Ship rolling state prediction method

Abstract

The invention discloses a ship rolling state prediction method, and belongs to the field of ship information processing. By constructing a data driving model, short-term prediction of the rolling state of the ship can be realized. Through an encoder-decoder structure, the input information is encoded into a context vector to be transmitted to the decoder, so that the lengths of input and output data can be different, the output data can be predicted by using longer input data, the model can learn more knowledge, and the prediction accuracy is improved. The method can be applied to a ship stabilization scene, the ship rolling state is predicted in advance, a rolling estimated value is provided for a stabilization system, and therefore design of a stabilization algorithm is supported.

Description

Technical Field

[0001] The invention belongs to the field of ship information processing, and in particular relates to a method for predicting a ship rolling state. Background Art

[0002] Smart ships have become a major trend, and how to use computer technology to reduce ship navigation risks and improve efficiency has become a concern for researchers. The ship's roll state is an important indicator for measuring ship safety, which can directly reflect the stability and navigation safety of the ship. Therefore, this example predicts the ship's roll angle by analyzing and modeling the ship's historical data, thereby improving the safety of marine operations.

[0003] The influence of the interaction of uncertain factors in the marine environment, the design characteristics of the ship itself, load distribution and other factors on the ship cannot be ignored, and it is very important to ensure the safety of the ship. At present, there are two main ways to build ship models: mathematical models and data-driven models. The roll motion of a ship involves the interaction and coupling effects of factors such as the ship's geometry, ship mass distribution, fluid motion, ship operation, and marine environment, and contains a large number of nonlinear factors. Therefore, it is difficult to accurately describe the roll motion of a ship with a digital model, and it is mainly achieved by using a data-driven model. For example: In 2023, Dan used the LSTM and CNN network parallel method in the article Adata driven method for multi-step prediction of ship roll motion in high seastates to achieve a ten-second prediction of ship roll. Yunyu used multi-objective fusion prediction and multi-layer error correction in the article Deterministic ship roll forecasting model based on multi-objective data fusion and multi-layer error correction to achieve a ten-second prediction based on the LSTM network.

[0004] At present, the prediction of ship roll angle is usually within the time range of ten seconds, and the prediction time length and accuracy are limited. The present invention proposes a method for constructing a ship roll prediction model, which analyzes and extracts features from a large amount of historical data of ships, constructs a full sea state data set, and thus builds and trains a ship roll prediction model, which can achieve accurate prediction of the ship's roll state in the next 120 seconds. Summary of the invention

[0005] The present invention aims to construct a ship short-term prediction model by extracting the relationship between the ship's rolling state and various characteristic data from historical data, and to make short-term predictions on the ship's rolling state.

[0006] In order to solve the deficiencies of the prior art, an embodiment of the present invention provides a data-driven construction method for a ship rolling state prediction model. To achieve the above-mentioned purpose, the present invention adopts the following technical solutions:

[0007] The present invention provides a method for predicting the rolling state of a ship, comprising the following steps:

[0008] Step 1: Obtain original ship data, including ship status, control instructions and ship environment information;

[0009] Step 2: Construct a ship state response training data set; preprocess the data in step 1, use the preprocessed data as the input sequence, and output a three-dimensional array;

[0010] Step 3: Build the data input layer; serialize the data set in step 2 into a form suitable for the time series model;

[0011] Step 4: Construct an encoder; input the serialized time series data obtained in step 3 into the encoder, convert it into a time series feature sequence representing the context information of the input sequence, and obtain a context vector;

[0012] Step 5: Construct a decoder; use the context vector in step 4 to calculate the initial state, and the decoder completes the decoding process by gradually generating the output of each time step of the target sequence; at each time step, the decoder passes the output of the previous time step and the input of the current time step to itself, and finally outputs a vector;

[0013] Step 6: Construct a fully connected layer; perform a linear transformation on the output vector obtained in step 5 through the fully connected layer, and map the transformed features to the new representation space through the nonlinear function tanh to obtain the target sequence generation result

[0014] Step 7: Training phase: Use forced learning training method to input training data train_X through the data input layer and encoder to obtain context vector v; the context vector is used as the initial state of the decoder, and the prediction result is generated through the decoder and fully connected layer. Through continuous training iterations, the network gradually approaches the target output train_Y;

[0015] Step 8: Input the ship data to be predicted into the trained prediction model and output the predicted value of the ship's rolling state.

[0016] Furthermore, in step 1, the ship data includes the ship's roll angle and its angular velocity, pitch angle and its angular velocity, surge velocity, sway velocity, rudder angle, rotation speed, wind direction, wind speed, sea wave direction, and wave height.

[0017] Furthermore, in step 2, before the constructed data set is input into the network, the data in step 1 needs to be preprocessed to regularize the data:

[0018]

[0019] Among them, μ is the data mean, μ=0; σ is the data variance, σ=1.

[0020] Furthermore, in step 4, the encoder is constructed by a Bi-LSTM layer; the Bi-LSTM is composed of a forward and a reverse LSTM layer, each LSTM layer is composed of a plurality of LSTM units, and each unit has a memory unit and an output gate; the serialized time series data is processed step by step through the Bi-LSTM layer, and the input of each time step is passed to the LSTM unit to capture the timing information.

[0021] Furthermore, in step 4, the context vector v:

[0022] v=(v1,v2)

[0023] Among them, v1 is the front eigenvector; v2 is the back eigenvector.

[0024] Furthermore, in step 5, the decoder is constructed by an LSTM layer, and the LSTM layer is composed of multiple LSTM units; the number of LSTM units in the decoder is twice that of the encoder.

[0025] Furthermore, in step 5, an attention mechanism v is added to the context vector i :

[0026]

[0027] Among them, each position h j The weight α ij ;

[0028]

[0029] Among them, e ij =a(s i-1 ,h j ) is the influence of the input around position j on the output position i, s i-1 is the hidden state of the encoder;

[0030] The present invention also provides a computer device / equipment / system, comprising a memory, a processor and a computer program stored in the memory, wherein the processor implements the steps of any of the above-mentioned ship roll state prediction methods when executing the computer program.

[0031] The present invention also provides a computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the steps of any of the above-mentioned ship roll state prediction methods.

[0032] The present invention also provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of any of the above-mentioned ship roll state prediction methods.

[0033] The beneficial effects of the present invention are:

[0034] (1) Compared with the existing prediction methods, the ship rolling state prediction method proposed in the present invention improves the learning ability of the model and prolongs the prediction time. Due to the complexity and uncertainty of the ship rolling motion itself, prolonging the prediction time has higher requirements on the prediction model.

[0035] (2) The ship roll state prediction method proposed in the present invention is more robust than the publicly available prediction methods. The prediction model is established based on a large amount of rich sea condition data and can be applied to relatively complex marine environments.

[0036] (3) Compared with the publicly available prediction methods, the ship roll state prediction method proposed in the present invention can change the length of the input and output sequences. Due to the prediction principle of LSTM, the input and output data must be of the same length. The present invention uses an encoder-decoder structure to encode the input information into a context vector and transmit it to the decoder, so that the length of the input and output data can be different. Therefore, longer input data can be used to predict the output data, and the model can learn more knowledge and improve the prediction accuracy. The length of the input data has a certain impact on the prediction ability of the model. If the input data is too short, it cannot provide the model with enough prediction information. However, if the input is too long, the previous information correlation is weak, which will interfere with the prediction results to a certain extent.

[0037] (4) The present invention can be applied in ship roll reduction scenarios to predict the ship's roll state in advance and provide a roll estimation value for the roll reduction system, thereby supporting the design of the roll reduction algorithm. In addition, by predicting the ship's roll state in advance, a dangerous warning for offshore operations can be issued in advance, thereby improving the safety of offshore operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A schematic block diagram of the process structure of a ship rolling behavior prediction model proposed by the present invention;

[0039] Figure 2 Schematic diagram of how the dataset is constructed;

[0040] Figure 3 This is a schematic diagram of the construction method of the encoder;

[0041] Figure 4 Schematic diagram of the construction method of the decoder;

[0042] Figure 5 Schematic diagram of the construction method of the fully connected layer;

[0043] Figure 6 This is a schematic diagram of the model training phase;

[0044] Figure 7 This is a schematic diagram of the model deduction stage. DETAILED DESCRIPTION

[0045] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0046] The present invention discloses a method for constructing a ship rolling behavior prediction model, comprising the following steps: S1: collecting ship historical data, collecting environmental information of the offshore waters of the ship, the position information of the ship, the control instructions of the ship, and the state information of the six degrees of freedom of the ship, including the speed and steering angle in each direction, based on radar and various sensors. Using the sensitivity analysis method Sobol, the influence of the characteristics on the ship rolling is analyzed.

[0047] S2: Construct training data set. Use the collected information to construct a ship state prediction training data set, adopt the three-dimensional array format set by the network, and extract data features based on the observation time length to predict the next ship roll state, so as to train the ship prediction model.

[0048] S3: Build the data input layer. The data input layer is mainly responsible for converting the original data into a format acceptable to the neural network and passing it to the next layer of the network for learning and processing.

[0049] S4: Construct an encoder. The encoder converts the input sequence into a fixed-length encoding representation to extract the time series data features. The encoder is mainly built based on the Bi-LSTM layer. Combined with forward and backward information transmission, it can effectively solve the gradient vanishing and gradient exploding problems in long sequence training, and can remember and capture long-term dependencies in the sequence for a long time.

[0050] S5: Construct the decoder. The decoder uses the encoding vector and the process of gradually generating the target sequence, which can use the context information extracted by the encoder and the predicted true value to generate an accurate target sequence. In the process of building the decoder, the attention mechanism is used for the context information, the soft connection is used for the impact of the historical time sequence on the future roll state, and the forced learning method is used to accelerate the convergence of the model.

[0051] S6: Construct a fully connected layer. Connect each neuron in the input layer to each neuron in the output layer to achieve nonlinear combination and transformation of features. The fully connected layer maps the high-dimensional features of the input to the low-dimensional representation of the output, thereby achieving dimensionality reduction and feature compression.

[0052] S7: Training phase. Use the prepared data set for training, use the data from the past period of time as the input sequence, and use the time series of the preset prediction time length as the target output. Use the forced learning training method, and use the real value instead of the decoder's own prediction result as the input.

[0053] S8: Model deduction stage. The prediction stage requires historical data of the same length, splits and reorganizes the trained model, and uses the model's own prediction results as the input of the decoder to iteratively generate prediction results.

[0054] Example

[0055] S1: Collect historical data. Collect ship sensor data, including global satellite navigation system data, inertial navigation data, radar data, etc. Use the Sobol sensitivity analysis network to analyze the data features that affect roll. Since the evaluation method of sensitivity analysis is the evaluation of scalars, it is difficult to use a scalar to represent a time series information feature. Therefore, after constructing multiple single-step models for sensitivity analysis, the results are statistically analyzed. The experimental results show that the statistical results of 5 steps are the same as those of 10 steps. Therefore, this result is used to construct the roll model. The ship data mainly include the ship's roll angle and speed, pitch angle and speed, bow angle and speed, surge speed, sway speed, heave speed, longitudinal displacement, transverse displacement, heave, control instructions including rudder angle, rotation speed and environmental information including wind direction, wind speed, sea wave direction and wave height (the impact of flow on ship behavior can be temporarily ignored because flow information is difficult to measure). From the 18 items of data, the ship's roll angle and speed, pitch angle and speed, surge speed, sway speed as well as control instructions and environmental information are found. A total of 12 dimensions of features have a greater impact on the roll.

[0056] S2: Data preprocessing. The ship rolling state prediction belongs to the time series prediction problem, which uses the data of a period of time in the past as the input sequence to predict multiple time steps in the future. The time series data consists of multiple time steps, each of which contains a feature vector. The time series data is divided into windows. The window size determines the input and output sequence lengths of the samples, such as Figure 2 As shown, set the observation time length t0 and the prediction time length t p, construct a training data set. The input data train_X is a three-dimensional array, and the shape of the data set is (number of samples, time interval, feature dimension). The number of samples refers to the number of independent samples contained in the data set; the time interval refers to the observation time length t0 of the input sample; the feature dimension refers to the number of features that affect the roll state in each sample, and 12-dimensional feature data is taken here. The output data train_Y is the ship roll angle in the next time step, which is also a three-dimensional array. The number of samples is the same as the input data, and the time interval refers to the predicted roll angle time length t p , the number of features is set to 1, because the output is only the roll angle after the current time step. Before the constructed dataset is input into the network, the data needs to be preprocessed and regularized, using the formula Where μ represents the mean of the data, σ represents the variance of the data, so that the variance of the X data set is 1 and the mean is 0.

[0057] S3: Construct the data input layer. As the starting point of the neural network, the data input layer mainly determines the format of the input data and ensures the correct transmission and processing of the data in the network. Each node in the data input layer corresponds to a feature dimension of the input data. In this example, the data input layer mainly defines 12 input feature dimensions. The input layer serializes the training data according to time steps to ensure the order and continuity of the data in the time dimension, and passes the serialized time series data S to the next layer of the neural network.

[0058] S4: Construct the encoder. The encoder is constructed by the Bi-LSTM layer. The Bi-LSTM consists of two LSTM layers, one for forward propagation and the other for backward propagation. Each LSTM layer consists of multiple LSTM units, each unit has a memory unit and an output gate to control the memory and output of information. The time series data T is processed step by step through the Bi-LSTM layer. The input of each time step is passed to the LSTM unit to capture the timing information, such as Figure 3 As shown in the figure, the encoder input is the serialized time series data output by the data input layer, and the output is the last step output information retained by the LSTM unit after it processes the input sequence step by step at each time step, which is called the context vector v = (v1, v2). v1 is the forward feature sequence, and v2 is the backward feature sequence, which contains the contextual time series features of the input sequence.

[0059] S5: Construct the decoder. The decoder uses the context vector v to calculate the initial state, which contains the information of the entire input sequence in the encoder, and is used to guide the decoder to generate the target sequence. Since the rolling state of the ship is affected by the time series data to different degrees, an attention mechanism is added to the context vector to determine the degree of correlation between the input and output positions. The specific implementation method is as follows:

[0060]

[0061] Each position h j The weight α ij

[0062]

[0063] in,

[0064] e ij =a(s i-1 ,h j )

[0065] Indicates the degree of influence of the input around position j on the output position i, the score e ij is based on the hidden state s of the encoder i-1 and the value of position j in the input sequence. The number of LSTM units used to construct the decoder is twice that of the encoder because the encoder produces twice the temporal information extracted by forward and backward propagation. The decoder completes the decoding process by gradually generating the output of each time step of the target sequence. At each time step, the decoder passes the output of the previous time step and the input of the current time step to the LSTM layer. The final output vector D of the decoder is as follows: Figure 4 shown.

[0066] S6: Construct a fully connected layer. The fully connected layer receives the decoder output vector D as input, performs a linear transformation through the fully connected layer, and maps the transformed features to a new representation space through the nonlinear function tanh, such as Figure 5 As shown, the final target sequence generation result is obtained

[0067] S7: Training phase. Use the training data set train_X and the corresponding target output train_Y. Input the training data train_X through the data input layer and the encoder to obtain the context vector v, which contains the context information and feature extraction results of the input sequence. The training phase uses a forced learning training method. For each training sample, the input of the decoder is the true value of the previous time step in the target sequence. That is, each time step has the correct sequence information to guide the decoder prediction, such as Figure 6 As shown in Figure 1, the context vector is used as the initial state of the decoder. The decoder and fully connected layer generate prediction results. Through continuous training iterations, the network gradually approaches the target output train_Y.

[0068] S8: Model deduction stage. In the model deduction stage, the input is a three-dimensional array test_X, which has the same shape as train_X, the number of samples is set to 1, the time interval is the observation time t0, the feature dimension is 12, and the output is the model's prediction based on the input data of the historical time step. Since the true value of the prediction cannot be obtained during the deduction process, the model layer iteration method is adopted. The data input layer and encoder are the same as the trained model. The layer structure of the decoder is reconstructed. At each time step, the output of the decoder in the previous time step is used as the input of the current time step, such as Figure 7 As shown in the figure, the 12-dimensional test data test_X with a preset observation time length t0 is input into the reconstruction model. The encoder will generate a context vector v, which is used as the initial state of the decoder. The decoder layer iterates and passes through the fully connected layer to output the predicted value of the ship's roll state at the current time step.

[0069] In particular, in some preferred embodiments of the present invention, a computer device is also provided, including a memory and a processor and a computer program stored in the memory, and when the processor executes the computer program, the steps of the ship rolling behavior prediction method in any of the above embodiments are implemented.

[0070] In some other preferred embodiments of the present invention, a computer-readable storage medium is provided, on which a computer program / instructions are stored. When the computer program is executed by a processor, the steps of the ship rolling behavior prediction method described in any of the above embodiments are implemented.

[0071] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiment method can be implemented by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes in the above-mentioned embodiment of the ship rolling behavior prediction method, which will not be repeated here.

[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting a ship's rolling state, characterized in that: The following steps are involved: Step 1: Obtain original ship data, including ship status, control instructions and ship environment information; Step 2: Construct a ship state response training data set; preprocess the data in step 1, use the preprocessed data as the input sequence, and output a three-dimensional array; Step 3: Build the data input layer; serialize the data set in step 2 into a form suitable for the time series model; Step 4: Construct an encoder; input the serialized time series data obtained in step 3 into the encoder, convert it into a time series feature sequence representing the context information of the input sequence, and obtain a context vector; Step 5: Construct a decoder; use the context vector in step 4 to calculate the initial state, and the decoder completes the decoding process by gradually generating the output of each time step of the target sequence; at each time step, the decoder passes the output of the previous time step and the input of the current time step to itself, and finally outputs a vector; Step 6: Construct a fully connected layer; perform a linear transformation on the output vector obtained in step 5 through the fully connected layer, and map the transformed features to a new representation space through the nonlinear function tanh to obtain the target sequence generation result y; Step 7: Training phase; Use the forced learning training method to pass the training data train_X through the data input layer, The encoder input is used to obtain the context vector v; the context vector is used as the initial state of the decoder, and the prediction result is generated through the decoder and the fully connected layer. Through continuous training iterations, the network gradually approaches the target output train_Y; Step 8: Input the ship data to be predicted into the trained prediction model and output the predicted value of the ship's rolling state.

2. A method for predicting ship rolling state according to claim 1, characterized in that: In step 1, the ship data includes the ship's roll angle and its angular velocity, pitch angle and its angular velocity, surge velocity, sway velocity, rudder angle, propeller speed, wind direction, wind speed, sea wave direction, and wave height.

3. A ship rolling state prediction method according to claim 1, characterized in that: In step 2, before the constructed data set is input into the network, the data in step 1 needs to be preprocessed to regularize the data: Among them, μ is the data mean, μ=0; σ is the data variance, σ=1.

4. A method for predicting ship rolling state according to claim 1, characterized in that: In step 4, the encoder is constructed by a Bi-LSTM layer; the Bi-LSTM is composed of a forward and a reverse LSTM layer, each LSTM layer is composed of a plurality of LSTM units, and each unit has a memory unit and an output gate; the serialized time series data is processed step by step through the Bi-LSTM layer, and the input of each time step is passed to the LSTM unit to capture the timing information.

5. A method for predicting ship rolling state according to claim 1, characterized in that: In step 4, the context vector v: v=(v1,v2) Among them, v1 is the front eigenvector; v2 is the back eigenvector.

6. A method for predicting ship rolling state according to claim 1, characterized in that: In step 5, the decoder is constructed by an LSTM layer, and the LSTM layer is composed of multiple LSTM units; the number of LSTM units in the decoder is twice that of the encoder.

7. A method for predicting ship rolling state according to claim 1, characterized in that: In step 5, an attention mechanism v is added to the context vector i : Among them, each position h j The weight α ij ; Among them, e ij =a(s i-1 ,h j ) is the influence of the input around position j on the output position i, s i-1 is the hidden state of the encoder.

8. A computer device / equipment / system comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images