Method and system for inverting three-dimensional sea wave spectrum along with ship motion based on deep learning model

Abstract

The invention belongs to the technical field of ship and ocean engineering, and discloses a method and a system for inverting a three-dimensional sea wave spectrum along with ship motion based on a deep learning model. According to the method, cross spectrum calculation is carried out among different ship motion time sequences to obtain a ship motion cross spectrum; carrying out data shuffling processing on the ship motion cross spectrums obtained in different time periods, and constructing a training set and a test set; and according to the constructed training set, inputting the training set into a CNN-ED-DE model, and carrying out CNN-ED-DE model training. And inputting the obtained input test data set into the trained CNN-ED-DE model, and solving to obtain a test data set inversion result. The method provided by the invention is completely based on motion sensor data such as a shipborne inertial measurement unit (IMU), does not need external remote sensing equipment or communication support, can estimate sea wave spectrum information in real time in any sailing state, and is suitable for online deployment in scenes such as complex course and unsteady sea conditions.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of ship and ocean engineering, and particularly relates to a method and system for inverting a three-dimensional sea wave spectrum based on ship motion based on a deep learning model. BACKGROUND

[0002] The acquisition of sea wave information is crucial for the operational safety, efficiency and sustainability of many marine-related activities. Real-time encounter wave information perception can provide decision support for autonomous ship navigation. For example, avoiding ship collisions, optimizing ship speed to reduce fuel consumption, and safe ship operation in rough seas. However, the lack of sea wave information in local waters poses a significant risk to ships navigating in these areas. In these areas, satellites cannot provide detailed ocean conditions to meteorological agencies. Although weather buoys can provide the required data, for cost and benefit, buoys are usually limited to near-shore locations. On the other hand, numerous ships are constantly navigating in various sea areas, using real-time, high-resolution sea wave information observed by the navigating ships to improve the reliability of global or regional wave atlas statistical data and global environmental measurement programs, and to ensure the overall safety and efficiency of maritime navigation. Common directional wave spectrum measurement methods include wave buoys, satellite observations, and shipborne wave radars. However, these methods have high initial costs and cannot accurately estimate the wave conditions encountered by a specific ship in real time. In the field of ship and ocean engineering, if the floating body in the wave is considered as a linear time-invariant system, then the ship's six-degree-of-freedom motion can be linked to the wave through a transfer function (also known as an amplitude-frequency response operator). Therefore, by using appropriate theoretical methods, the encountered sea wave information can be inverted from the ship's oscillation motion response. Therefore, as an alternative, using shipborne motion sensors to perceive the encountered sea wave spectrum is of great significance for shipborne wave information perception.

[0003] Related researchers have proposed wave frequency domain inversion techniques and wave time domain inversion techniques based on the idea of ship analogy to buoys. Among them, wave frequency domain inversion techniques mainly rely on ship motion response spectrum and transfer function to realize sea wave spectrum perception. Depending on whether the model is assumed in advance, it is divided into parameterized inversion methods and non-parameterized inversion methods. This type of technology is heavily dependent on the accuracy of ship motion response transfer function calculation. Wave time domain inversion techniques directly estimate the sea wave from the motion time history measured by sensors. They mainly include model-based methods and data-driven methods. Model-based methods mainly include Kalman filter methods and step-by-step estimation methods, which still cannot fundamentally get rid of the limitations of transfer functions.

[0004] Through the above analysis, the problems and defects of the prior art are: the existing published patents or papers mainly focus on the traditional physical model method inversion wave spectrum field, and these methods depend on the calculation accuracy of the ship motion response transfer function (RAO). The deep learning method is mainly for inverting wave time history, but the wave time history cannot provide more effective wave parameters. Compared with the time history, the three-dimensional sea wave spectrum information encountered in the field of ocean engineering is more concerned. SUMMARY

[0005] In order to overcome the problems in the related art, the present application discloses a method and system for inverting three-dimensional sea wave spectrum based on deep learning model.

[0006] The technical solution is as follows: the method for inverting three-dimensional sea wave spectrum based on deep learning model includes the following steps:

[0007] S1, obtain ship six-degree-of-freedom data, calculate the cross spectrum between different ship motion time series, and obtain ship motion cross spectrum;

[0008] S2, reconstruct the obtained ship motion cross spectrum data set, mix the ship motion cross spectrum obtained in different time periods, and construct a training set and a test set;

[0009] S3, input the constructed training set into the CNN-ED-DE model, and train the CNN-ED-DE model;

[0010] S4, input the constructed test data set into the trained CNN-ED-DE model, input the test data set into the trained CNN-ED-DE model, and obtain the test data set inversion result.

[0011] Step S1 specifically includes:

[0012] The ship six-degree-of-freedom motion time history data is X k , X k The superscript k in the formula represents different degrees of freedom of the ship, k=1 for heave, k=2 for pitch, k=3 for roll, k=4 for yaw, k=5 for surge, and k=6 for sway; the expression is:

[0013]

[0014] In the formula, represents the ship motion time history in the heave degree of freedom, the time interval between each data point is Δt, and n is the number of time series points, The superscript k in the formula represents the heave motion, and the subscript n represents the heave motion displacement at time n.

[0015] Further, the ship six-degree-of-freedom motion time history data X k Further, the ship six-degree-of-freedom motion time history data X

[0016]

[0017] wherein, is the k-degree-of-freedom ship motion time history after de-meaning, is the k-degree-of-freedom ship motion time history without de-meaning;

[0018] The window function is selected n = 0, 1, 2, …, N-1, w hw (n) is the window function, the subscript kw is the Hanning window function, and the windowed signal is:

[0019]

[0020] wherein, is the result of the de-meaning after the k-degree-of-freedom ship motion time history is windowed and processed;

[0021] The discrete frequency point is:

[0022]

[0023] wherein, ω q is the discrete frequency index, q is the discrete frequency point index, and N is the total number of samples;

[0024] The DFT change is performed on the windowed signal to obtain:

[0025]

[0026] wherein, x k [q] is the complex spectrum after the discrete Fourier transform, e is the exponential calculation, j is the imaginary unit, ω q is the discrete frequency index, t n is the time corresponding to the nth sampling point, t n = n·Δt.

[0027] Further, the calculation formula of the ship motion cross spectrum is:

[0028]

[0029] wherein, is the k-degree-of-freedom motion cross spectrum, x k is the k-degree-of-freedom ship motion time history, q is the discrete frequency point, q = 0, 1, … N-1; U is a constant term, U = 0.375;

[0030] The ship motion cross spectrum of different time periods is calculated by the above formula.

[0031] In step S2, the ship motion cross spectrum obtained in different time periods is subjected to data shuffling processing to construct a training set and a test set, including:

[0032] The ship motion cross spectrum calculated in different time periods and the corresponding two-dimensional wave spectrum are recorded as a data set D, expressed as:

[0033] D = {(x1, y1), (x2, y2), …, (xN, yN)} N N )

[0034] In the formula, (x1, y1) is the ship motion cross spectrum (x1) and the two-dimensional sea wave spectrum (y1) under a certain time period, N is the total number of samples; (x N N ) is the Nth ship motion cross spectrum and the two-dimensional sea wave spectrum set;

[0035] Define a permutation operation Ψ: {1, 2, …, N} → {1, 2, …, N}, which is a random permutation on the set {1, 2, …, N}, satisfying Ψ ∈ Sym(N), which is a set composed of all permutations of N elements, and Ψ is uniformly sampled;

[0036] After shuffling, the data set becomes:

[0037] D shuffled = {(x Ψ(1) ,y Ψ(1) ),(x Ψ(2) ,y Ψ(2) ),…,(x Ψ(N) ,y Ψ(N) )}

[0038] In the formula, (x Ψ(1) ,y Ψ(1) ) is the first ship motion cross spectrum and the two-dimensional sea wave spectrum set after shuffling, (x Ψ(N) ,y Ψ(N) ) is the Nth ship motion cross spectrum and the two-dimensional sea wave spectrum set after shuffling;

[0039] Take 80% of the data set D shuffled as the training set and 20% as the test set

[0040] In step S3, the CNN-ED-DE model training includes:

[0041] The obtained data is input into the training set, expressed as: ​​

[0042]

[0043] wherein, is the input of the i-th sample, is the ship motion cross spectrum, the number of channels is 25, and each channel contains F in frequency points, is the i-th three-dimensional sea wave spectrum, is a real set;

[0044] frequency direction Dir, the CNN-ED-DE model is trained to construct the mapping relationship from the ship motion cross spectrum to the three-dimensional sea wave spectrum, and to learn the nonlinear mapping function f θ , the expression is:

[0045]

[0046] wherein, is the input space of the model, that is, the real value matrix form of the ship motion cross spectrum data, wherein 25 is the number of input feature channels, and F in is the number of input frequency sampling points; is the output space of the model, that is, the discrete representation of the target two-dimensional sea wave spectrum, F out is the number of output sea wave spectrum frequency sampling points, and Dir is the number of output direction sampling points.

[0047] Further, the CNN-ED-DE model training further includes:

[0048] The convolutional encoding and decoding layer is four-scale splicing, which inputs the ship motion cross spectrum X (i) into a one-dimensional convolution kernel, and the expression is:

[0049] F z =Conv1D(X (i) ,z),z∈{1,3,5,7}

[0050] wherein, Conv1D is a one-dimensional convolution kernel, and z is the convolution kernel size;

[0051] The convolution results of the four scales are spliced, and the expression is:

[0052]

[0053] wherein, U0 is the spliced tensor after convolution, Concat is tensor splicing, is the dimension of the spliced tensor after convolution;

[0054] Batch normalization and nonlinear activation are performed, and the expression is:

[0055] U1=σ(BN(U0))

[0056] where U1 is the calculated tensor, σ is the ReLU activation function, and BN is the batch normalization;

[0057] The encoded features are flattened into a one-dimensional vector, and the expression is:

[0058] U = Flatten(U1)

[0059] where U is the one-dimensional vector after calculation, Flatten is the flattening process,

[0060] The tensor dimension is reshaped, and the expression is:

[0061] U dec = Unflatten(U)

[0062] where U dec is the reshaped tensor, and Unflatten is the dimension reshaping operation.

[0063] Further, the obtained U0, U1, U, and U dec are passed into the convolutional decoder layer, and are gradually upsampled by multiple two-dimensional transpose convolution layers, and the expression is:

[0064] U l+1 = σ(BN(ConvTranspose2D(U l ))), l = 0, 1, 2, …

[0065] where U l+1 is the feature tensor output by the l+1th layer of the decoder, ConvTranspose2D is the two-dimensional transpose convolution operation (deconvolution operation), U l is the input tensor of the lth layer of the decoder, and U0 = U dec ;

[0066] The final output is a single-channel two-dimensional tensor:

[0067]

[0068] where Y is the obtained three-dimensional sea wave spectrum, σ is the activation function, and Conv2D is the two-dimensional convolution operation;

[0069] The real three-dimensional sea wave spectrum Y (i) is trained using the mean square error loss, and the expression is:

[0070]

[0071] where χ is the loss value, F inis the number of frequency sampling points, is the number of direction sampling points, g is an index variable in the frequency dimension, h is an index variable in the direction dimension, is the value of the sea wave spectrum obtained by inversion of the model for the ith sample at frequency index g and direction index h, (i) (g,h) is the value of the real sea wave spectrum of the corresponding ith sample at frequency index g and direction index h.

[0072] Step S4 specifically comprises:

[0073] The ship motion cross spectrum data of the test set is input, and the expression is:

[0074]

[0075] The trained model is input with different test set samples, and a three-dimensional sea wave spectrum array of the ship motion cross spectrum inversion of the test set is solved, and the expression is:

[0076] Y (i) =S(ω,θ) (i) ,i∈1,2,…,N×0.2

[0077] In the formula, S(ω,θ) (i) is the three-dimensional sea wave spectrum corresponding to the ith sample, ω is the frequency, θ is the angle, and S is the three-dimensional sea wave spectrum function.

[0078] Another object of the present application is to provide a ship motion inversion three-dimensional sea wave spectrum system based on a deep learning model, which implements the ship motion inversion three-dimensional sea wave spectrum method based on the deep learning model, and the system comprises:

[0079] A ship motion cross spectrum obtaining module is configured to obtain ship six-degree-of-freedom data, and obtain ship motion cross spectrum by cross spectrum calculation between different ship motion time series;

[0080] A training set and test set constructing module is configured to reconstruct data sets by using the obtained ship motion cross spectrum data, and perform data shuffling processing on the ship motion cross spectrum obtained in different time periods to construct training sets and test sets;

[0081] A CNN-ED-DE model training module is configured to input the constructed training sets into the CNN-ED-DE model to perform CNN-ED-DE model training;

[0082] A test data set inversion result obtaining module is configured to input the constructed test data sets into the trained CNN-ED-DE model, and solve the test data set inversion result by using the trained CNN-ED-DE model.

[0083] In combination with all the above technical solutions, the application has the following beneficial effects:

[0084] Firstly, the application proposes a method for inverting three-dimensional sea wave spectrum based on ship motion based on a deep learning model. The method uses ship motion cross spectrum and uses convolutional coding and decoding deep learning model to realize real-time inversion of three-dimensional sea wave spectrum information. The inversion of three-dimensional sea wave spectrum information can obtain detailed statistical parameters of sea waves (significant wave height, characteristic period, spectral peak frequency, wave main direction, etc.). The existing patents and documents have not adopted the deep learning method to realize the inversion of ship motion to three-dimensional sea wave spectrum. There is no published literature and patent for the method of inverting three-dimensional sea wave spectrum based on deep learning model and using ship motion cross spectrum.

[0085] The application first combines multi-scale convolutional encoder (CNN Encoder) and two-dimensional deconvolution decoder (Decoder) to construct an end-to-end CNN-ED-DE deep learning neural network framework for directly inverting three-dimensional sea wave spectrum from ship motion cross spectrum. Compared with the traditional method relying on transfer function or numerical model, the method does not need prior physical model information, and has stronger adaptability and generalization ability.

[0086] Secondly, the method proposed by the application is completely based on shipborne inertial measurement unit (IMU) and other motion sensor data, without external remote sensing equipment or communication support, and can estimate sea wave spectrum information in real time under any navigation state, and is suitable for online deployment in complex navigation, unsteady sea conditions and other scenes.

[0087] The three-dimensional sea wave spectrum obtained by the method proposed by the application contains more detailed sea state parameters. For example: significant wave height, characteristic period, spectral peak frequency, wave main direction, etc. It can better provide detailed input information for intelligent navigation, collision avoidance control and operation decision of ships at sea.

[0088] Thirdly, the application can be widely used in autonomous navigation ships, offshore operation platforms, ocean observation networks and other fields. By using shipborne inertial or motion sensors, low-cost, real-time and high-precision shipborne sea wave spectrum sensing is realized, avoiding the deployment and maintenance cost of expensive wave buoys and radar systems. In commercial applications, the technology can provide key wave environment input for ship intelligent navigation, collision avoidance control, energy-saving speed optimization, etc., significantly improving navigation safety and operation efficiency; at the same time, it can form long-term accumulation of global high-resolution wave data, provide value-added data services for shipping companies, meteorological agencies and ocean big data platforms, promote the development of commercial meteorological and sea state information business, and has significant economic benefits and sustainable market competitiveness. Therefore, it has high practical application value.

[0089] Fourth, the current domestic and foreign patents and literature are mostly concentrated on the wave spectrum inversion method based on the physical model (transfer function) or the wave time history inversion method based on deep learning, and there is no research and patent on the combination of ship motion cross spectrum and deep convolutional encoding-decoding network (CNN-ED-DE) to realize the real-time inversion of sea wave spectrum of the ship in any navigation state. The present application first proposes a deep learning inversion framework without relying on the transfer function of the ship, which can automatically learn the nonlinear mapping relationship between the ship motion and the sea wave spectrum, significantly improving the applicability of the model, and is an innovative method in the current field of research.

[0090] For a long time, the field of ocean engineering and ship has always hoped to obtain the encountered sea wave spectrum in any navigation state in real time and accurately without relying on expensive wave measuring equipment and accurate modeling of the transfer function. However, the traditional method is always difficult to realize due to the limitations of transfer function calculation error, harsh model assumption conditions and high hardware cost. The present application uses the CNN-ED-DE deep learning model to automatically extract the wave spectrum features from the ship motion cross spectrum data, realizes the high-precision, low-cost and real-time wave spectrum inversion, and successfully breaks through this long-existing technical bottleneck. BRIEF DESCRIPTION OF DRAWINGS

[0091] The accompanying drawings, which are incorporated into and form part of the specification, illustrate embodiments consistent with the present disclosure and, together with the specification, serve to explain the principles of the present disclosure;

[0092] Figure 1 is a flow chart of a method for inverting a three-dimensional sea wave spectrum based on a deep learning model according to an embodiment of the present application;

[0093] Figure 2 is a principle diagram of a method for inverting a three-dimensional sea wave spectrum based on a deep learning model according to an embodiment of the present application;

[0094] Figure 3 is a schematic diagram of ship motion time history data according to an embodiment of the present application;

[0095] Figure 4 is a schematic diagram of a simulated sea wave spectrum according to an embodiment of the present application;

[0096] Figure 5 is a schematic diagram of ship motion cross spectrum (heave-pitch) according to an embodiment of the present application;

[0097] Figure 6 is a schematic diagram of a model inversion sea wave spectrum according to an embodiment of the present application;

[0098] Figure 7 is a network diagram of a CNN-ED-DE model according to an embodiment of the present application. DETAILED DESCRIPTION

[0099] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0100] The innovation of this invention lies in its pioneering use of the cross-spectrum of multi-degree-of-freedom ship motion as input to a deep learning model. Employing a convolutional encoder-decoder (CNN-ED-DE) network structure, it automatically learns the complex nonlinear mapping relationship between ship motion and wave spectrum, achieving real-time, high-precision inversion of three-dimensional wave spectra under different navigation states and sea states without relying on the ship's transfer function. This invention overcomes the limitations of traditional methods, such as dependence on the accuracy of transfer function calculation and the high cost of wave measurement equipment. It boasts advantages such as strong model robustness, wide applicability, and fast inversion speed. It can construct a high-resolution wave observation network globally using ships, significantly improving the safety and intelligence of maritime operations and autonomous navigation.

[0101] Example 1, such as Figure 1 As shown in the embodiment of the present invention, the method for inverting three-dimensional wave spectra based on deep learning models during ship motion includes:

[0102] S1. Obtain the six degrees of freedom data of the ship and calculate the cross spectrum of the ship motion by performing cross spectrum calculation between different ship motion time series;

[0103] For example, specifically including: the known motion history data of a ship with six degrees of freedom is X. k X k The superscript k in the equation represents different degrees of freedom of the ship: k=1 for heave, k=2 for pitch, k=3 for roll, k=4 for bow roll, k=5 for pitch, and k=6 for sway; the expression is:

[0104]

[0105] In the formula, The time history represents the ship's motion under heave degrees of freedom, with the time interval between each data point being Δt, and n being the number of time series points. In this context, the superscript k represents heave motion, and the subscript n represents the displacement of the heave motion at time n.

[0106] The motion history data of the ship's six degrees of freedom X k The mean-reduction process is performed using the following expression:

[0107]

[0108] wherein, is the k degree of freedom ship motion time history after de-meaning, is the k degree of freedom ship motion time history without de-meaning;

[0109] The window function is selected n = 0, 1, 2, …, N-1, w hw (n) is the window function, the subscript hw is the Hanning window function, and the windowed signal is:

[0110]

[0111] wherein, is the result of the windowed processing of the de-meaned ;

[0112] The discrete frequency points are:

[0113]

[0114] wherein, ω q is the discrete frequency index, q is the discrete frequency point index, and N is the total number of samples;

[0115] The DFT change of the windowed signal is:

[0116]

[0117] wherein, x k [q] is the complex spectrum after the discrete Fourier transform, e is the exponential calculation, j is the imaginary unit, and ω q is the discrete frequency index, t n is the time corresponding to the nth sampling point, t n = n·Δt.

[0118] The calculation formula of the ship motion cross spectrum is:

[0119]

[0120] wherein, is the k degree of freedom and the k degree of freedom motion cross spectrum, x k is the k degree of freedom ship motion time history, q is the discrete frequency point, q = 0, 1, … N-1; U is a constant term, U = 0.375;

[0121] The ship motion cross spectrum of different time periods of six degrees of freedom is calculated through the above formula.

[0122] As can be seen, this invention innovatively proposes the above formulas. These respectively represent the demeaning processing of the ship motion history for different degrees of freedom and the DFT transformation after windowing using a window function to obtain the ship motion cross spectrum.

[0123] S2, the obtained ship motion cross spectrum data is reconstructed into a dataset, and the ship motion cross spectrum obtained from different time periods is mixed and processed to construct a training set and a test set;

[0124] For example, the specific method of step S2 is as follows: the cross spectrum of ship motion calculated at different time periods and the corresponding two-dimensional wave spectrum are denoted as dataset D, D={(x1,y1),(x2,y2),…,(x N ,y N )},

[0125] In the formula, (x1, y1) represents the cross spectrum of ship motion (x1) and the two-dimensional wave spectrum (y1) over a certain time period, and N is the total number of samples; (x N ,y N ) represents the set of the Nth ship motion cross spectrum and the two-dimensional wave spectrum;

[0126] Define a permutation operation Ψ: {1,2…,N}→{1,2…,N}, which is a random permutation on the set {1,2…,N}, satisfying Ψ∈Sym(N), which is the set of all permutations of N elements, and Ψ is uniformly sampled;

[0127] This invention innovatively proposes that the dataset becomes the following after shuffling:

[0128] D shuffled ={)x Ψ(1) ,y Ψ(1) ),(x Ψ(2) ,y Ψ(2) ),…,(x Ψ(N) ,y Ψ(N) )}

[0129] In the formula, (x Ψ(1) ,y Ψ(1) (x) represents the set of the first ship motion cross spectrum after mixing and the two-dimensional wave spectrum. Ψ(N) ,y Ψ(N) ) represents the set of the Nth ship motion cross spectrum after mixing and the two-dimensional wave spectrum;

[0130] Take dataset D shuffled 80% as training set 20% as test set

[0131] S3, input the training set obtained by construction into the CNN-ED-DE model, and perform CNN-ED-DE model training;

[0132] For example, the specific steps are as follows: according to the input training set obtained in step S2 The input of the i-th sample is a ship motion cross spectrum, the number of channels is 25, each channel contains F in frequency points, is the i-th three-dimensional sea wave spectrum, is a real set;

[0133] frequency direction Dir, the CNN-ED-DE model training constructs a mapping relationship from the ship motion cross spectrum to the three-dimensional sea wave spectrum, and learns a nonlinear mapping function f θ , the expression is:

[0134]

[0135] In the formula, is a real value matrix form of the model input space, that is, the ship motion cross spectrum data, wherein 25 is the number of input feature channels, F in is the number of input frequency sampling points; is a discrete representation of the model output space, that is, the target two-dimensional sea wave spectrum, F out is the number of output sea wave spectrum frequency sampling points, and Dir is the number of output direction sampling points.

[0136] It can be known that the nonlinear mapping function f θ of the present application represents the nonlinear mapping function relationship between the ship motion cross spectrum and the two-dimensional sea wave spectrum, and represents a real set, and the superscript is a specific data dimension form according to the ship motion inversion sea wave spectrum.

[0137] The CNN-ED-DE model training further includes:

[0138] The convolutional encoding and decoding layer is four scales splicing, the ship motion cross spectrum X (i) is input into a one-dimensional convolution kernel, and the expression is:

[0139] F z =Conv1D(X (i) ,z),z∈{1,3,5,7}

[0140] In the formula, Conv1D is a one-dimensional convolution kernel, and z is a convolution kernel size;

[0141] The convolution results of the four scales are spliced, and the expression is:

[0142]

[0143] wherein U0 is the concatenated tensor after convolution, Concat is the tensor concatenation, is the dimension of the concatenated tensor after convolution;

[0144] Batch normalization and nonlinear activation are performed, and the expression is:

[0145] U1 = σ(BN(U0))

[0146] wherein U1 is the calculated tensor, σ is the ReLU activation function, and BN is the batch normalization;

[0147] The encoded features are flattened into a one-dimensional vector, and the expression is:

[0148] U = Flatten(U1)

[0149] wherein U is the one-dimensional vector after calculation, Flatten is the flattening process,

[0150] Tensor dimension reshaping is performed, and the expression is:

[0151] U dec = Unflatten(U)

[0152] wherein U dec is the reshaped tensor, and Unflatten is the dimension reshaping operation.

[0153] The obtained U0, U1, U, and U dec are transmitted to the convolutional decoder layer, and are gradually upsampled by multiple two-dimensional transpose convolutions, and the expression is:

[0154] U l+1 = σ(BN(ConvTranspose2D(U l ))), l = 0, 1, 2, …

[0155] wherein U l+1 is the feature tensor output by the l+1th layer of the decoder, ConvTranspose2D is the two-dimensional transpose convolution operation (deconvolution operation), U l is the input tensor of the lth layer of the decoder, and U0 = U dec ;

[0156] The final output is a single-channel two-dimensional tensor:

[0157]

[0158] wherein is the three-dimensional wave spectrum obtained by inversion, ReLU is the activation function, and Conv2D is the two-dimensional convolution operation;

[0159] with real three-dimensional sea spectrum Y (i) The training is performed by using a mean square error loss, and the expression is:

[0160]

[0161] wherein x is a loss value, F in is the number of input frequency sampling points, Dir is the number of direction sampling points, g is an index variable in the frequency dimension, h is an index variable in the direction dimension, is a value of the sea spectrum obtained by the model for the i-th sample at the frequency index g and the direction index h, Y (i) (g, h) is a value of the real sea spectrum of the i-th sample at the frequency index g and the direction index h.

[0162] S4, the test data set constructed is input into the trained CNN-ED-DE model, and the test data set is solved by the trained CNN-ED-DE model to obtain the inversion result of the test data set.

[0163] For example, the specific method of the step S4 is that, according to the application, the ship motion cross spectrum data of the test set is input, and the expression is:

[0164]

[0165] The ship motion cross spectrum data of the test set is input into the trained model, and a three-dimensional sea spectrum array of the ship motion cross spectrum inversion of the test set is solved, and the expression is:

[0166] Y (i) = S(ω, θ) (i) , i ∈ 1, 2,..., N x 0.2

[0167] wherein S(ω, θ) (i) is a three-dimensional sea spectrum corresponding to the i-th sample, ω is a frequency, θ is an angle, and S is a three-dimensional sea spectrum function.

[0168] It can be known that the ship motion cross spectrum data of the test set represented by the ship motion cross spectrum data according to the application is 0.2 times, i.e., 20%, of the total data amount N;

[0169] The three-dimensional sea spectrum array Y (i) of the ship motion cross spectrum inversion according to the application represents an inversion two-dimensional sea spectrum, and there are i two-dimensional arrays in total.

[0170] Another example is shown in FIG. 1. Figure 2 FIG. 1 shows a schematic diagram of a method for inverting a three-dimensional sea spectrum according to ship motion based on a deep learning model.

[0171] The embodiment 2 provides a ship motion inversion three-dimensional sea wave spectrum system based on a deep learning model.

[0172] The ship motion cross spectrum obtaining module is used for obtaining ship six-degree-of-freedom data, and obtaining ship motion cross spectrum through cross spectrum calculation between different ship motion time sequences.

[0173] The training set and test set constructing module is used for reconstructing a data set through the obtained ship motion cross spectrum data, performing data mixing processing on the ship motion cross spectrum obtained in different time periods, and constructing a training set and a test set.

[0174] The CNN-ED-DE model training module is used for inputting the constructed training set into the CNN-ED-DE model, and performing CNN-ED-DE model training.

[0175] The test data set inversion result obtaining module is used for inputting the constructed test data set into the trained CNN-ED-DE model, performing calculation on the input test data set through the trained CNN-ED-DE model, and obtaining a test data set inversion result.

[0176] It is known from the above embodiment that the wave state has a direct influence on the safety and operation efficiency of ship navigation, especially in the strong wind and wave sea state, and real-time acquisition of sea wave spectrum information encountered by the ship is of great significance for autonomous navigation, collision avoidance decision, energy consumption optimization and the like. However, traditional wave measurement methods such as wave buoys and satellite remote sensing have problems such as high cost, poor timeliness, insufficient spatial resolution and the like, and it is difficult to meet the perception needs of specific ships in the navigation process for real-time and local sea wave information. The present application provides a method for inversion of three-dimensional sea wave spectrum based on a deep learning model, which uses the multi-degree-of-freedom motion cross spectrum data recorded by the ship-borne motion sensor, automatically learns the complex nonlinear mapping relationship between the ship motion and the encountered sea wave spectrum through the convolutional neural network, realizes real-time and high-precision estimation of the three-dimensional sea wave spectrum of the ship in any navigation state. The method does not depend on the prior ship transfer function information, has the advantages of strong model robustness, wide application range and high inversion precision, and can provide key wave environment input for intelligent navigation, collision avoidance control and operation decision of the ship at sea, and significantly improve the safety and intelligent level of the sea operation.

[0177] ​​The application provides a method for inversing three-dimensional sea wave spectrum based on ship motion by using a deep learning model. By inputting ship motion cross spectrum, the deep learning model is used to automatically learn the complex nonlinear mapping relationship between ship motion and encountered sea wave spectrum, so that real-time and high-precision estimation of sea wave spectrum of the ship under any navigation state is realized. The three-dimensional sea wave spectrum obtained by the method can provide detailed statistical parameters (significant wave height, characteristic period, spectral peak frequency, and wave main direction) of the sea wave, and the method has strong robustness, can be quickly transplanted to different ship types, and has high application value.

[0178] To further illustrate the effects of the embodiments of the application, the following experiments are performed.

[0179] Experimental example: In this example, ship motion simulation data of a 160m-long ship at zero speed and corresponding sea wave spectrum data are used. The ship motion data is generated by using a two-dimensional half method and linear superposition, the sea wave spectrum is selected as a Jonswap spectrum, and the direction function is a simple direction function. The working conditions of the simulated sea wave data are set as: wave height H s ={1m, 6m}, period T p ={4s, 14s}, and wave direction β={0°, 360°}. Random sampling is performed on the three parameters, and the sample number is 1000. Ship motion simulation (1200s) and sea wave spectrum simulation are performed on the data under 1000 working conditions. The ship motion simulation is as shown in Figure 3 , and the sea wave spectrum simulation result is as shown in Figure 4 .

[0180] Specifically, the method for inversing three-dimensional sea wave spectrum based on a deep learning model includes the following steps: Figure 2

[0181] Step 1: Obtain ship six-degree-of-freedom data, and obtain ship motion cross spectrum by cross spectrum calculation between different time sequences. The detailed method of step 1 is as follows: given ship six-degree-of-freedom motion time history data X k is X k The superscript k in the formula represents different degrees of freedom of the ship, k=1 represents heave, k=2 represents pitch, k=3 represents roll, k=4 represents yaw, k=5 represents surge, and k=6 represents sway;

[0182] represents ship motion time history of the heave degree of freedom, the time interval between each data point is Δt, and n is the number of time sequence points, The superscript k in the formula represents heave motion, and the subscript n represents heave motion displacement at the n time. The ship six-degree-of-freedom motion time history data X k is subjected to mean removal processing, and a window function ​n = 0, 1, 2, ..., N-1, w hw (n) represents the window function, and the subscript hw indicates the Hanning window function. The windowed signal after windowing is: Discrete frequency points are q = 0, 1, ..., N-1; DFT transformation is performed on the windowed signal to obtain... The formula for calculating the cross spectrum of ship motion is: k = 1, 2, 3, 4, 5, 6 Let x be the cross spectrum of motions with k degrees of freedom and k degrees of freedom. k Let be the time history of ship motion for k degrees of freedom, q be discrete frequency points, q = 0, 1, ..., N-1; U be a constant term, U = 0.375; the cross-spectrum of ship motion for different time periods of the six degrees of freedom is calculated using the above formula, as follows: Figure 5 As shown.

[0183] Step 2: Reconstruct the dataset from the data in Step 1. Perform data shuffling on the cross-spectrums of ship motion obtained from different time periods to construct training and testing sets. Specifically, Step 2 involves denoting the cross-spectrums calculated from different time periods and their corresponding three-dimensional wave spectra as dataset D, where D = {(x1,y1),(x2,y2),…,(x…}. N ,y N )}, where (x1,y1) represents the cross spectrum of ship motion (x1) and the two-dimensional wave spectrum (y1) over a certain time period, and N is the total number of samples; (x N ,y N Let be the set of the Nth ship motion cross spectrum and the two-dimensional wave spectrum; define a permutation operation Ψ: {1,2…,N}→{1,2…,N}, which is a random permutation on the set {1,2…,N} satisfying Ψ∈Sym(N), and is the set of all permutations of N elements, where Ψ is uniformly sampled; after shuffling, the dataset becomes D. shuffled ={(x Ψ(1) ,y Ψ(1) ),(x Ψ(2) ,y Ψ(2) ),…,(x Ψ(N) ,y Ψ(2) )}, take dataset D shuffled 80% as training set 20% as test set

[0184] Step 3: Based on the training set obtained in Step 2 The input is fed into the CNN-ED-DE model for model training, such as... Figure 7 As shown. The specific steps are as follows: based on the input training set obtained in step 2... In the formula, The input for the i-th sample is a cross-spectrum of ship motion, with 25 channels, each channel containing F. in Each frequency point, For the i-th three-dimensional wave spectrum, For the set of real numbers; frequency In the direction Dir, the CNN-ED-DE model training constructs a mapping relationship between the cross spectrum of ship motion and the three-dimensional wave spectrum, learning the nonlinear mapping function f between them. θ The expression is

[0185] Step 4: Input test dataset obtained in Step 2 The input is fed into the model trained in step 3, and the input test dataset is used to solve the model to obtain the test dataset inversion result. The specific method of step 4 is as follows: input the ship motion cross-spectral data of the test set: For i∈1,2,…,N×0.2, different test set samples can be solved using the model to obtain the three-dimensional wave spectrum array Y of the ship motion cross spectrum inversion of the test set. (i) =S(ω,θ) (i) For each i ∈ 1, 2, ..., N × 0.2, the following shows the results of plotting the wave spectrum array retrieved using the CNN-ED-DE model. For example... Figure 6 As shown.

[0186] The two-dimensional array of wave spectrum was used to invert the cross spectrum of ship motion for post-processing calculations. The statistical results of the sea state parameter error calculations for some post-processing are shown in Table 1.

[0187] Table 1. Statistical Table of Errors in Three-Dimensional Wave Spectrum Inversion from Ship Motion Cross-Spectrum

[0188] H s ]]> [CAT p ]]> β True value 5.26 8.14 50 Inverted value 5.1 8.52 50 Relative error 3％ 4.7％ 0％

[0189] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for retrieving three-dimensional wave spectra based on ship motion using a deep learning model, characterized in that, The method includes the following steps: S1. Obtain the six degrees of freedom data of the ship and calculate the cross spectrum of the ship motion by performing cross spectrum calculation between different ship motion time series; S2, the obtained ship motion cross spectrum data is reconstructed into a dataset, and the ship motion cross spectrum obtained from different time periods is mixed and processed to construct a training set and a test set; S3, input the constructed training set into the CNN-ED-DE model to train the CNN-ED-DE model; S4. Input the constructed test dataset into the trained CNN-ED-DE model. The input test dataset is solved by the trained CNN-ED-DE model to obtain the test dataset inversion result.

2. The method for retrieving three-dimensional wave spectra based on ship motion according to claim 1, characterized in that, Step S1 specifically includes: The motion history data of the ship's six degrees of freedom is X k X k The superscript k in the equation represents different degrees of freedom of the ship: k=1 for heave, k=2 for pitch, k=3 for roll, k=4 for bow roll, k=5 for pitch, and k=6 for sway; the expression is: In the formula, The time series represents the ship's motion under heave degrees of freedom, with the time interval between each data point being Δt, and n being the number of time series points. In this context, the superscript k represents heave motion, and the subscript n represents the displacement of the heave motion at time n.

3. The method for retrieving three-dimensional wave spectra based on ship motion according to claim 2, characterized in that, The motion history data of the ship's six degrees of freedom X k The mean-reduction process is performed using the following expression: In the formula, The time history of ship motion under k degrees of freedom after mean normalization. The time history of ship motion under k degrees of freedom without demeaning processing; Select window function n = 0, 1, 2, ..., N-1, w hw (n) represents the window function, and the subscript hw represents the Hanning window function. The windowed signal after windowing is: In the formula, For the mean-reduced processing The result of windowing post-processing; Discrete frequency points are: In the formula, ω q is the discrete frequency index, g is the discrete frequency point index, and N is the total number of samples; The windowed signal was transformed using a DFT to obtain: In the formula, x k [q] represents the complex spectrum after the discrete Fourier transform, e is the exponent, j is the imaginary unit, and t... n Let t be the time corresponding to the nth sampling point. n = n·Δt.

4. The method for retrieving three-dimensional wave spectrum based on ship motion according to claim 3, wherein the calculation formula for the ship motion cross spectrum is: In the formula, Let x be the cross spectrum of motions with k degrees of freedom and k degrees of freedom. k Let be the time history of ship motion with k degrees of freedom, q be discrete frequency points, q = 0, 1, ..., N-1; and U be a constant term. The above formula is used to calculate the cross spectrum of ship motion at different time periods for six degrees of freedom.

5. The method for retrieving three-dimensional wave spectra based on ship motion according to claim 1, characterized in that, In step S2, the cross-spectrum of ship motion obtained from different time periods is shuffled to construct training and testing sets, including: The cross-spectrum of ship motion calculated at different time periods and the corresponding two-dimensional wave spectrum are denoted as dataset D, and the expression is: D={(x1,y1),(x2,y2),…,(x N ,y N )} In the formula, (x1, y1) represents the cross spectrum of ship motion (x1) and the two-dimensional wave spectrum (y1) over a certain time period, and N is the total number of samples; (x N ,y N ) represents the set of the Nth ship motion cross spectrum and the two-dimensional wave spectrum; Define a permutation operation Ψ: {1,2…,N}→{1,2…,N}, which is a random permutation on the set {1,2…,N}, satisfying Ψ∈Sym(N), which is the set of all permutations of N elements, and Ψ is uniformly sampled; After shuffling, the dataset becomes: D shuffled ={(x Ψ(1) ,and Ψ(1) ),(x Ψ(2) ,and Ψ(2) ),…,(x Ψ(N) ,and Ψ(N) )} In the formula, (x Ψ(1) ,y Ψ(1) (x) represents the set of the first ship motion cross spectrum after mixing and the two-dimensional wave spectrum. Ψ(N) ,y Ψ(N) ) represents the set of the Nth ship motion cross spectrum after mixing and the two-dimensional wave spectrum; Take dataset D shuffled 80% as training set 20% as test set 6. The method for retrieving three-dimensional wave spectra based on ship motion according to claim 1, characterized in that, In step S3, training the CNN-ED-DE model includes: Input the obtained data into the training set, as shown in the expression: In the formula, The input for the i-th sample is a cross-spectrum of ship motion, with 25 channels, each channel containing F. in Each frequency point, For the i-th three-dimensional wave spectrum, It is the set of real numbers; frequency In the direction Dir, the CNN-ED-DE model training constructs a mapping relationship between the cross spectrum of ship motion and the three-dimensional wave spectrum, learning the nonlinear mapping function f between them. θ The expression is: In the formula, The input space of the model is a real-valued matrix of the cross-spectral data of ship motion, where 25 represents the number of input feature channels, and F... in The number of frequency sampling points for input; F is the discrete representation of the target two-dimensional wave spectrum in the model output space. out Dir represents the number of frequency sampling points for the output wave spectrum, and Dir represents the number of directional sampling points for the output.

7. The method for retrieving three-dimensional wave spectra based on ship motion according to claim 6, characterized in that, Training the CNN-ED-DE model further includes: The convolutional encoder-decoder layer is stitched together at four scales to cross-spectrum X of ship motion. (i) The input to a one-dimensional convolution kernel is expressed as: F z =Conv1D(X (i) ,z),z∈{1,3,5,7} In the formula, Conv1D is a one-dimensional convolution kernel, and z is the kernel size; The concatenation results of the four scales are concatenated as follows: In the formula, U0 is the concatenated tensor after convolution, and Concat is the tensor concatenation. The concatenation of tensors after convolution; Batch normalization and nonlinear activation are performed, and the expression is: U1=σ(BN(U0)) In the formula, U1 is the calculated tensor, σ is the ReLU activation function, and BN is the batch normalization; The encoded features are flattened into a one-dimensional vector, as shown in the expression: U = Flaatten(U1) In the formula, U is the calculated one-dimensional vector, and Flatten represents the flattening process. To reshape the tensor dimensions, the expression is: U dec =Unflatten(U) In the formula, U dec For the reshaped tensor, Unflatten is the dimension reshaping operation.

8. The method for retrieving three-dimensional wave spectra based on ship motion according to claim 7, characterized in that, The obtained U0, U1, U, U dec The data is fed into the convolutional decoder layer and upsampled progressively through multiple layers of 2D transposed convolutions, as shown in the expression: IN l+1 =σ(BN(ConvTranspose2D(U l ))),l=0,1,2,… In the formula, U l+1 U is the feature tensor output from the (l+1)th layer of the decoder. ConvTranspose2D is the two-dimensional transpose convolution operation (deconvolution operation). l Let U0 be the input tensor of the l-th layer of the decoder. dec ; The final output is a single-channel two-dimensional tensor: In the formula, The three-dimensional wave spectrum obtained by inversion is represented by ReLU as the activation function and Conv2D as the two-dimensional convolution operation. With real three-dimensional ocean wave spectrum Y (i) Training is performed using mean squared error loss, expressed as: In the formula, χ is the loss value, and F in Here, represents the number of frequency sampling points, Dir represents the number of direction sampling points, g represents the index variable in the frequency dimension, and h represents the index variable in the direction dimension. For the wave spectrum retrieved by the model for the i-th sample, the values ​​of Y at the frequency index g and direction index h are given. (i) (g,h) represents the values ​​of the actual wave spectrum of the corresponding i-th sample at the frequency index g and the direction index h.

9. The method for retrieving three-dimensional wave spectra based on ship motion according to claim 1, characterized in that, Step S4 specifically includes: The input test set contains ship motion cross-spectral data, expressed as follows: Different test set samples are input into the trained model, and the three-dimensional wave spectrum array derived from the cross-spectral inversion of ship motion in the test set is obtained, expressed as: Y (i) =S(ω,θ) (i) ,i∈1,2,…,N×0.2 In the formula, S(ω,θ) (i) Let ω be the three-dimensional wave spectrum corresponding to the i-th sample, θ be the frequency, and S be the three-dimensional wave spectrum function.

10. A system for retrieving three-dimensional wave spectra based on ship motion using a deep learning model, characterized in that, The method for retrieving three-dimensional wave spectra based on a deep learning model according to any one of claims 1-9, the system comprising: The ship motion cross spectrum acquisition module is used to acquire six-degree-of-freedom ship data and obtain the ship motion cross spectrum by performing cross spectrum calculations between different ship motion time series. The training and test set construction module is used to reconstruct the dataset from the obtained ship motion cross spectrum data, perform data shuffling on the ship motion cross spectra obtained from different time periods, and construct the training and test sets. The CNN-ED-DE model training module is used to train the constructed training set. The input is fed into the CNN-ED-DE model for training. The module for obtaining the test dataset inversion results is used to convert the constructed test dataset. The input is fed into the trained CNN-ED-DE model. The input test dataset is then processed by the trained CNN-ED-DE model to obtain the inversion result of the test dataset.

Images