Very short-term real-time prediction method for ship motion posture

Abstract

The present invention provides a very short-term real-time prediction method for hull motion posture, which comprises: under the condition of satisfying Shannon sampling theorem, collecting hull motion posture data with a large sampling period, establishing an AR model, using the Bayesian information BIC criterion to determine the order of the AR model, using a limited memory recursive least squares algorithm to perform online update of AR model parameters, using a rolling prediction method to perform multi-step prediction of hull motion posture, combining the predicted data with historical data of a certain length, establishing an RBF neural network model for fitting interpolation, and obtaining very short-term prediction data of hull motion posture. The present invention uses a limited memory recursive least squares algorithm for online updating, establishes an RBF neural network model for fitting interpolation, can effectively improve the real-time performance of the prediction algorithm, and has high prediction accuracy.

Description

Technical Field

[0001] The invention belongs to the technical field of on-line prediction of ship motion posture, in particular to an extremely short-term real-time prediction method of ship motion posture. Background Art

[0002] As countries attach more and more importance to marine resources, maritime activities are becoming more frequent, such as the development of marine resources, maritime rescue and cargo transportation, warship cruises and special operations, etc. As an important means of transportation for people to carry out marine operations, ships produce 6 degrees of freedom under the action of waves, which brings inconvenience and danger to marine operations.

[0003] The shipborne 6-DOF stable platform provides a stable working environment for offshore operations. In order to keep the platform stable, it is necessary to control the stable platform in time according to the motion control of each degree of freedom of the hull. However, when the platform load is large or the control system performance is poor, the control system will have a large lag and cannot actively compensate for the hull motion well, resulting in poor stability of the stable platform. In this case, if the hull motion is predicted in advance according to the response lag time of the system, and then active compensation control is performed in time, the stability of the stable platform will be greatly improved. Since the response lag time of the system is short, the prediction time is also short, generally 1s-2s, and the prediction value is updated in real time, so this form of prediction is called the extremely short-term online prediction of the hull motion posture. At present, the time series analysis method is commonly used for the extremely short-term online prediction of the hull motion posture, but due to the large amount of calculated data and the need to invert the data matrix during the calculation process, the calculation process is time-consuming and the real-time performance is relatively poor. Therefore, in order to improve the real-time performance of the prediction while ensuring the accuracy of the prediction, it is very urgent and necessary to seek an extremely short-term real-time prediction method for the hull motion posture. Summary of the invention

[0004] In view of the defects in the above-mentioned prior art, the present invention proposes a very short-term real-time prediction method for the hull motion posture. The method includes collecting hull motion posture data with a large sampling period under the condition of satisfying Shannon sampling theorem, establishing an AR model, using the Bayesian information BIC criterion to determine the order of the AR model, using a limited memory recursive least squares algorithm to perform online updates of AR model parameters, using a rolling forecast method to perform multi-step prediction of the hull motion posture, combining the predicted data with historical data of a certain length, establishing an RBF neural network model for fitting interpolation, and obtaining very short-term prediction data of the hull motion posture. The present invention uses a limited memory recursive least squares algorithm for online updating, establishes an RBF neural network model for fitting interpolation, can effectively improve the real-time performance of the prediction algorithm, and has a high prediction accuracy.

[0005] The present invention provides a method for extremely short-term real-time prediction of a ship motion posture, which comprises the following steps:

[0006] S1. Under the condition of satisfying Shannon sampling theorem, the hull motion posture data is collected with a larger sampling period;

[0007] S2. Use the collected time series data of the ship motion posture to establish an AR model:

[0008] x t =φ1x t-1 +φ2x t-2 +…+φ p x t-p +ε t (1)

[0009] Among them, x t represents the pose value at the current moment; p represents the order of the AR model, that is, the number of independent variables; ε t represents error; φ represents model coefficient;

[0010] S3. Use the Bayesian information BIC criterion to determine the order p of the AR model:

[0011] BIC(p)=lge 2 +(p+1)(lgN) / N (3)

[0012] Where N represents the number of samples; e 2 represents the mean square error and Represents the predicted value of posture;

[0013] S4, using the limited memory recursive least squares algorithm to perform online update of the AR model parameter Φ;

[0014] S41. When there are n data in the data window, the AR model order p is determined based on the Bayesian information BIC criterion, and the sample input matrix X constituting the training set is:

[0015] X=[Θ1 Θ2 … Θ n-p ] T (4)

[0016] Among them, Θ n-p represents a single input sample and Θ n-p =[x n-p x n-p+1 … x n-1 ] T , the corresponding sample output Y is:

[0017] Y=[x p+1 x p+2 … x n ]T (5);

[0018] S42, fix the window length of the data used to estimate the parameters, and obtain a new sample input matrix each time a new set of data is collected for:

[0019]

[0020] Corresponding to the new sample output matrix for:

[0021]

[0022] S43, using the limited memory recursive least squares algorithm to perform online update of the AR model parameter Φ:

[0023]

[0024] in, represents the estimated value of the AR model parameter Φ obtained using the restricted memory method after collecting new data; Represents the gain coefficient; P n+1 represents the covariance matrix; is the estimated value of the AR model parameter Φ obtained using the recursive least squares algorithm after collecting new data;

[0025] S5. Use rolling forecast to make multi-step forecast of the ship motion position: forecast the data at the next moment based on the historical data of the ship motion position, and then forecast the data at the next moment based on the historical data until the forecast duration is reached;

[0026] S6. Combine the forecasted data with the historical data of a certain length, establish an RBF neural network model for fitting and interpolation, and obtain the extremely short-term forecast data of the ship's motion posture.

[0027] Further, the step S1 specifically includes the following steps:

[0028] S11. Based on Shannon sampling theorem, select several sampling periods between the maximum sampling period and the system control period, and select the one with the best prediction performance as the final sampling period;

[0029] S12, collect data on the motion posture of the 6 degrees of freedom of the hull respectively, and record the collected time series data as {x i ,i=1~n}, where i represents time.

[0030] Further, the AR model parameter Φ of the AR model in step S2 is:

[0031] Φ=[φ1 φ2 … φ p ]T (2).

[0032] Preferably, the step S5 specifically includes the following steps:

[0033] S51, predicting the hull motion posture data at the next moment according to the historical data of the hull motion posture;

[0034] S52, combining the predicted ship motion posture with historical data to predict the ship motion posture at the next moment, until the requirement for the forecast duration is met;

[0035] S53, obtaining the predicted value of the ship's motion posture at the lth moment after the current moment for:

[0036]

[0037] Among them, l is the number of steps that need to be predicted.

[0038] Preferably, the radial basis function of the hidden layer of the RBF neural network model adopts a Gaussian kernel function:

[0039]

[0040] Among them, ||XC i || represents the vector norm, that is, X and C i The distance between two vectors; C i represents the center of the i-th Gaussian kernel function, which is uniformly selected from the input samples; σ i Represents the width coefficient of the i-th Gaussian kernel function;

[0041] The output layer of the RBF neural network model is a weighted sum of the output values ​​of the hidden layer:

[0042] f(X)=ΨW+b (11)

[0043] Wherein, Ψ represents the output value of the hidden layer of the RBF neural network model; W represents the weighted value; b represents the bias term; and the batch least squares algorithm is used to perform online estimation of the output layer parameters of the RBF neural network model.

[0044] Preferably, the limited memory is that each time a new set of data is collected, a set of old data is discarded to ensure that the current model parameters are estimated based on the latest section of hull motion posture data.

[0045] Preferably, the interval between adjacent moments predicted in step S5 is equal to the sampling period, the period of the extremely short-term hull motion posture prediction data in step S6 is consistent with the system control period, and the sampling period is greater than the system control period.

[0046] Preferably, the step S3 uses a grid search algorithm to perform optimization with BIC(p) as the target value.

[0047] Compared with the prior art, the technical effects of the present invention are:

[0048] 1. A very short-term real-time prediction method for hull motion posture designed by the present invention first collects data on hull motion with a larger sampling period to reduce the amount of calculation; then an AR model is established to predict the hull motion online, the model order is determined according to the Bayesian information BIC criterion, and the model parameters are updated online using the recursive least squares algorithm with limited memory to shorten the calculation time, and multi-step prediction is performed by rolling prediction to meet the requirements of the forecast time.

[0049] 2. The present invention designs an extremely short-term real-time prediction method for hull motion posture. In order to make the prediction data and the system control data have the same sampling period, a RBF neural network model is established to fit and interpolate the data set consisting of the prediction data and part of the historical data. Time is used as the sample input and the hull motion posture is used as the sample output. A Gaussian function is used as the radial basis kernel function, and the output layer parameters are updated online using a batch processing algorithm. The real-time performance of the prediction algorithm can be effectively improved, and the prediction accuracy is relatively high. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Other features, objects and advantages of the present application will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings.

[0051] Figure 1 It is a flow chart of the extremely short-term real-time prediction method of the hull motion posture of the present invention;

[0052] Figure 2 It is a schematic diagram of the RBF neural network structure of the present invention;

[0053] Figure 3 It is a schematic diagram of different time periods of the hull motion curve of the present invention. DETAILED DESCRIPTION

[0054] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It should also be noted that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.

[0055] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0056] Taking the rise and fall of the ship as an example, the very short-term forecast duration is 1s, and the period of system control data is relatively small, generally 0.01s. If this period is used as the sampling period for data collection and calculation and prediction of the ship's motion posture, the amount of calculation is large. Therefore, the present invention adopts the method of increasing the sampling period to reduce the amount of calculation and improve real-time performance.

[0057] Figure 1 The present invention shows a very short-term real-time prediction method for the hull motion posture, which comprises the following steps:

[0058] S1. Under the condition of satisfying Shannon sampling theorem, the hull motion posture data is collected with a larger sampling period.

[0059] S11. According to Shannon's sampling theorem, in order to restore the analog signal without distortion, the sampling frequency should be no less than twice the highest frequency in the analog signal spectrum. Select several sampling periods between the maximum sampling period and the system control period, and select the one with the best forecast performance as the final sampling period. In a specific embodiment, in order to determine the best sampling period, four different sampling periods of 0.01s, 0.05s, 0.1s, and 0.5s are selected for forecast analysis, and a sliding data window of 20s with the same time length is taken, and a forecast analysis is performed on 200s of data. The same duration is to ensure that the sample data used for modeling contains the same amount of information, and to ensure that the prerequisites for accuracy comparison are consistent.

[0060] S12, collect data on the motion posture of the 6 degrees of freedom of the hull respectively, and record the collected time series data as {x i ,i=1~n}, where i represents time.

[0061] S2. Use the collected time series data of the ship motion posture to establish an AR model:

[0062] x t =φ1x t-1 +φ2x t-2 +…+φ p x t-p +ε t (1)

[0063] Among them, x t represents the pose value at the current moment; p represents the order of the AR model, that is, the number of independent variables; ε t represents the error; φ represents the model coefficient.

[0064] The AR model parameter Φ of the AR model is:

[0065] Φ=[φ1 φ2 … φ p ] T (2).

[0066] S3. Use the Bayesian information BIC criterion to determine the order p of the AR model:

[0067] BIC(p)=lge 2 +(p+1)(lgN) / N (3)

[0068] Where N represents the number of samples; e 2 represents the mean square error and Represents the predicted value of posture. The grid search algorithm is used to optimize with BIC(p) as the target value.

[0069] S4. Use the recursive least squares algorithm with limited memory to update the AR model parameters Φ online; limited memory means that every time a new set of data is collected, a set of old data is discarded to ensure that the current model parameters are estimated based on the latest section of hull motion posture data.

[0070] S41. When there are n data in the data window, the AR model order p is determined based on the Bayesian information BIC criterion, and the sample input matrix X constituting the training set is:

[0071] X=[Θ1 Θ2 … Θ n-p ] T (4)

[0072] Among them, Θ n-p represents a single input sample and Θ n-p =[x n-p x n-p+1 … x n-1 ] T , the corresponding sample output Y is:

[0073] Y=[x p+1 x p+2 … x n ] T (5).

[0074] The estimated values ​​of the model parameters obtained by the batch algorithm of the least squares method are:

[0075]

[0076] However, since the inverse operation of the matrix is ​​time-consuming, if batch processing is required every time the parameters are updated, the real-time performance of the prediction algorithm will be poor. Therefore, the recursive least squares algorithm can be used for parameter estimation. n+1 And when it is composed into a new sample, the sample input matrix becomes The sample output matrix becomes Y n+1 =[Y x n+1 ] T, then the expression of the recursive algorithm can be derived as:

[0077]

[0078] in, It represents the estimated value of the AR model parameter Φ obtained by using the recursive least squares algorithm after collecting new data. However, this recursive algorithm only considers the impact of new data on the model parameters and does not process the old data. As the data continues to increase, the amount of information obtained from the new data will decrease relatively, causing the algorithm to gradually lose its correction ability, which is the data saturation phenomenon.

[0079] In order to overcome this defect, the recursive least squares algorithm with limited memory is used to update the parameters online, as follows:

[0080] S42, fix the window length of the data used to estimate the parameters, and obtain a new sample input matrix each time a new set of data is collected for:

[0081]

[0082] Corresponding to the new sample output matrix for:

[0083]

[0084] S43, using the limited memory recursive least squares algorithm to perform online update of the AR model parameter Φ:

[0085]

[0086] in, represents the estimated value of the AR model parameter Φ obtained using the restricted memory method after collecting new data; Represents the gain coefficient; P n+1 represents the covariance matrix;

[0087] S5. Use rolling forecast method to make multi-step forecast of ship motion position: forecast the data at the next moment based on the historical data of ship motion position, and then forecast the data at the next moment based on the historical data, until the forecast duration.

[0088] S51. Predict the hull motion posture data at the next moment based on the historical data of the hull motion posture.

[0089] S52, combining the predicted ship motion posture with historical data to predict the ship motion posture at the next moment, until the requirement for the forecast duration is met, and the interval between adjacent forecast moments is equal to the sampling period.

[0090] S53, obtaining the predicted value of the ship's motion posture at the lth moment after the current moment for:

[0091]

[0092] Among them, l is the number of steps that need to be predicted.

[0093] In a specific embodiment, a batch algorithm and a bounded memory algorithm are used for forecast analysis respectively. The optimization results and forecast results of the model order of 1s hull heave displacement under different sampling periods of the batch algorithm are shown in Table 1, and the optimization results and forecast results of the model order of 1s hull heave displacement under different sampling periods of the bounded memory algorithm are shown in Table 2. When the sampling period is 0.01s, due to the long calculation time, the model order is not optimized and is set according to debugging experience. PE describes the proportion of the mean absolute error MAE in the absolute mean of the true value, indicating the accuracy of the forecast:

[0094]

[0095]

[0096] Table 1

[0097]

[0098]

[0099] Table 2

[0100] Comparing the data in Table 1 and Table 2, it can be seen that although the prediction accuracy of the bounded memory algorithm is lower than that of the batch processing algorithm, except for the sampling period of 0.01s, the error is less than 1%, and the accuracy is very high. Moreover, the calculation speed of the bounded memory method is significantly faster than that of the batch processing algorithm. At the same time, as the sampling period increases, the calculation time is shorter and the real-time performance is higher.

[0101] S6. By increasing the sampling period for data collection, the period of the forecast data obtained is inconsistent with the control period of the system. To solve this problem, the forecast data is combined with the historical data of a certain length, and an RBF neural network model is established for fitting interpolation. The structure is as follows: Figure 2 As shown, the extremely short-term prediction data of the hull motion posture is obtained, and the period of the extremely short-term prediction data of the hull motion posture is consistent with the system control period.

[0102] The hidden layer radial basis function of the RBF neural network model uses a Gaussian kernel function:

[0103]

[0104] Among them, ||XC i || represents the vector norm, that is, X and C i The distance between two vectors; C i represents the center of the i-th Gaussian kernel function, which is uniformly selected from the input samples; σ i Represents the width coefficient of the i-th Gaussian kernel function, which can be set empirically or optimized through an optimization algorithm.

[0105] The output layer of the RBF neural network model is a weighted sum of the output values ​​of the hidden layer:

[0106] f(X)=ΨW+b (11)

[0107] Where Ψ represents the output value of the hidden layer of the RBF neural network model and Ψ=[ψ1 ψ2 … ψ i ]; W represents the weighted value and W=[w1 w2 … w i ] T ; b represents the bias term. The batch least squares algorithm is used to estimate the output layer parameters of the RBF neural network model online.

[0108] When the forecast duration is 1s, the number of data predicted by the sampling periods of 0.05s, 0.1s and 0.5s are 20, 10 and 2 respectively. In order to increase the fitting accuracy, the data values ​​predicted at this moment are combined with the historical data of a certain length to increase the number of fitting data. Through debugging, 1s historical data is selected for combined interpolation. The shape of the hull motion curve can be divided into two periods: the approximate linear period and the curved period, such as Figure 3 As shown, the RBF neural network model is used to perform fitting interpolation and convert the time into the time used to process 200s data. The results of the linear period are shown in Table 3, and the results of the bending period are shown in Table 4.

[0109]

[0110] Table 3

[0111]

[0112] Table 4

[0113] By comparing the two tables, it can be found that the fitting interpolation error in the curved period is slightly larger than that in the approximate linear period, but overall, the errors in both periods are very small; at the same time, the calculation time decreases with the increase of the sampling period. The time spent on the prediction and fitting interpolation of 200s data at 0.05s, 0.1s and 0.5s sampling periods is counted together, as shown in Table 5.

[0114]

[0115] Table 5

[0116] When the sampling period is 0.01s, which is the period of the system control data, the time consumed for forecasting is 54.568s; when the method of increasing the sampling period for forecasting and then performing fitting and interpolation on the forecast data is adopted, it is obvious that the time consumed is significantly shortened, and at the same time, the accuracy of the forecast and fitting and interpolation is also maintained at a high level, proving the effectiveness of the invention.

[0117] The present invention designs a very short-term real-time prediction method for hull motion posture. Firstly, the hull motion data is collected with a relatively large sampling period to reduce the amount of calculation. Then, an AR model is established to predict the hull motion online. The model order is determined according to the Bayesian information BIC criterion. The model parameters are updated online by a recursive least squares algorithm with limited memory to shorten the calculation time. Multi-step prediction is performed by a rolling prediction method to meet the requirement of the prediction time. In order to make the prediction data and the system control data have the same sampling period, a RBF neural network model is established to fit and interpolate the data set composed of the prediction data and part of the historical data. Time is used as the sample input and the hull motion posture is used as the sample output. A Gaussian function is used as the radial basis kernel function. The output layer parameters are updated online by a batch processing algorithm. The real-time performance of the prediction algorithm can be effectively improved and the prediction accuracy is high.

[0118] Finally, it should be noted that the above embodiments are only intended to illustrate rather than limit the technical solutions of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the present invention can still be modified or replaced by equivalents. Any modification or partial replacement that does not depart from the spirit and scope of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for extremely short-term real-time prediction of ship motion posture, characterized in that: It includes the following steps: S1. Under the condition of satisfying Shannon sampling theorem, the hull motion posture data is collected with a larger sampling period; S2. Use the collected time series data of the ship motion posture to establish an AR model: x t =φ1x t-1 +φ2x t-2 +…+φ p x t-p +e t (1) Among them, x t represents the pose value at the current moment; p represents the order of the AR model, that is, the number of independent variables; ε t represents error; φ represents model coefficient; S3. Use the Bayesian information BIC criterion to determine the order p of the AR model: BIC(p)=lge 2 +(p+1)(lgN) / N (3) Where N represents the number of samples; e 2 represents the mean square error and Represents the predicted value of posture; S4, using the limited memory recursive least squares algorithm to perform online update of the AR model parameter Φ; S41. When there are n data in the data window, the AR model order p is determined based on the Bayesian information BIC criterion, and the sample input matrix X constituting the training set is: X=[Θ1 Θ2…Θ n-p ] T (4) Among them, Θ n-p represents a single input sample and Θ n-p =[x n-p x n-p+1 …x n-1 ] T , the corresponding sample output Y is: Y=[x p+1 x p+2 …xn]Tx n ] T (5); S42, fix the window length of the data used to estimate the parameters, and obtain a new sample input matrix each time a new set of data is collected for: Corresponding to the new sample output matrix for: S43, using the limited memory recursive least squares algorithm to perform online update of the AR model parameter Φ: in, represents the estimated value of the AR model parameter Φ obtained using the restricted memory method after collecting new data; Represents the gain coefficient; P n+1 represents the covariance matrix; is the estimated value of the AR model parameter Φ obtained using the recursive least squares algorithm after collecting new data; S5. Use rolling forecast to make multi-step forecast of the ship motion position: forecast the data at the next moment based on the historical data of the ship motion position, and then forecast the data at the next moment based on the historical data until the forecast duration is reached; S6, combining the forecasted data with the historical data of a certain length, establishing an RBF neural network model for fitting and interpolation, and obtaining extremely short-term forecast data of the ship's motion posture; The radial basis function of the hidden layer of the RBF neural network model adopts a Gaussian kernel function: Among them, ||XC i || represents the vector norm, that is, X and C i The distance between two vectors; C i represents the center of the i-th Gaussian kernel function, which is uniformly selected from the input samples; σ i Represents the width coefficient of the i-th Gaussian kernel function; The output layer of the RBF neural network model is a weighted sum of the output values ​​of the hidden layer: f(X)=ΨW+b (11) Wherein, Ψ represents the output value of the hidden layer of the RBF neural network model; W represents the weighted value; b represents the bias term; and the batch least squares algorithm is used to perform online estimation of the output layer parameters of the RBF neural network model.

2. The method for extremely short-term real-time prediction of ship motion posture according to claim 1 is characterized in that: The step S1 specifically includes the following steps: S11. Based on Shannon sampling theorem, select several sampling periods between the maximum sampling period and the system control period, and select the one with the best prediction performance as the final sampling period; S12, collect data on the motion posture of the 6 degrees of freedom of the hull respectively, and record the collected time series data as {x i ,i=1~n}, where i represents time.

3. The method for extremely short-term real-time prediction of ship motion posture according to claim 1 is characterized in that: The AR model parameter Φ of the AR model in step S2 is: Φ=[φ1 φ2…φ p ] T (2)。 4. The method for extremely short-term real-time prediction of ship motion posture according to claim 1, characterized in that: The step S5 specifically comprises the following steps: S51, predicting the hull motion posture data at the next moment according to the historical data of the hull motion posture; S52, combining the predicted ship motion posture with historical data to predict the ship motion posture at the next moment, until the requirement for the forecast duration is met; S53, obtaining the predicted value of the ship's motion posture at the lth moment after the current moment for: Among them, l is the number of steps that need to be predicted.

5. The method for extremely short-term real-time prediction of ship motion posture according to claim 1, characterized in that: The limited memory is that each time a new set of data is collected, a set of old data is discarded to ensure that the current model parameters are estimated based on the latest section of hull motion posture data.

6. The method for extremely short-term real-time prediction of ship motion posture according to claim 1, characterized in that: The interval between adjacent moments predicted in step S5 is equal to the sampling period, the period of the extremely short-term ship motion posture prediction data in step S6 is consistent with the system control period, and the sampling period is greater than the system control period.

7. The method for extremely short-term real-time prediction of ship motion posture according to claim 1 is characterized in that: The step S3 uses a grid search algorithm to optimize with BIC(p) as the target value.

Images