction model to obtain the temporal features of the trajectory sequence. The attention module, connected to the output of the LSTM prediction model, takes the features of the trajectory sequence in the time dimension as input data, and adjusts the attention coefficient through the attention module to obtain the initial predicted drift trajectory of the crashed AUV.
4. The method according to claim 3, characterized in that For the trained LSTM prediction model, each time a feature column of the trajectory sequence in the time dimension is selected, randomly shuffled, and input into the model to predict the corresponding value score. Record each feature column that changes and its corresponding feature value score, and then normalize it.
5. The method according to claim 3, characterized in that, The output of the hidden layer of the LSTM prediction model is input into the attention module, and the weight coefficients are obtained through additive attention. The output of each hidden layer of the LSTM prediction model is evaluated for value, and a value score is obtained.
6. The method according to claim 3, characterized in that, Fine-tuning the initial trajectory prediction model using the optimal target domain dataset includes: Freeze all layers in the initial trajectory prediction model except for the last fully connected layer. Start from the shallowest layer and gradually thaw the current layer and the layer above it. The test set of the optimal target domain is used for verification. If the effect of the current layer is not improved, the weight of the current layer is not retained. If the effect is improved after loading another layer, the current layer and the previous layer are retained, and the weight parameters are updated. The target trajectory prediction model is obtained by retaining the model structure and hyperparameters that minimize the prediction error.
7. The method according to claim 1, characterized in that, The location data of the crashed AUV is a sequence of drift trajectory points returned by the crashed AUV's sensor at that time; The ocean drift object location data is a sequence of historical drift trajectory points returned by the wrecked vessel through the AIS positioning system and sensors. The first set of marine meteorological data includes at least: real ocean currents and wave data near the date of the AUV's accident. The second marine meteorological data includes at least: real ocean currents and wave data near the date of the accident.
8. The method according to claim 7, characterized in that, Obtain ocean current and wave data near the event dates over the years, as well as reanalysis data of ocean currents and waves. The obtained wave and current data near the event date in previous years are compared with the reanalysis data of current and wave, and the difference is used as the influencing factor of other currents and waves near the event date. By incorporating other current and wave influencing factors into the current and wave data constructed around the date of the incident, the true current and wave data around the date of the incident are obtained.
9. The method according to claim 1, characterized in that, Also includes: Based on the measured drift trajectory of the AUV and the final predicted drift trajectory of the crashed AUV, the target trajectory prediction model is evaluated using mean square error, root mean square error, and / or mean absolute error as evaluation indicators.
10. A device for predicting the drift trajectory of a crashed AUV, characterized in that, The method for predicting the drift trajectory of a crashed AUV as described in any one of claims 1-9 includes at least: The dataset construction module is used to collect the location data of the wrecked AUV and the first marine meteorological data to construct the initial target domain dataset. The first marine meteorological data is the marine meteorological data that has a direct impact on the drift motion of the wrecked AUV. as well as Data on the location of ocean drifting objects and second-generation ocean meteorological data were collected to construct an initial source domain dataset. The second-generation ocean meteorological data refers to ocean meteorological data that directly influences the drift motion of ocean drifting objects. Feature similarity analysis is performed on the initial target domain dataset and the initial source domain dataset. Sequence subsets with similarity exceeding a similarity threshold are removed, and sequence subsets with similarity not exceeding the similarity threshold are retained to generate the optimal target domain dataset and the optimal source domain dataset. The initial model building module is used to train the basis prediction model on the optimal source domain dataset to generate the initial trajectory prediction model; The model optimization module is used to fine-tune the initial trajectory prediction model using the optimal target domain dataset to obtain the target trajectory prediction model, which outputs the final predicted drift trajectory of the crashed AUV.
Citation Information
Patent Citations
Autonomous underwater vehicle track predicating method based on marine environment information
CN104268625A
Data and model coupling analysis method for marine drift trajectory prediction
CN112115598A