A method for predicting the rest period of ship motion

Abstract

The present invention discloses a method for predicting the resting period of ship motion, comprising the following steps: S1. Data acquisition: Obtain sea surface image data through a stereo vision camera, obtain a wave radar image through a shipborne maritime radar, then convert the sea surface image data into a three-dimensional sea surface topography through stereo photogrammetry software, and obtain wave motion characteristic data through inversion and correction; S2. Data sequence processing: Combine the wave motion characteristic data and the ship motion characteristic parameters according to time series to form a ship-wave motion characteristic duration data sequence, and perform normalization processing in combination with zoning management; S3. Establish a two-layer long short-term memory (LSTM) neural network model; S4. Train and verify the LSTM neural network model based on the sliding window segmentation method; S5. Predict the resting period of ship motion under specific sea states through the LSTM neural network model in combination with the ship-wave real-time motion characteristic data sequence. The present invention can realize the prediction of the resting period window of ship motion.

Description

Technical Field

[0001] The present invention relates to the technical field of advanced prediction of ship motion in open sea areas, and particularly relates to a method for predicting the quiescent period of ship motion. Background Art

[0002] The quiescent period (QP) refers to "the time interval during which all motions of the ship are within acceptable limits to perform the required activities". In the NATO standard protocol STANAG4154, it is stipulated that "the allowable limits for ship motion during vertical and short takeoff and vertical landing operations are: roll angle 2.5°, pitch angle 1.5°, and vertical speed (Vs) 1.0 m / s"; when the ship's attitude is maintained within this motion range, it is the quiescent state that meets the requirements for aircraft landing. The associated operations between the ship platform and other systems have requirements for the duration of the quiescent period. For example, for tasks such as fixed-wing aircraft landing, helicopter takeoff, missile launch, and delivery, it is required that QP reaches at least 30 s; for fire control, general landing, maintenance, small manned operation recovery, material replenishment, etc., it is required that QP is about 1 min.

[0003] The technology of ultra-short-term prediction (6 - 11 s) of ship motion has been relatively well studied. Traditional methods include: using statistical (regression analysis) methods to find the time-domain model that best matches the input data of the wave sensor and the output data of the ship response, so as to realize the output prediction of the ship motion attitude in the next few seconds; this method has good effects on linear systems, but the time series modeling and prediction accuracy for non-linear systems are not high; an improved strategy is to use dynamic neural networks to design neural network models to realize the ultra-short-term prediction of ship motion based on past time series data by taking advantage of the characteristic that they can approximate non-linear functions with arbitrary accuracy through learning.

[0004] The prediction of the quiescent period window of ship motion is the detection and prediction of QP for a relatively long time (>30 s) of the ship motion situation. However, in open sea areas, under the action of various factors such as ocean currents, meteorology, and other navigation tools, the wave forces encountered by the ship are very complex, and even an experienced captain sometimes cannot accurately capture the quiescent period of the ship motion. Summary of the Invention

[0005] The present invention is to overcome the above problems in the prior art and provides a method for predicting the quiescent period of ship motion, which can realize the prediction of the quiescent period window of ship motion.

[0006] The present invention provides a method for predicting the quiescent period of ship motion, including the following steps:

[0007] S1. Data acquisition: Obtain sea surface image data through a stereo vision camera, obtain a sea wave radar image through a shipborne maritime radar, then convert the sea surface image data into a three-dimensional sea surface topography through stereo photogrammetry software, and obtain a sea wave motion characteristic parameter array through inversion and calibration;

[0008] S2. Data sequence processing: Combine the sea wave motion characteristic parameter array and the ship motion characteristic parameter array into a ship-sea wave motion characteristic parameter array according to time sequence, and perform normalization processing on the ship-sea wave motion characteristic parameter array in combination with zoning management to obtain one-dimensional serialized data;

[0009] S3. Establish a long short-term memory neural network prediction model;

[0010] S4. Train and verify the long short-term memory neural network model based on the sliding window segmentation method in combination with the normalized ship-sea wave motion characteristic parameter array;

[0011] S5. Instantly predict the ship motion rest period under a specific sea state through the verified long short-term memory neural network prediction model in combination with the real-time processing result of the ship-sea wave motion characteristic historical data.

[0012] For a method for predicting the ship motion rest period according to the present invention, as a preferred mode, step S1 further includes the following steps:

[0013] S11. Obtain sea surface image data within 300 m near the ship through a stereo vision camera. There are 3 groups of stereo vision cameras, which are respectively arranged at the bow, the left side of the ship and the right side of the ship. When synchronously exposing, at least 60% of the overlapping range exists in the two images obtained;

[0014] S12. Analyze the sea surface image data into the sea wave propagation phase and the sea wave propagation amplitude through stereo photogrammetry software, and calculate the sea wave propagation speed, the wave steepness η and the wave tilt angle γ when encountering the ship, and generate a three-dimensional sea surface topography;

[0015] S13. Invert the three-dimensional sea surface topography and perform slope steepness correction through the homologous point matching technology and the resection method in the overlapping area of the stereo image pair to obtain the wavefront amplitude A, the frequency f and the surface flow velocity B;

[0016] S14. Obtain a sea wave radar image of 240 - 3000 m through a shipborne maritime radar, and then obtain the significant wave height H of the sea wave through inversion S , the peak period T P , the main wave direction Θ P and the main wave wavelength L P ;

[0017] S15. The wave steepness η, the wave tilt angle γ, the wavefront amplitude A, the frequency f, the surface flow velocity B, the significant wave height H of the sea waveS , peak period T P , main wave direction Θ P and main wave wavelength L P constitute the sea wave motion characteristic parameter array MCt.

[0018] The radar antenna of the X-band maritime radar emits electromagnetic waves towards the sea surface at a low grazing angle. The small-scale rough waves and capillary waves generated by the wind on the sea surface undergo Bragg resonance scattering with the electromagnetic waves, and this scattering is modulated by the long gravity waves on the sea surface, thereby imaging the sea waves (displayed as "sea clutter" in the radar image); the pixel values in the X-band echo image represent the echo values of the radar in terms of distance and scanning angle; the classical wave-current inversion algorithm assumes the stationarity of the wave field and current field in the inversion area, applies three-dimensional Fourier transform to the radar image sequence, obtains the three-dimensional wave number-frequency radar image spectrum distribution of the sea waves, and inverses the sea wave and sea current information based on the "dispersion relationship" existing among the sea wave frequency, sea wave number, and sea current of the gravity wave. Ideally, the average image intensity of the radar should decay in the form of r (-7 / 4) , but the actual echo situation of the radar is often very complex. For example, sometimes affected by the weather, the echo signal often decays faster on rainy days; due to the relationship of the wind direction, there are also varying degrees of differences in the pixel values of the image in the scanning angle. Therefore, the present invention uses the sea wave inversion result of the optical stereo camera to correct the inversion result of the radar image in real time to ensure the quality of obtaining sea wave-current data. When there is thick fog and at night, the use of the optical camera is restricted, and at this time, the system inverses the spatio-temporal characteristics of the sea waves solely based on the X-band maritime radar.

[0019] In a preferred manner, in the method for predicting the rest period of a ship's motion according to the present invention, in step S1, the inversion result of the sea wave radar image is corrected in real time through the inversion result of the optical stereo camera.

[0020] In a preferred manner, in the method for predicting the rest period of a ship's motion according to the present invention, step S2 further includes the following steps:

[0021] S21. Obtain the angular velocity of the ship through a gyroscope, and obtain the rotation angles in three directions after one integration in combination with the geometric and physical parameters of the ship: roll θ C ,, pitch ψ C and yaw K C ;

[0022] S22. Obtain the motion acceleration of the ship through an accelerometer, and obtain the translational velocities in three directions after one integration in combination with the geometric and physical parameters of the ship: sway HD, surge ZD, and heave CD;

[0023] S23. Obtain the ship motion parameter array M t , M t= [HD, ZD, CD, K C , Ψ C , θ C ;

[0024] S24. Combine the ship motion parameter array M at time t t and the ocean wave motion characteristic parameter array MC t to obtain the ship-ocean wave coupled motion characteristic parameter array X t , where X t = [M t , MC t , MC t = [A, B, η, T P , Θ P , λ P ;

[0025] S25. With the ship's heading being 0°, divide a circular sea area with a radius of P within the effective detection range of the radar into N circular sectors. The central angle α of each sector i-1 = 360° / N, where i = 1, 2,..., N; the ship's sailing speed is V m / s, and define the radius of each circle L = 2V. Then the surrounding sea area wave current field is divided into P / L×N circular sectors. The weight of the Q0 area consistent with the heading is ≥ the weights of the Q1 areas on the left and right of the Q0 area, and the weight of Q N-1 ≥ the weights of the remaining areas. When the ship is stationary, all weights are 1;

[0026] S26. Serialize one-dimensionally the sea area wave current field weights of the two-dimensional circular area to obtain the one-dimensional weight sequence WQ = [Q0, Q1, Q N-1 ,,, Q N ,,,].

[0027] In a preferred embodiment of the method for predicting the ship motion rest period according to the present invention, in steps S21 and S22, the ship geometric and physical parameters include: length, molded breadth, bottom length, deck depth, gross tonnage, designed speed, and designed draft.

[0028] A method for predicting the rest period of ship motion according to the present invention, as a preferred mode, in step S3, the long short-term memory neural network model includes N time steps and 2 LSTM network layers (the number of network layers is determined according to the complexity of the problem, and 2 layers are used in the present invention); the LSTM network layer includes a working unit cell, and the working unit cell realizes the persistence and suppression of information through the structures of a forgetting gate f, an input gate i, and a control gate o. The long short-term memory neural network LSTM is a specific form of recurrent neural network; it not only establishes weight connections between layers, but also establishes weight links between the same layers, effectively solving the problem of gradient disappearance in recurrent neural networks when processing long sequence data. The "gate" consists of a sigmoid network layer and a bitwise multiplication operation. When an information enters the LSTM network, only the information that meets the algorithm authentication will be retained, and the unqualified information will be forgotten through the forgetting gate. When the network model parameters are fixed, the weight scales at different times can be dynamically changed.

[0029] A method for predicting the rest period of ship motion according to the present invention, as a preferred mode, the cell in the LSTM is updated by the following four-step calculation to obtain the state of the LSTM unit at each moment:

[0030] Step 1. The forgetting gate f filters the h t-1 and x t information and decides which information to discard from the previous moment's state.

[0031] f t =σ(W f [h t-1 , x t +b f )

[0032] where: f t is the forgetting gate; σ() is the activation function, W f is the weight; b f is the bias term of the forgetting gate;

[0033] Step 2. The input gate selects which information to retain through the sigmoid function and generates a candidate vector value through the activation function tanh.

[0034]

[0035] where: i t is the input gate; W i and W c are the weights; b i and b c are the bias terms of the input gate and the input node; h t-1 is the output at time t-1; x tis the new variable value input at time t; tanh is the hyperbolic tangent function;

[0036] Step 3. Update the cell state. Multiply the old state by the forget gate to discard the information that needs to be discarded, and then add the filtered new information to obtain the current state of the working cell;

[0037]

[0038] where: C t is the cell state at time t; C t-1 is the cell state at time t-1; is the input state of the memory cell;

[0039] Step 4. The layer processing value passes through the tanh function to determine the final output value.

[0040] O t =σ(W o [h t-1 , x t +b o ), h t =O t *tanh(C t )

[0041] where: O t is the output gate; W o is the weight; b o is the bias term of the output gate; h t is the output at time t.

[0042] In a preferred embodiment, the method for training the long short-term memory neural network learning model in step S4 of the method for predicting the static period of ship motion according to the present invention is as follows:

[0043] Denote the number of ship-wave historical observation motion sequences as T, and the data sequence after step S2 is [X i (1≤i≤T); Select a sliding window of length N to limit the length of the input training sequence by the length of the sliding window, and use the iterative method to update the samples in the sliding window; Group [X i (1≤i≤T) into sliding windows, with N+1 data in each group. The first N data are used as input data to predict the value of the N+1th data Compare the error between the predicted value of the N+1th data and the true value of the N+1th data, and iteratively adjust the network parameters;

[0044] Define the loss function Use the gradient descent method to optimize the model. When the value of the loss function <0.00005, the training is completed.

[0045] A method for predicting the resting period of ship motion according to the present invention, as a preferred embodiment, step S5 further includes the following steps:

[0046] S51. Assume that the motion states of the ship need to be predicted for N steps. After iteration, replace the oldest data in the sliding window with the predicted value of the i-th step, where i = 1, 2, 3, …, N. Each time a replacement is made, let the LSTM network perform a new learning to update the network structure, and use the new network structure for the next prediction;

[0047] S52. Take the three parameters of the roll angle, pitch angle, and heave velocity output by the LSTM model as the combined judgment variables, that is

[0048] S53. Define the judgment threshold Y τ , when 15 consecutive output judgment variables Y are less than the predetermined threshold Y τ , the system initiates the prediction of the moment when the ship enters the resting period.

[0049] A method for predicting the resting period of ship motion according to the present invention, as a preferred embodiment, when the sea state changes, repeat steps S1 - S5 to complete the prediction of the resting period of ship motion under the new sea state.

[0050] The beneficial effects of the present invention compared with the prior art are as follows:

[0051] (1) The present invention uses an X - band marine radar to extract the wave characteristics in the range of 240 - 3000 m around the ship, uses a stereophotogrammetry system to obtain the wave characteristic data hitting the bow within 300 m near the ship, and uses the hull attitude sensors installed on the ship to record the hull attitude motion data; these data themselves contain the characteristic laws of the interaction between the sea wave current motion and the ship motion reaching a momentary "resting" balance; making good use of these data can solve the problem of predicting the resting period of ship motion.

[0052] (2) The prediction system of the present invention applies artificial intelligence deep learning technology. By establishing a long - short - term memory neural network and training it with the "sea wave - ship" linkage historical data sequence, the system continuously recognizes the motion characteristics of the waves around the ship, continuously learns the response of a specific ship to the wave forces under specific sea conditions, and gradually masters the characteristic laws of the interaction between wave motion and ship motion, and finally captures the time window of the ship motion resting period.

[0053] (3) On the basis of comprehensively analyzing all factors affecting the ship motion situation, through professional sensor data collection and professional inversion and calibration processing, the system of the present invention constructs a big data sequence processing mechanism for the motion characteristics of the ship and the sea waves, provides rich and accurate training data input for the long - short - term memory neural network model, and ensures the effectiveness of model establishment and training. Brief Description of the Drawings

[0054] Figure 1 It is a flowchart of a method for predicting the rest period of ship motion;

[0055] Figure 2 It is a flowchart of step S1 of a method for predicting the rest period of ship motion;

[0056] Figure 3 It is a schematic diagram of a method for obtaining ship motion and sea wave motion data;

[0057] Figure 4 It is a flowchart of inverting wave characteristics from a sequence of maritime radar images;

[0058] Figure 5 It is a flowchart of step S2 of a method for predicting the rest period of ship motion;

[0059] Figure 6 It is a schematic diagram of zoning management;

[0060] Figure 7 It is an internal structure diagram of an LSTM neural network;

[0061] Figure 8 It is a framework of a two-layer LSTM neural network learning model;

[0062] Figure 9 It is a schematic diagram of a method for training and validating an LSTM model based on a sliding window;

[0063] Figure 10 It is a flowchart of step S5 of a method for predicting the rest period of ship motion. Detailed Implementation Manner

[0064] Example 1

[0065] As Figure 1 shown, a method for predicting the rest period of ship motion includes the following steps:

[0066] S1. Data acquisition: Obtain sea surface image data through a stereo vision camera, obtain sea wave radar images through a shipborne maritime radar, then convert the sea surface image data into a three-dimensional sea surface terrain through stereo photogrammetry software, and obtain sea wave motion characteristic data through inversion, and correct the inversion result of the sea wave radar image in real time through the inversion result of the optical stereo camera; As Figure 2 shown, step S1 further includes the following steps:

[0067] S11. Obtain sea surface image data within 300 m near the ship through a stereo vision camera. There are 3 groups of stereo vision cameras, which are respectively set at the bow, the left side of the ship, and the right side of the ship. At least 60% of the overlapping range is obtained for the two images obtained during synchronous exposure;

[0068] S12. Parse the sea surface image data into wave propagation phase and wave propagation amplitude through stereophotogrammetry software, calculate the wave propagation speed, wave steepness η and wave tilt angle γ when encountering the ship, and generate a three-dimensional sea surface topography. The schematic diagram of the sea surface image data acquisition method is as Figure 3 shown;

[0069] S13. As Figure 4 shown, invert the three-dimensional sea surface map through the homologous point matching technology and resection method in the overlapping area of the stereoscopic image pair and perform slope steepness correction to obtain the wavefront amplitude A, frequency f and surface flow velocity B;

[0070] S14. Obtain the sea wave radar image of 240 - 3000 m through the shipborne maritime radar, and then obtain the significant wave height H S , peak period T P , main wave direction Θ P and main wave wavelength LP of the sea waves. The radar antenna is installed near the center of the hull, and the antenna height is 10 - 45 m. The schematic diagram of the sea wave radar image acquisition method is as Figure 3 shown. The main technical characteristics of the X-band maritime radar are shown in the following table:

[0071]

[0072] S15. Combine the wave steepness η, wave tilt angle γ, wavefront amplitude A, frequency f, surface flow velocity B, significant wave height H S , peak period T P , main wave direction Θ P and main wave wavelength L P to form a sea wave motion characteristic parameter array MCt;

[0073] S2. Data sequence processing: Combine the sea wave motion characteristic parameter array with the ship motion characteristic parameter array to form a ship-sea wave motion characteristic parameter array, and perform normalization processing on the ship-sea wave motion characteristic data in combination with zoning management to obtain one-dimensional serialized data. As Figure 5 shown, step S2 further includes the following steps:

[0074] S21. Obtain the angular velocity of the ship through the gyroscope, and obtain the rotation angles in three directions after one integration in combination with the ship's geometric and physical parameters: roll θ C , pitch ψ C and yaw K C ; The ship's geometric and physical parameters include: length, molded breadth, bottom length, deck depth, gross tonnage, design speed and design draft;

[0075] S22. Obtain the motion acceleration of the ship through the accelerometer, and after one integration in combination with the geometric and physical parameters of the ship, obtain the translational velocities in three directions: sway HD, surge ZD, and heave CD;

[0076] S23. Obtain the ship motion parameter array M at time t t , M t = [HD, ZD, CD, K C , Ψ C , θ c ;

[0077] S24. Combine the ship motion parameter array M t and the ocean wave motion characteristic parameter array MC t to obtain the ship-ocean wave coupled motion characteristic parameter array X t , where X t = [M t , MC t , MC t = [A, B, η, T P , Θ P , λ P ;

[0078] S25. As shown in Figure 6 , with the ship's heading as 0°, divide a circular sea area with a radius of P within the effective detection range of the radar into N circular ring sectors. The central angle α of each sector i-1 = 360° / N, where i = 1, 2,..., N; the ship's sailing speed is V m / s, and define the radius of each circular ring L = 2V. Then the surrounding sea area wave flow field is divided into P / L × N circular ring sectors. The weight of the Q0 area consistent with the heading ≥ the Q1 areas on the left and right of the Q0 area, and the weight of Q N-1 ≥ the weights of the remaining areas. When the ship is stationary, all weights are 1;

[0079] S26. Serialize the weights of the sea area wave flow field in the two-dimensional circular ring area into a one-dimensional weight sequence WQ = [Q0, Q1, Q N-1 ,,, Q N ,,,];

[0080] S3. Establish a long short-term memory neural network model; as shown in Figures 7 - 8 , the long short-term memory neural network learning model includes N time steps and 2 LSTM network layers; the LSTM network layer includes a working unit cell, and the working unit cell realizes the persistence and inhibition of information through the structure of the forget gate f, input gate i, and control gate o;

[0081] The cell in the LSTM is updated with the state of the LSTM unit at each moment through the following four-step calculation:

[0082] Step 1. The forget gate f filters h t-1 and x t information and decides which information to discard from the previous moment's state.

[0083] f t = σ(W f [h t-1 , x t + b f )

[0084] where: f t is the forget gate; σ() is the activation function, W f is the weight; b f is the bias term of the forget gate;

[0085] Step 2. The input gate selects which information to retain through the sigmoid function and generates a candidate vector value through the activation function tanh.

[0086]

[0087] where: i t is the input gate; W i and W c are the weights; b i and b c are the bias terms of the input gate and the input node; h t-1 is the output at time t-1; x t is the new variable value input at time t; tanh is the hyperbolic tangent function;

[0088] Step 3. Update the cell state. Multiply the old state by the forget gate, discard the information that is determined to be discarded, and then add the filtered new information to obtain the state of the current working unit.

[0089]

[0090] where: C t is the cell state at time t; C t-1 is the cell state at time t-1; is the input state of the memory unit;

[0091] Step 4. The layer processing value passes through the tanh function to determine the final output value.

[0092] O t = σ(W o [h t-1 , x t + b o ), h t = O t * tanh(Ct )

[0093] Wherein: O t is the output gate; W o is the weight; b o is the bias term of the output gate; h t is the output at time t;

[0094] S4. Train and validate the long short - term memory neural network model based on the sliding window segmentation method combined with the normalized ship - sea wave motion characteristic parameter array; as Figure 9 shown, the method for training the long short - term memory neural network learning model is as follows:

[0095] Denote the number of ship - sea wave historical observation motion sequences as T, and the data sequence after step S2 is [X i (1 ≤ i ≤ T); Select a sliding window with a length of N, limit the length of the input training sequence by the length of the sliding window, and use the iterative method to update the samples in the sliding window; Group [X i (1 ≤ i ≤ T) into sliding windows, with N + 1 data in each group. The first N data are used as input data to predict the value of the (N + 1) - th data Compare the error between the predicted value and the true value of the (N + 1) - th data, and iteratively adjust the network parameters;

[0096] Define the loss function Use the gradient descent method to optimize the model. When the value of the loss function < 0.00005, the training is completed;

[0097] S5. Through the verified long short - term memory neural network learning model combined with the real - time processing results of ship - sea wave motion characteristic historical data, immediately predict the ship motion rest period under specific sea states; as Figure 10 shown, step S5 further includes the following steps:

[0098] S51. Assume that it is necessary to predict the motion state of the ship for N steps. After iteration, replace the oldest data in the sliding window with the predicted value of the i - th step, where i = 1, 2, 3, …, N. Each time a replacement is made, let the LSTM network perform a new learning, update the network structure, and make the next prediction with the new network structure;

[0099] S52. Take the roll angle, pitch angle, and heave of the LSTM model output as the combined judgment variables, that is

[0100] S53. Define the judgment threshold Y τ , when 15 consecutive outputs of the judgment variable Y are less than the predetermined threshold Y τ , then the system initiates a forecast of the moment when the ship enters the rest period;

[0101] S6. When the sea state changes, repeat steps S1 - S5 to complete the prediction of the ship's motion rest period under the new sea state.

[0102] The above description is illustrative rather than restrictive to the present invention. Those of ordinary skill in the art understand that any modification, change or equivalence made without departing from the spirit and scope defined by the claims will fall within the protection scope of the present invention.

Claims

1. A method for predicting the rest period of a ship's motion, characterized in that: Including the following steps: S1. Data acquisition: Obtain sea surface image data through a stereo vision camera, obtain a sea wave radar image through a shipborne maritime radar, then convert the sea surface image data into a three-dimensional sea surface topography through stereo photogrammetry software, and obtain an array of sea wave motion characteristic parameters through inversion and calibration; S11. Obtain sea surface image data within 300 m near the ship through a stereo vision camera. There are 3 groups of the stereo vision cameras, which are respectively arranged at the bow, the starboard side, and the port side of the ship; S12. Analyze the sea surface image data into sea wave propagation phase and sea wave propagation amplitude through stereo photogrammetry software, calculate the sea wave propagation speed, wave steepness η and wave tilt angle γ when meeting the ship, and generate a three-dimensional sea surface topography; S13. Invert the three-dimensional sea surface topography and perform slope steepness correction through the same-name point matching technology and resection method in the overlapping area of the stereo image pair to obtain the wavefront amplitude A, frequency f, and surface flow velocity B; S14. Obtain a sea wave radar image of 240 - 3000 m through the shipborne maritime radar, and then inversely derive the significant wave height H of the sea wave S , peak period T P , main wave direction Θ P and main wave wavelength L P ; S15. Combine the wave steepness η, the wave tilt angle γ, the wave front amplitude A, the frequency f, the surface current velocity B, the significant wave height H of the ocean wave S , the peak period T P , the main wave direction Θ P and the main wave length L P to form the ocean wave motion characteristic parameter array MC t ; S2. Data sequence processing: Combine the array of sea wave motion characteristic parameters and the array of ship motion characteristic parameters into an array of ship-sea wave motion characteristic parameters according to time sequence, and perform normalization processing on the array of ship-sea wave motion characteristic parameters in combination with zoning management to obtain one-dimensional serialized data; S3. Establish a long short-term memory neural network prediction model; S4. Train and verify the long short-term memory neural network model based on the sliding window segmentation method combined with the normalized array of ship-sea wave motion characteristic parameters; S5. Instantly predict the ship motion rest period under a specific sea state through the verified long short-term memory neural network prediction model combined with the real-time processing result of the ship-sea wave motion duration data.

2. The method for predicting the rest period of a ship's motion according to claim 1, characterized in that: In step S1, the inversion result of the sea wave radar image is corrected in real time through the inversion result of the optical stereo camera.

3. The method for predicting the rest period of a ship's motion according to claim 1, characterized in that: Step S2 further includes the following steps: S21. Obtain the angular velocity of the ship through the gyroscope, and after a single integration in combination with the geometric and physical parameters of the ship, obtain the rotation angles in three directions: roll θ C , pitch ψ C and yaw K C ; S22. Obtain the motion acceleration of the ship through an accelerometer, and obtain the translational velocities in three directions: sway HD, surge ZD, and heave CD after one integration in combination with the geometric and physical parameters of the ship; S23. Obtain the ship motion parameter array M at time t t , M t = [HD, ZD, CD, K C , Ψ C , θ C ; S24. Combine the ship motion parameter array M at time t t and the ocean wave motion characteristic parameter array MC t to obtain the ship-ocean wave coupled motion characteristic parameter array X t , where x t = [M t , MC t , and MC t = [A, B, η, T P , Θ P , λ P ; S25. With the ship's heading being 0°, divide a circular sea area with a radius of P within the effective detection range of the radar into N circular ring sectors, and the central angle α of each sector i-1 = 360° / N, where i = 1, 2,..., N; the ship's sailing speed is V m / s, and define the radius L of each circular ring as L = 2V, then the surrounding sea wave flow field is divided into P / L × N circular ring sectors, and the weight of the Q0 area consistent with the heading ≥ the weights of the Q1 and Q areas on the left and right of the Q0 area N-1 ≥ the weights of the remaining areas. When the ship is stationary, the weight of all areas is 1; S26. Serialize the sea area wave current field weights in the two-dimensional circular ring region into one dimension to obtain a one-dimensional weight sequence WQ = [Q0, Q1, Q N-1 ,,, Q N ,,,].

4. The method for predicting the rest period of a ship's motion according to claim 3, characterized in that: The geometric and physical parameters of the ship in step S21 and step S22 include: length, molded breadth, bottom length, deck depth, gross tonnage, designed speed, and designed draft.

5. The method for predicting the rest period of a ship's motion according to claim 1, characterized in that: The long short-term memory neural network model in step S3 includes N time steps and 2 LSTM network layers; the LSTM network layer includes a working unit cell, and the working unit cell realizes the persistence and inhibition of information through the structure of a forget gate f, an input gate i, and a control gate o.

6. The method for predicting the rest period of a ship's motion according to claim 5, characterized in that: The cell in the LSTM is updated at each moment through the following four-step calculation to obtain the state of the LSTM unit: Step 1. The forget gate f filters h t-1 and x t information and determines which information to discard from the previous moment's state. f t = σ(W f [h t-1 , x t + b f ) Among them: f t is the forget gate; σ() is the activation function, W f is the weight; b f is the bias term of the forget gate; Step 2. The input gate selects which information to retain through the sigmoid function and generates a candidate vector value through the activation function tanh; i t = σ(W i [h t-1 ,x t + b i ), where: i t is the input gate; W i and W c are the weights; b i and b c are the bias terms of the input gate and the input node; h t-1 is the output at time t-1; x t is the new variable value input at time t; tanh is the hyperbolic tangent function; Step 3. Update the cell state, multiply the old state by the forget gate, discard the information that is determined to be discarded, and then add the filtered new information to obtain the state of the current working unit; Where: C t is the unit state at time t; C t-1 is the unit state at time t-1; is the input state of the memory unit; Step 4. The layer processing value passes through the tanh function to determine the final output value. O t = σ(W o [h t-1 , x t + b o ), h t = O t * tanh(C t ) Where: O t is the output gate; W o is the weight; b o is the bias term of the output gate; h t is the output at time t.

7. A method for predicting the rest period of a ship's motion according to claim 1, characterized in that: The method for training the long short-term memory neural network model in step S4 is as follows: Let the number of the ship-wave time history observation motion sequences be T, and after step S2, the data sequence is [X i (1 ≤ i ≤ T); Select a sliding window of length N to limit the length of the input training sequence by the length of the sliding window, and use the iterative method to update the samples in the sliding window; group [X i (1 ≤ i ≤ T) into sliding window groups, with N + 1 data in each group. The first N data are used as input data to predict the value of the (N + 1)-th data Compare the error with the true value of the (N + 1)-th data, and iteratively adjust the network parameters; Define the loss function Use the gradient descent method to optimize the model. When the value of the loss function < 0.00005, the training is completed.

8. A method for predicting the rest period of a ship's motion according to claim 1, characterized in that: Step S5 further includes the following steps: S51. Assume that the motion state of the ship needs to be predicted for N steps. After iteration, the oldest data in the sliding window is replaced with the predicted value of the i-th step, where i = 1, 2, 3, …, N. Each time a replacement is made, the LSTM network is made to perform a new learning to update the network structure, and the next prediction is made with the new network structure. S52. Take the output roll angle, pitch angle, and heave of the LSTM model as the combined judgment variables, that is S53. Define the judgment threshold Y τ When the judgment variable Y of 15 consecutive outputs is less than the predetermined threshold Y τ the system initiates a prediction of the moment when the ship enters the resting period.

9. A method for predicting the rest period of a ship's motion according to any one of claims 1 to 8, characterized in that: When the sea state changes, repeat steps S1 to S5 to complete the prediction of the ship's motion rest period under the new sea state.

Images