Ship motion attitude prediction device based on neural network

Through the empirical modal decomposition and gated graph neural network model to process ship motion attitude data, the shortcomings of neural network in prediction accuracy and real-time are solved, and efficient ship motion attitude prediction is achieved.

CN118468024BActive Publication Date: 2024-11-29HARBIN INST OF TECH
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202410595401.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-14
Publication Date
2024-11-29
Estimated Expiration
2044-05-14

AI Technical Summary

Technical Problem

Existing neural networks are difficult to select appropriate models and ensure training speed in ship motion posture prediction, resulting in insufficient prediction accuracy and real-time performance.

Method used

Empirical modal decomposition (EMD) is used to process the time series data of ship motion postures, decompose it into multiple eigenmodal functions and residual components, and predict it through the gated graph neural network (GGNNs) model to construct the graph structure to Improve prediction accuracy and shorten prediction time.

Benefits of technology

It improves prediction accuracy, ensures real-time data, reduces the difficulty of model construction, and realizes real-time prediction of ship motion posture.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118468024B_ABST
    Figure CN118468024B_ABST
Patent Text Reader

Abstract

The ship motion attitude prediction device based on neural network belongs to the technical field of marine electronic equipment, and particularly relates to the prediction of ship motion attitude; it solves the problem of selecting and constructing a suitable network model existing in the prior art, and the problem that it is difficult to ensure the timeliness of data while improving the accuracy of ship motion attitude prediction, the training speed of the neural network is uncertain, and the real-time prediction cannot be guaranteed; the device includes the following modules: the EMD processing module is used to perform EMD processing on the input ship motion attitude time series data, and decompose the ship motion attitude time series data into multiple IMF component sequences; the GGNNs module is used to use the best GGNNs model according to the graph structure to obtain the prediction results of each IMF component sequence. The ship motion attitude prediction device based on neural network is suitable for the prediction of ship motion attitude.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of marine electronic equipment, and in particular to the prediction of ship motion posture. Background Art

[0002] Ship motion attitude prediction (forecast) refers to the use of certain theories and technologies to predict the motion attitude of a ship in the future. Predicting the motion attitude of a ship in the future effective time can greatly improve the seaworthiness and safety of the ship, and can also improve the combat effectiveness of warships and relax the performance requirements for carrier-based aircraft, which has significant social and military value.

[0003] When using a neural network to predict the motion of a ship, you only need to input a piece of historical information into the network. The artificial neural network will automatically learn, adjust the network parameters according to the rules, and find the rules, so that you can get the expected output of the input information, that is, the predicted value. Compared with other prediction methods, the artificial neural network method can not only efficiently process and analyze large-scale data sets, but also model nonlinear relationships, and is suitable for complex systems and multivariate problems. It does not need to know the specific relationship between input and output, so it has great advantages in the field of ship motion prediction.

[0004] However, there are also some problems with the existing prediction methods using neural networks:

[0005] First, there are many types of neural network models, and how to select and build a suitable network model is a difficult problem;

[0006] Second, the training speed of neural networks is uncertain. How to train quickly and ensure the timeliness of data is also a problem that needs to be solved urgently.

[0007] Therefore, there is an urgent need for a ship motion attitude prediction system or method that can simultaneously improve prediction accuracy and prediction real-time performance. Summary of the invention

[0008] The present invention proposes a ship motion attitude prediction device based on a neural network, which solves the problem of selecting and constructing a suitable network model in the prior art, as well as the problem that while the accuracy of ship motion attitude prediction is improved, it is difficult to ensure the timeliness of data, the training speed of the neural network is uncertain, and the real-time prediction cannot be guaranteed.

[0009] The ship motion posture prediction device based on neural network described in the present invention has the following technical solutions:

[0010] A ship motion posture prediction device based on a neural network, the device comprises the following modules:

[0011] The data preprocessing module is used to process the outliers and standardize the collected original ship motion attitude information to obtain the ship motion attitude time series data;

[0012] The result prediction module is used to input the time series data of the ship motion posture into the EMD-GGNNs combined prediction model to obtain the final prediction result; wherein the EMD-GGNNs combined prediction model includes an EMD processing module, a time window division module, a graph structure construction module, a GGNNs module and a reconstruction module;

[0013] The EMD processing module is used to perform EMD processing on the input ship motion attitude time series data, and decompose the ship motion attitude time series data into multiple IMF component sequences; the IMF component sequence is an intrinsic mode function or a residual component;

[0014] The time window division module is used to divide each IMF component sequence into multiple time windows, each time window contains the same number of consecutive time points;

[0015] The graph structure building module is used to build the graph structure of each IMF component sequence according to the result after time window division;

[0016] The GGNNs module is used to obtain the prediction results of each IMF component sequence according to the graph structure using the best GGNNs model;

[0017] The reconstruction module is used to reconstruct the prediction results of all IMF component sequences to obtain the final prediction results.

[0018] Further, a preferred embodiment is provided, wherein the EMD processing module performs EMD processing on the input ship motion attitude time series data by the following method, and decomposes the ship motion attitude time series data into multiple IMF component sequences:

[0019] An outer layer iterative processing method is used to obtain multiple intrinsic mode functions and residual components; the i-th outer layer iterative process of the outer layer iterative processing method is as follows:

[0020] S1, substitute the residual signal of the last outer iterative process into the inner iterative processing method to obtain the i-th intrinsic mode function; wherein, the residual signal of the last outer iterative process of the first outer iterative process is the time series data of the ship's motion attitude;

[0021] S2, using the residual signal in the last outer layer iteration process minus the i-th intrinsic mode function to obtain the residual signal in the i-th outer layer iteration process;

[0022] S3, judging whether all the intrinsic mode functions obtained in all the previous outer layer iteration processes have reached a given threshold, and judging whether the residual signal of the i-th outer layer iteration process can be decomposed into an intrinsic mode function:

[0023] If all the intrinsic mode functions obtained in all the previous outer layer iteration processes reach a given threshold, or the residual signal of the i-th outer layer iteration process cannot be decomposed into an intrinsic mode function, the residual signal of the i-th outer layer iteration process is used as the residual component, and the outer layer iteration processing method ends;

[0024] Otherwise, continue with the next outer iteration process.

[0025] Further, a preferred embodiment is provided, wherein the j-th inner layer iteration process of the inner layer iteration processing method in the i-th outer layer iteration process is as follows:

[0026] SS1, search for maximum and minimum points of the original signal of the previous inner iteration process, and record the position of each maximum and minimum point; the original signal of the previous inner iteration process in the first inner iteration process is the residual signal of the i-1th outer iteration process;

[0027] SS2, according to the searched maximum point and minimum point, obtain the upper envelope and lower envelope of the original signal of the last inner layer iteration process;

[0028] SS3, using the original signal of the last inner iteration process minus the average value of its upper envelope and lower envelope to obtain the jth calculation result of the i-th intrinsic mode function;

[0029] SS4. Determine whether the j-th calculation result of the i-th intrinsic mode function satisfies the definition of IMF:

[0030] If the definition of IMF is satisfied, the j-th calculation result of the i-th intrinsic mode function is used as the i-th intrinsic mode function, and the inner layer iteration processing method in the i-th outer layer iteration process ends;

[0031] Otherwise, the j-th calculation result of the i-th intrinsic mode function is used as the original signal of the j-th inner layer iteration process, and the next inner layer iteration process is continued.

[0032] Furthermore, a preferred embodiment is provided, wherein the graph structure building module is used to build the graph structure of each IMF component sequence according to the result after time window division by using the following method:

[0033] Each time window of each IMF component series is taken as a graph node of the graph structure;

[0034] The distance between any two time windows of each IMF component sequence is used as the edge weight of the graph structure;

[0035] According to the graph nodes and edge weights, the graph structure of each IMF component sequence is constructed.

[0036] Further, a preferred embodiment is provided, wherein the optimal GGNNs model is obtained by the following method:

[0037] Setting initial values ​​for the hyperparameters of the GGNNs model; the initial hyperparameters include optimizer, learning rate, loss function, batch size, and number of iterations;

[0038] Using the training set, the GGNNs model with the initial value of the hyperparameter set is trained according to the back propagation algorithm to obtain a trained GGNNs model; wherein the training set is obtained by dividing the time series data of the ship motion posture;

[0039] The validation set is used to select the best GGNNs model from the trained GGNNs models according to the model accuracy evaluation index; wherein the validation set is obtained by dividing the time series data of the ship motion posture.

[0040] Furthermore, a preferred implementation is provided, using a training set, and training the GGNNs model with the initial values ​​of the hyperparameters set according to the back propagation algorithm, and obtaining the trained GGNNs model includes:

[0041] Step 1: The training set includes the input value and the true label value corresponding to each input value;

[0042] Input the input values ​​in the training set into the GGNNs model with the initial values ​​of the hyperparameters set, obtain the label prediction values ​​through the forward propagation algorithm, and calculate the loss value between the label prediction values ​​and the true label values ​​in the training set;

[0043] Step 2: For each time step, use the error back propagation algorithm to calculate the error signal of that time step, and then calculate the gradient information based on the error signal and the input and output of the current time step;

[0044] Step 3: For each time step, update the parameters of that time step according to the gradient descent algorithm;

[0045] Step 4: Repeat steps 1 to 3 until the error signal on the training set is small enough or reaches a preset number of training times, the training of the GGNNs model is completed, and a trained GGNNs model is obtained.

[0046] Further, a preferred embodiment is provided, wherein the length of the time window is determined by the following method:

[0047] Time windows of different lengths are selected, and based on the time series data of ship motion posture, the accuracy of the EMD-GGNNs combined prediction model is tested using the model accuracy evaluation index. The length of the time window corresponding to the EMD-GGNNs combined prediction model with the highest accuracy is selected as the determined time window length.

[0048] Further, a preferred embodiment is provided, wherein the device further comprises a GUI interaction module;

[0049] The GUI interaction module includes a parameter setting unit, a data reading unit, a preprocessing unit, an image observation unit and a prediction result output unit;

[0050] The parameter setting unit is used to set the parameters of the EMD-GGNNs combined prediction model in a visual way;

[0051] The data reading unit is used to read the time series data of the ship's motion posture;

[0052] The pre-processing unit is used to display the result image of EMD processing on the time series data of the ship's motion posture in a visual way;

[0053] The image observation unit is used to display the image of the final prediction result of the ship's motion posture in a visual manner;

[0054] The prediction result output unit is used to display the error value of the final prediction result in a visual manner.

[0055] The present invention also proposes a computer program product, and its technical solution is as follows:

[0056] A computer program product comprises: one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors control the above-mentioned neural network-based ship motion posture prediction device.

[0057] The present invention also proposes a computer storage medium, and its technical solution is as follows:

[0058] A computer storage medium, wherein a computer program is stored in the computer-readable storage medium, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to control the above-mentioned neural network-based ship motion posture prediction device.

[0059] The present invention has the following beneficial effects:

[0060] 1. The neural network-based ship motion attitude prediction device described in the present invention decomposes the ship motion attitude time series data into multiple intrinsic mode functions through the empirical mode decomposition algorithm, and then predicts and reconstructs them separately to obtain the final prediction result, which can improve the prediction accuracy of the model.

[0061] 2. The neural network-based ship motion attitude prediction device described in the present invention recursively predicts the ship motion attitude data through a gated graph neural network model, which can effectively shorten the prediction time and ensure the real-time nature of the data.

[0062] 3. The ship motion attitude prediction device based on neural network described in the present invention, because the ship motion attitude time series data contains very rich feature information and shows strong nonlinear and non-stationary characteristics, it is difficult to extract its intrinsic characteristics using traditional methods; through the method of empirical mode decomposition, the intrinsic characteristics in the ship motion attitude time series data can be extracted; at the same time, through empirical mode decomposition, the difficulty of building a prediction model can be reduced, and it is expected to improve the prediction accuracy and extend the prediction time.

[0063] 4. The neural network-based ship motion attitude prediction device described in the present invention can realize real-time prediction of ship motion attitude and is used to assist ship control.

[0064] The ship motion posture prediction device based on neural network described in the present invention is suitable for predicting the ship motion posture. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0066] Figure 1 This is a schematic diagram of the structure of the EMD-GGNNs combined prediction model in one embodiment of the present invention. DETAILED DESCRIPTION

[0067] In order to make the technical solutions and advantages of the present invention more clearly described, the specific implementation methods of the present invention will be further described in detail and completely in conjunction with the accompanying drawings. The various implementation methods described below are only part of the preferred solutions of the present invention, rather than all implementation methods; the various implementation methods described below are intended to explain the present invention and cannot be understood as limitations on the present invention; the reasonable combination of technical features defined in the various implementation methods of the present invention, and all other implementation methods obtained by ordinary technicians in the field without creative work based on the implementation methods of the present invention, are within the scope of protection of the present invention.

[0068] Implementation Method 1: Combination Figure 1 This embodiment describes a ship motion posture prediction device based on a neural network, and the specific implementation contents are as follows:

[0069] A ship motion posture prediction device based on a neural network, the device comprises the following modules:

[0070] The data preprocessing module is used to process the outliers and standardize the collected original ship motion attitude information to obtain the ship motion attitude time series data;

[0071] The result prediction module is used to input the time series data of the ship motion posture into the EMD-GGNNs combined prediction model to obtain the final prediction result; wherein the EMD-GGNNs combined prediction model includes an EMD processing module, a time window division module, a graph structure construction module, a GGNNs module and a reconstruction module;

[0072] The EMD processing module is used to perform EMD processing on the input ship motion attitude time series data, and decompose the ship motion attitude time series data into multiple IMF component sequences; the IMF component sequence is an intrinsic mode function or a residual component;

[0073] The time window division module is used to divide each IMF component sequence into multiple time windows, each time window contains the same number of consecutive time points;

[0074] The graph structure building module is used to build the graph structure of each IMF component sequence according to the result after time window division;

[0075] The GGNNs module is used to obtain the prediction results of each IMF component sequence according to the graph structure using the best GGNNs model;

[0076] The reconstruction module is used to reconstruct the prediction results of all IMF component sequences to obtain the final prediction results.

[0077] In this embodiment, GGNNs is a gated graph neural network model. The core idea of ​​GGNNs is to use a gated unit (GRU) in a graph neural network (GNN) to iterate a fixed number of propagation steps T to update the hidden state of each node.

[0078] In this embodiment, the neural network-based ship motion attitude prediction device can achieve real-time prediction of the ship motion attitude to assist in ship control.

[0079] In this implementation manner, in English, the IMF is the abbreviation of the Intrinsic Mode Function.

[0080] In this implementation, the EMD is the empirical mode decomposition (algorithm). EMD processing is processing using the empirical mode decomposition algorithm.

[0081] In this embodiment, the ship motion attitude time series data is decomposed into multiple intrinsic mode functions through the empirical mode decomposition algorithm, and then predicted and reconstructed separately to obtain the final prediction result, which can improve the prediction accuracy of the model.

[0082] In this embodiment, the ship motion posture data is recursively predicted through a gated graph neural network model (i.e., a GGNNs model), which can effectively shorten the prediction time and ensure the real-time nature of the data.

[0083] In this implementation manner, each time window contains the same number of consecutive time points:

[0084] The ship motion attitude time series data is data arranged in the order of sampling time points; similarly, the IMF component sequence is also data arranged in the order of sampling time points.

[0085] Here, the sampling time points are referred to as time points. Each time window contains the same number of consecutive time points, that is, each time window contains the same number of data of consecutive sampling time points.

[0086] In this implementation, outlier processing is to remove missing values ​​or values ​​with obvious variations in the data or to perform mean value interpolation processing.

[0087] In this implementation, the normalization operation is to scale the original data to a certain interval according to a certain ratio, and usually the data is converted to the interval [0, 1] or [-1, 1].

[0088] In this implementation, the reconstruction module is the end data processing operation of the empirical mode decomposition algorithm, and uses the reconstruction algorithm to reconstruct the components into one piece of data.

[0089] Implementation Method 2: Combination Figure 1This embodiment is described. This embodiment is a further limitation of the vessel motion posture prediction device based on neural network described in the first embodiment. The specific implementation contents are as follows:

[0090] The EMD processing module performs EMD processing on the input ship motion attitude time series data by the following method, and decomposes the ship motion attitude time series data into multiple IMF component sequences:

[0091] An outer layer iterative processing method is used to obtain multiple intrinsic mode functions and residual components; the i-th outer layer iterative process of the outer layer iterative processing method is as follows:

[0092] S1, substitute the residual signal of the last outer iterative process into the inner iterative processing method to obtain the i-th intrinsic mode function; wherein, the residual signal of the last outer iterative process of the first outer iterative process is the time series data of the ship's motion attitude;

[0093] S2, using the residual signal in the last outer layer iteration process minus the i-th intrinsic mode function to obtain the residual signal in the i-th outer layer iteration process;

[0094] S3, judging whether all the intrinsic mode functions obtained in all the previous outer layer iteration processes have reached a given threshold, and judging whether the residual signal of the i-th outer layer iteration process can be decomposed into an intrinsic mode function:

[0095] If all the intrinsic mode functions obtained in all the previous outer layer iteration processes reach a given threshold, or the residual signal of the i-th outer layer iteration process cannot be decomposed into an intrinsic mode function, the residual signal of the i-th outer layer iteration process is used as the residual component, and the outer layer iteration processing method ends;

[0096] Otherwise, continue with the next outer iteration process.

[0097] In this embodiment, since the time series data of the ship's motion attitude contains very rich feature information and exhibits strong nonlinear and non-stationary characteristics, it is difficult to extract its intrinsic features using traditional methods. Through the method of empirical mode decomposition, the intrinsic features in the time series data of the ship's motion attitude can be extracted; at the same time, through empirical mode decomposition, the difficulty of building a prediction model can be reduced, and it is expected to improve the prediction accuracy and extend the prediction time.

[0098] Implementation Method 3: Combination Figure 1 This embodiment is described as a further limitation of the vessel motion posture prediction device based on neural network described in the second embodiment. The specific implementation contents are as follows:

[0099] The jth inner layer iteration process of the inner layer iteration processing method in the i-th outer layer iteration process is as follows:

[0100] SS1, search for maximum and minimum points of the original signal of the previous inner iteration process, and record the position of each maximum and minimum point; the original signal of the previous inner iteration process in the first inner iteration process is the residual signal of the i-1th outer iteration process;

[0101] SS2, according to the searched maximum point and minimum point, obtain the upper envelope and lower envelope of the original signal of the last inner layer iteration process;

[0102] SS3, using the original signal of the last inner iteration process minus the average value of its upper envelope and lower envelope to obtain the jth calculation result of the i-th intrinsic mode function;

[0103] SS4. Determine whether the j-th calculation result of the i-th intrinsic mode function satisfies the definition of IMF:

[0104] If the definition of IMF is satisfied, the j-th calculation result of the i-th intrinsic mode function is used as the i-th intrinsic mode function, and the inner layer iteration processing method in the i-th outer layer iteration process ends;

[0105] Otherwise, the j-th calculation result of the i-th intrinsic mode function is used as the original signal of the j-th inner layer iteration process, and the next inner layer iteration process is continued.

[0106] In this implementation, the original signal of the previous inner layer iteration process is searched for maximum and minimum points, and the position of each maximum and minimum point is recorded, as follows:

[0107] Starting from the second sampling time point of the original signal of the last inner iteration process, determine whether the data at each sampling time point is a local extreme point;

[0108] If the data at the current sampling time point is a local extreme point, then determine whether the data at the current sampling time point is a maximum point or a minimum point:

[0109] If the data at the current sampling time point is a maximum point, mark the data maximum point at the current sampling time point and record its position;

[0110] If the data at the current sampling time point is a minimum point, mark the data minimum point at the current sampling time point and record its position.

[0111] In this implementation, the upper envelope and lower envelope of the original signal of the last inner layer iteration process are obtained according to the searched maximum and minimum points, as follows:

[0112] According to the intermediate value between the maximum point and the minimum point, the upper envelope and lower envelope of the original signal of the last inner iteration process are constructed. The specific method is to perform linear interpolation on the data between the maximum point and the minimum point to obtain the upper envelope and lower envelope of the original signal of the last inner iteration process.

[0113] In this implementation, the IMF is defined as: the extreme points (ie, the maximum point and the minimum point) are equal to the zero point or differ by no more than one, or the overall average value is not zero.

[0114] Implementation Method 4: Combination Figure 1 This embodiment is described. This embodiment is a further limitation of the vessel motion posture prediction device based on neural network described in the first embodiment. The specific implementation contents are as follows:

[0115] The graph structure building module is used to build the graph structure of each IMF component sequence according to the result after time window division by using the following method:

[0116] Each time window of each IMF component series is taken as a graph node of the graph structure;

[0117] The distance between any two time windows of each IMF component sequence is used as the edge weight of the graph structure;

[0118] According to the graph nodes and edge weights, the graph structure of each IMF component sequence is constructed.

[0119] Implementation Method 5: Combination Figure 1 This embodiment is described. This embodiment is a further limitation of the vessel motion posture prediction device based on neural network described in the first embodiment. The specific implementation contents are as follows:

[0120] The optimal GGNNs model is obtained by the following method:

[0121] Setting initial values ​​for the hyperparameters of the GGNNs model; the initial hyperparameters include optimizer, learning rate, loss function, batch size, and number of iterations;

[0122] Using the training set, the GGNNs model with the initial value of the hyperparameter set is trained according to the back propagation algorithm to obtain a trained GGNNs model; wherein the training set is obtained by dividing the time series data of the ship motion posture;

[0123] The validation set is used to select the best GGNNs model from the trained GGNNs models according to the model accuracy evaluation index; wherein the validation set is obtained by dividing the time series data of the ship motion posture.

[0124] In this implementation, RMSE is selected as the model accuracy evaluation indicator.

[0125] In this implementation, the validation set is used to select the best GGNNs model from the trained GGNNs models according to the model accuracy evaluation index, as follows:

[0126] The iterative method is used to select the best GGNNs model; assuming that the GGNNs model trained before the iterative method starts is the current optimal GGNNs model, then each iteration process is as follows:

[0127] Input the validation set into the GGNNs model trained in this iteration process to obtain the prediction results output by this iteration process;

[0128] According to the validation set and the prediction results output by this iteration process, the model accuracy evaluation index value corresponding to the prediction results output by this iteration process is obtained:

[0129] Determine whether the model accuracy evaluation index value (such as RMSE) corresponding to the prediction result output by this iteration process is the current minimum value:

[0130] If it is the current minimum value, the GGNNs model trained in this iteration is the current optimal GGNNs model;

[0131] Otherwise, the current optimal GGNNs model remains unchanged;

[0132] Adjust the hyperparameters of the GGNNs model trained in this iteration process, obtain the GGNNs model trained in the next iteration process, and start the next iteration process;

[0133] The iterative method ends when the number of iterations reaches a given threshold, and the current optimal GGNNs model in the last iteration process is the best GGNNs model.

[0134] In this implementation, the basic idea of ​​the back propagation algorithm is to transmit the error signal from the output end back to each time step in the network through the chain rule, and update the model parameters according to the error signal.

[0135] In this implementation, the optimizer refers to an optimization algorithm used to train the model, such as Adam, SGD, etc.

[0136] In this implementation, the learning rate refers to a parameter used to control the weight update step size, which affects the training speed and accuracy of the model.

[0137] In this implementation, the loss function is used to evaluate the gap between the model prediction result and the true value, such as mean square error (MSE), mean absolute error (MAE), etc.

[0138] In this implementation, batch size refers to the number of samples input into the model in each iteration, which affects the training speed and generalization ability of the model.

[0139] In this implementation, the number of iterations refers to the number of rounds of training the model, which affects the fitting ability and generalization ability of the model.

[0140] In this implementation, the GGNNs model selects Adam as the optimizer, MSE as the loss function, and adds a dropout layer to prevent overfitting.

[0141] Implementation Method 6: Combination Figure 1 This embodiment is described. This embodiment is a further limitation of the ship motion posture prediction device based on neural network described in Embodiment 5. The specific implementation contents are as follows:

[0142] The training set is used to train the GGNNs model with the initial values ​​of the hyperparameters set according to the back propagation algorithm. The trained GGNNs model includes:

[0143] Step 1: The training set includes the input value and the true label value corresponding to each input value;

[0144] Input the input values ​​in the training set into the GGNNs model with the initial values ​​of the hyperparameters set, obtain the label prediction values ​​through the forward propagation algorithm, and calculate the loss value between the label prediction values ​​and the true label values ​​in the training set;

[0145] Step 2: For each time step, use the error back propagation algorithm to calculate the error signal of that time step, and then calculate the gradient information based on the error signal and the input and output of the current time step;

[0146] Step 3: For each time step, update the parameters of that time step according to the gradient descent algorithm;

[0147] Step 4: Repeat steps 1 to 3 until the error signal on the training set is small enough or reaches a preset number of training times, the training of the GGNNs model is completed, and a trained GGNNs model is obtained.

[0148] In this implementation, the training set includes input values ​​and true label values, which is a conventional technical approach.

[0149] In the training of neural networks, labels (the true label values ​​here) refer to the identifiers assigned to each sample or data point (the input values ​​here) in the data set (the training set here). In supervised learning tasks, labels provide the correct target output for the training data, helping the model learn the relationship between input and output.

[0150] Implementation Method 7: Combination Figure 1 This embodiment is described. This embodiment is a further limitation of the vessel motion posture prediction device based on neural network described in the first embodiment. The specific implementation contents are as follows:

[0151] The length of the time window is determined by the following method:

[0152] Time windows of different lengths are selected, and based on the time series data of ship motion posture, the accuracy of the EMD-GGNNs combined prediction model is tested using the model accuracy evaluation index. The length of the time window corresponding to the EMD-GGNNs combined prediction model with the highest accuracy is selected as the determined time window length.

[0153] In this implementation, RMSE can also be selected as a model accuracy evaluation indicator.

[0154] Implementation Method 8: Combination Figure 1 This embodiment is described. This embodiment is a further limitation of the vessel motion posture prediction device based on neural network described in the first embodiment. The specific implementation contents are as follows:

[0155] The device also includes a GUI interaction module;

[0156] The GUI interaction module includes a parameter setting unit, a data reading unit, a preprocessing unit, an image observation unit and a prediction result output unit;

[0157] The parameter setting unit is used to set the parameters of the EMD-GGNNs combined prediction model in a visual way;

[0158] The data reading unit is used to read the time series data of the ship's motion posture;

[0159] The pre-processing unit is used to display the result image of EMD processing on the time series data of the ship's motion posture in a visual way;

[0160] The image observation unit is used to display the image of the final prediction result of the ship's motion posture in a visual manner;

[0161] The prediction result output unit is used to display the error value of the final prediction result in a visual manner.

[0162] In this implementation manner, the visualization method is, for example, a graphic and text display page.

[0163] In this embodiment, the parameters of the EMD-GGNNs combined prediction model include the hyperparameters of the GGNNs model.

[0164] The technical solution provided by the present invention is further described in detail above through several specific implementation modes in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the several specific implementation modes described above are not intended to be used as limitations on the present invention. Any reasonable changes and improvements to the present invention, reasonable combinations of implementation modes and equivalent substitutions based on the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A ship motion attitude prediction device based on a neural network, characterized in that: The device comprises the following modules: The data preprocessing module is used to process the outliers and standardize the collected original ship motion attitude information to obtain the ship motion attitude time series data; A result prediction module is used to input the ship motion posture time series data into the EMD-GGNNs combined prediction model to obtain the final prediction result; wherein the EMD-GGNNs combined prediction model includes an EMD processing module, a time window division module, a graph structure construction module, a GGNNs module and a reconstruction module; The EMD processing module is used to perform EMD processing on the input ship motion attitude time series data, and decompose the ship motion attitude time series data into multiple IMF component sequences; the IMF component sequence is an intrinsic mode function or a residual component; The time window division module is used to divide each IMF component sequence into multiple time windows, each time window contains the same number of continuous time points; The graph structure building module is used to build the graph structure of each IMF component sequence according to the result after time window division; The GGNNs module is used to obtain the prediction results of each IMF component sequence according to the graph structure using the best GGNNs model; The reconstruction module is used to reconstruct the prediction results of all IMF component sequences to obtain the final prediction results; The optimal GGNNs model is obtained by the following method: Set initial values ​​for the hyperparameters of the GGNNs model; the initial hyperparameters include optimizer, learning rate, loss function, batch size, and number of iterations; Using the training set, the GGNNs model with the initial value of the hyperparameter set is trained according to the back propagation algorithm to obtain a trained GGNNs model; wherein the training set is obtained by dividing the time series data of the ship motion posture; The validation set is used to select the best GGNNs model from the trained GGNNs models according to the model accuracy evaluation index; wherein the validation set is obtained by dividing the time series data of the ship motion posture.

2. The ship motion posture prediction device based on neural network according to claim 1 is characterized in that: The EMD processing module performs EMD processing on the input ship motion attitude time series data by the following method, and decomposes the ship motion attitude time series data into multiple IMF component sequences: An outer layer iterative processing method is used to obtain multiple intrinsic mode functions and residual components; the i-th outer layer iterative process of the outer layer iterative processing method is as follows: S1, substitute the residual signal of the last outer iterative process into the inner iterative processing method to obtain the i-th intrinsic mode function; wherein, the residual signal of the last outer iterative process of the first outer iterative process is the time series data of the ship's motion attitude; S2, using the residual signal in the last outer layer iteration process minus the i-th intrinsic mode function to obtain the residual signal of the i-th outer layer iteration process; S3, judging whether all the intrinsic mode functions obtained in all the previous outer layer iteration processes have reached a given threshold, and judging whether the residual signal of the i-th outer layer iteration process can be decomposed into an intrinsic mode function: If all the intrinsic mode functions obtained in all the previous outer layer iteration processes reach a given threshold, or the residual signal of the i-th outer layer iteration process cannot be decomposed into an intrinsic mode function, the residual signal of the i-th outer layer iteration process is used as the residual component, and the outer layer iteration processing method ends; Otherwise, continue with the next outer iteration process.

3. The ship motion posture prediction device based on neural network according to claim 2 is characterized in that: The jth inner layer iteration process of the inner layer iteration processing method in the i-th outer layer iteration process is as follows: SS1, search for maximum and minimum points on the original signal of the previous inner iteration process, and record the position of each maximum and minimum point; the original signal of the previous inner iteration process in the first inner iteration process is the residual signal of the i-1th outer iteration process; SS2, according to the searched maximum point and minimum point, obtain the upper envelope and lower envelope of the original signal of the last inner layer iteration process; SS3, using the original signal of the last inner iteration process minus the average value of its upper envelope and lower envelope to obtain the jth calculation result of the i-th intrinsic mode function; SS4. Determine whether the j-th calculation result of the i-th intrinsic mode function satisfies the definition of IMF: If the definition of IMF is met, the j-th calculation result of the i-th intrinsic mode function is used as the i-th intrinsic mode function, and the inner layer iteration processing method in the i-th outer layer iteration process ends; Otherwise, the j-th calculation result of the i-th intrinsic mode function is used as the original signal of the j-th inner layer iteration process, and the next inner layer iteration process is continued.

4. The ship motion posture prediction device based on neural network according to claim 1, characterized in that: The graph structure building module is used to build the graph structure of each IMF component sequence according to the result after time window division by using the following method: Each time window of each IMF component series is taken as a graph node of the graph structure; The distance between any two time windows of each IMF component sequence is used as the edge weight of the graph structure; According to the graph nodes and edge weights, the graph structure of each IMF component sequence is constructed.

5. The ship motion posture prediction device based on neural network according to claim 1, characterized in that: The training set is used to train the GGNNs model with the initial values ​​of the hyperparameters set according to the back propagation algorithm. The trained GGNNs model includes: Step 1: The training set includes the input value and the true label value corresponding to each input value; Input the input values ​​in the training set into the GGNNs model with the initial values ​​of the hyperparameters set, obtain the label prediction values ​​through the forward propagation algorithm, and calculate the loss value between the label prediction values ​​and the true label values ​​in the training set; Step 2: For each time step, use the error back propagation algorithm to calculate the error signal of that time step, and then calculate the gradient information based on the error signal and the input and output of the current time step; Step 3: For each time step, update the parameters of that time step according to the gradient descent algorithm; Step 4: Repeat steps 1 to 3 until the error signal on the training set is small enough or reaches a preset number of training times, the training of the GGNNs model is completed, and a trained GGNNs model is obtained.

6. The ship motion posture prediction device based on neural network according to claim 1, characterized in that: The length of the time window is determined by the following method: Time windows of different lengths are selected, and based on the time series data of ship motion attitude, the accuracy of the EMD-GGNNs combined prediction model is tested using the model accuracy evaluation index. The length of the time window corresponding to the EMD-GGNNs combined prediction model with the highest accuracy is selected as the determined time window length.

7. The ship motion posture prediction device based on neural network according to claim 1, characterized in that: The device also includes a GUI interaction module; The GUI interaction module includes a parameter setting unit, a data reading unit, a preprocessing unit, an image observation unit and a prediction result output unit; The parameter setting unit is used to set the parameters of the EMD-GGNNs combined prediction model in a visual manner; The data reading unit is used to read the time series data of the ship's motion posture; The pre-processing unit is used to display the result image of EMD processing on the time series data of the ship's motion posture in a visual way; The image observation unit is used to display the image of the final prediction result of the ship's motion posture in a visual manner; The prediction result output unit is used to display the error value of the final prediction result in a visual manner.

8. A computer program product comprising: One or more processors and a memory, characterized in that the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors control the neural network-based ship motion posture prediction device as described in any one of claims 1 to 7.

9. A computer storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located controls the neural network-based ship motion posture prediction device according to any one of claims 1 to 7.

Citation Information

Patent Citations

  • Intelligent ship moving posture decomposition field forecasting method

    CN103926932A

  • Short-term ship attitude prediction method based on empirical mode decomposition and support vector regression

    CN112052623A

  • Ship rolling integrated prediction method and system based on deep learning

    CN114528756A

  • Ship motion attitude prediction method based on multiple combinations

    CN115994481A

  • Ship Motion Prediction Method Based on Long Short-Term Memory Network and Gaussian Process Regression

    US20210117796A1