A data and knowledge hybrid-driven ship speed prediction method

Abstract

The application belongs to the technical field of artificial intelligence, and in particular relates to a data and knowledge hybrid driving ship speed prediction method. The scheme takes industry knowledge data as one of data sets, adopts a 'data+knowledge' hybrid driving training mode to predict the ship speed, effectively deals with the small sample problem in the shipping business, significantly improves the accuracy of the speed prediction, and reduces the construction cost.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to a data and knowledge hybrid driven ship speed prediction method. BACKGROUND

[0002] An automatic identification system (AIS) transmits detailed information of a ship through radio, including a ship name, a position, a heading, a speed and the like, and provides a solid foundation for ship monitoring, collision avoidance, route optimization and behavior analysis. In combination with AIS data and a deep learning model, a ship speed can be accurately predicted, a space-time position of the ship can be predicted in advance, potential risks can be effectively avoided, sailing safety is improved, and an intelligent process of shipping is promoted.

[0003] In current business, ship speed prediction mainly relies on an experience statistical method, a statistical model is constructed for prediction based on analysis of a feature law of historical data. However, this method ignores complex interactions of environmental factors, lacks real-time performance and dynamic adjustment capability, and has a limited application range. With rapid development of artificial intelligence technology, a machine learning model has shown excellent performance in many fields due to its high efficiency and accuracy. However, a pure data driven network model often relies on a large amount of labeled data for training to ensure its accuracy and good generalization capability. In actual application of shipping business, it is difficult to obtain a sufficient number of high-quality data, and the cost is high. SUMMARY

[0004] The technical problem to be solved by the application is to provide a data and knowledge hybrid driven ship speed prediction method, which has a small sample quantity and low cost.

[0005] To solve the above problems, the following technical solutions are provided.

[0006] The data and knowledge hybrid driven ship speed prediction method has the following steps.

[0007] In a first step, original data is collected, and the original data is preprocessed to obtain standardized data. The original data includes meteorological data, AIS data, ship static data and industry knowledge data.

[0008] In a second step, all data are divided into multiple training sets and test sets, the training sets and the test sets each include feature data for inputting a learning model and label data for verifying a prediction result of the learning model.

[0009] In a third step, the training sets and the test sets are input into the learning model to train the learning model until an optimal parameter is found, and the optimal parameter is saved.

[0010] Fourthly, the new test set containing historical data and industry knowledge data is input into the prediction model to realize the prediction of the ship speed.

[0011] The industry knowledge data is obtained by an experience formula, and the experience formula includes a Russian Central Marine Research Institute formula, a Qingdao Meteorological Navigation Research Consortium formula, a National Meteorological Bureau National Meteorological Center formula and / or a semi-experience formula.

[0012]

[0013] wherein, is the actual speed of the ship in the wave; is the speed of the ship in still water; is the loss of speed direction factor changing with the wind wave direction and wind level; is the loss of speed correction coefficient changing with the block coefficient, loading condition and Froude number; is the ship type factor related to the ship type, wind level and ship displacement in units of tons. The first step is to preprocess the original data as follows:

[0014] S1: Eliminate abnormal data, and the specific rules are:

[0015] (1) If all the meteorological data in a certain data are 0, the data is eliminated;

[0016] (2) If the speed is lower than 4 knots or higher than 20 knots, the data is eliminated.

[0017] (3) If the wind level is less than 4, and the ship speed in the normal navigation process is between 4-5 knots, the data is eliminated.

[0018] (4) If the ship charter speed is 0, the data is eliminated. (5) If the wind level is equal to 0, the data with the difference between the ship speed and the charter speed greater than 8 is eliminated.

[0019]

[0020] S2: Encoding processing

[0021]

[0022] The 0-1 encoding is used to convert the ship loading state:

[0023]

[0024] ​​​​​​​wherein, empty, full, i.e., if the ship is empty, the code is 0; if the ship is full, the code is 1;

[0025] S3: normalization processing

[0026] The data is standardized by using the normalization method to convert the data to the data in the interval [0, 1] .

[0027] The normalization is a Min-Max normalization method, which converts the data to the interval [0, 1] by the following formula:

[0028]

[0029] wherein, is the original data; , The maximum and minimum values in the original data are respectively.

[0030] The training set and the test set each contain 40 feature data and 1 label data; the 40 feature data contained in the training set and the test set are respectively 20 feature data from meteorological data, 5 feature data from AIS data, 11 feature data from ship static data, 4 feature data from industry knowledge data, and 1 label data from AIS data; the 1 label data contained in the training set and the test set is the label data from the AIS data;

[0031] Among the meteorological data, the 20 feature data from the meteorological data are further subdivided into current meteorological data and meteorological data at a previous time in the same voyage by 10 meteorological data;

[0032] In the AIS data, the current speed is set as the label data, and the speed at a previous time in the same voyage is taken as 1 feature data from the AIS; at the same time, the ship's heading is also subdivided into the current heading and the heading at a previous time in the same voyage as 2 feature data from the AIS; in addition, the current ship's latitude and longitude position data are also taken as 2 feature data from the AIS.

[0033] The sample set is split by random division to ensure that there is no repeated sample in the training set and the test set, 90% of the samples in the preprocessed sample set are randomly selected as the data of the training set, and the remaining 10% of the samples are taken as the data of the test set.

[0034] The learning model includes but is not limited to LightGBM, XGBoost, random forest and BP neural network.

[0035] The learning model continuously optimizes the model parameters by minimizing the loss function until the optimal parameters are found.

[0036] The minimized loss function is:

[0037]

[0038] Wherein: is the weight of the data-based loss function, is the weight of the knowledge-based loss function; and can be various error measurement methods, including but not limited to root mean square error (RMSE), mean square error (MSE), and mean absolute error (MAE).

[0039] With the above solution, the following advantages are achieved:

[0040] 1. The present application combines meteorological data, AIS data, ship static data and industry knowledge data, and adopts a "data + knowledge" hybrid driving training mode to predict ship speed. This method effectively addresses the small sample problem in shipping business by combining data and knowledge in the industry, significantly improving the accuracy of speed prediction and reducing the construction cost.

[0041] 2. In recent years, ship speed prediction methods based on artificial intelligence have considered ship static water speed and environmental factors affecting the ship as model inputs, and the mapping relationship can be understood as "static water speed at point A + meteorological information at point A = actual speed at point A". Ship static water speed requires dynamic variables that reflect the ship's set cruising speed, such as engine speed and propeller shaft speed. Engine speed and propeller shaft speed are dynamic information during ship travel, and these parameters change with the running state of the ship. In terms of real shipping business, the model is often used to predict speed at a certain location point, and the output is the speed for multiple consecutive days in the future. It is impossible to obtain future dynamic information of the ship during the model prediction process, so it is difficult and unsuitable to use dynamic information in artificial intelligence training schemes based on real business needs of shipping companies. Based on the above problems, the present application ingeniously combines AIS real-time data information during the construction of the data set, specifically using AIS information at a certain time point (the "certain" represents a non-fixed time point) as a feature at a later time point, and the mapping relationship is "AIS information at a certain time point + meteorological information corresponding to this point + meteorological information at point A = actual speed at point A". This construction scheme no longer requires static water speed and can be applied to real GMDSS business, effectively extending the prediction time and ensuring accurate long-term prediction. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a flowchart of the data and knowledge hybrid driven ship speed prediction method of the present application;

[0043] Figure 2 is a learning flowchart of the learning model in the data and knowledge hybrid driven ship speed prediction method of the present application;

[0044] Figure 3 is a root mean square error (RMSE) comparison chart of the test set in the specific implementation case;

[0045] Figure 4 is a schematic diagram of Table 4 in the specific implementation case;

[0046] Figure 5 is a schematic diagram of Table 5 in the specific implementation case is a meaning table of the factor;

[0047] Figure 6 is a schematic diagram of Table 5 in the specific implementation case;

[0048] Figure 7 is a schematic diagram of Table 6 in the specific implementation case;

[0049] Figure 8 is a schematic diagram of Table 7 in the specific implementation case;

[0050] Figure 9 is a schematic diagram of the model and error measurement method selection table of the experiment in the specific implementation case. DETAILED DESCRIPTION

[0051] The present application will be further described in detail below with reference to the accompanying drawings.

[0052] As shown in Figure 1 , the data and knowledge hybrid driven ship speed prediction method of the present application includes the following steps:

[0053] First, collect raw data and pre-process the raw data to obtain standardized data .

[0054] The raw data collected in this embodiment includes meteorological data, AIS data, ship static data, and industry knowledge data, which are as follows:

[0055] 1. Meteorological data

[0056] Table 1 Meteorological feature data and data units

[0057] Weather feature data Data unit Wind direction degrees Wind speed knots Swell direction degrees Swell height meters Swell period seconds Sea direction degrees Sea height meters Sea period seconds Current direction degrees Current speed knots

[0058] 2. AIS data

[0059] Table 2 AIS data and data units

[0060] AIS data Data unit Speed knots Heading degrees Ship position - longitude degrees Ship position - latitude degrees

[0061] 3. Ship static data

[0062] Table 3. Static data of ships and data units

[0063] Ship static data Data unit Cargo status / Deadweight tons tons Ship length meters Ship width meters Ship height meters Ship displacement tons Ship waterline length meters Ship draft meters Ship gross tonnage / Ship net tonnage / Charter speed knots

[0064] 4. Industry knowledge data

[0065] Industry knowledge data is obtained by constructing empirical formulas based on statistical analysis of a large amount of measured and experimental data.

[0066] The empirical formulas that can be used include, but are not limited to, the following methods:

[0067] (1) Russian Central Maritime Research Institute

[0068] (1)

[0069] in: This represents the ship's actual speed in the waves; The speed of the ship in still water; For the high waves; The angle between the bow and the wave direction; This refers to the ship's actual displacement (in tons).

[0070] (2) Qingdao Meteorological Navigation Research Consortium

[0071] (2)

[0072] in: This represents the ship's actual speed in the waves; For significant wave height; The angle between the bow and the direction of the waves. This refers to the ship's actual displacement. The speed of the ship in still water; These are the various performance coefficients of the ship, where: =0.745, =0.05015, =0.0045, =1.35×10 -6 G is an empirical coefficient, and the specific values ​​of G are shown in Table 4, i.e., as follows: Figure 4 As shown:

[0073] (3) National Meteorological Center, China Meteorological Administration

[0074] ① Applicable to cargo ships of 20,000 to 50,000 tons:

[0075]

[0076]

[0077] (3)

[0078] ② Applicable to cargo ships of 50,000 to 100,000 tons:

[0079]

[0080]

[0081] (4)

[0082] in: Let wind direction (WD), wind speed (WS), wave height (SH), swell direction (SWD), swell height (SWH), displacement (TON), course (HDG), and speed (SPD) be the stall factor for ship navigation. Then, the above equations... The meaning of factor is as follows: Figure 5 As shown;

[0083] The final formula for speed is:

[0084] (5)

[0085] in: This represents the ship's actual speed in the waves; The speed of the ship in still water; The stall factor for ship navigation.

[0086] (4) Semi-empirical formula (Kwon, YJ estimation method)

[0087] (6)

[0088] in: This represents the ship's actual speed in the waves; The speed of the ship in still water; To follow the direction of wind and waves and wind force ( The varying stall direction factor is detailed in Table 5. Figure 6 As shown; The factors include the block coefficient, loading status, and Froude number. The stall correction factor varies accordingly; see Table 6 for details. Figure 7 As shown; To match the ship type and wind class ( ) and with The ship's displacement (D) is given in units of tons The ship type factor (K) is related to the ship's displacement (D) and is given in Table 7, as shown below. Figure 8

[0089] Where: Refers to the Beaufort wind scale, which is divided into 0~12 according to the strength of the wind, a total of 13 levels.

[0090] Where: The calculation formula of the Froude number is:

[0091] (7)

[0092] Where, The speed of the ship in still water; Is a fixed value, ; Is the length of the ship's waterline.

[0093] In this embodiment, the process of data preprocessing is as follows:

[0094] 1. Outlier processing

[0095] In the ship data set, the reasons for the occurrence of outliers include but are not limited to sensor failure, data transmission failure and human recording error, which pose a potential threat to the performance of the model. In view of the sufficient amount of collected data, the present application takes a data scientific and reasonable approach to deal with the problem of outliers. Specifically, the following rules are set to identify and exclude abnormal data:

[0096] If all the weather data in a certain data shows 0, the data is excluded;

[0097] Specifically, if all the weather data in a certain data shows 0, it usually means that the sensor may have failed, resulting in the inability to correctly collect and record weather information. Therefore, such data is considered as abnormal data and is deleted.

[0098] (2) If the speed is less than 4 knots or more than 20 knots, the data is excluded.

[0099] Specifically, generally speaking, when the speed is 3 knots, 2 knots or less, it is usually a human speed reduction operation when the ship is about to stop, so if the speed is less than 4 knots or more than 20 knots, the present application considers that it is beyond the normal speed range, therefore these data are also considered as abnormal and are deleted.

[0100] (3) If the wind scale is less than 4, the ship speed during normal navigation falls between 4~5 knots, the data is excluded. ​

[0101] Specifically, Beaufort wind scale Less than 4, indicating that the current is in a windless or light wind state. Under such weak wind force, if the ship speed during normal navigation falls between 4-5 knots, the present invention considers this to be an unreasonable phenomenon. Therefore, such data is also regarded as abnormal and deleted.

[0102] (4) If the ship charter speed is 0, the data is deleted.

[0103] Specifically, the ship charter speed refers to the sailing speed agreed upon in the lease contract. If the ship charter speed in a certain data is 0, it usually means that the data record is abnormal, so the present invention determines such data as abnormal and deletes it.

[0104] (5) If the wind scale is equal to 0, the data with a difference between the ship speed and the charter speed greater than 8 is deleted.

[0105] Specifically, when is equal to 0 (i.e., indicating that the current is in a windless state), the present invention considers the data with a difference between the ship speed and the charter speed greater than 8 as abnormal data and deletes it.

[0106] 2. Encoding processing

[0107] 0-1 encoding is used to convert the ship loading state:

[0108]

[0109] Among them, empty, full, i.e., if the ship is empty, the encoding is 0; if the ship is full, the encoding is 1, i.e., if the ship is empty, the encoding is 0; if the ship is full, the encoding is 1. In this way, the input data is properly formatted into a numerical type to meet the requirements of model training.

[0110] 3. Normalization processing

[0111] In order to eliminate the differences in dimensions and numerical ranges between features, the present invention uses the Min-Max normalization method, which converts data to the [0, 1] interval by the following formula:

[0112] (9)

[0113] Among them: is the standardized data; is the original data; , are the maximum and minimum values in the original data, respectively.

[0114] In summary, this process not only accelerates the process of the gradient descent algorithm to find the optimal solution, but also effectively improves the prediction accuracy of the model.

[0115] Secondly, all data are divided into multiple training sets and test sets, and both the training set and the test set contain feature data for inputting the learning model and label data for verifying the prediction result of the learning model.

[0116] Both the training set and the test set contain 40 feature data and 1 label data. The 40 feature data contained in the training set and the test set are respectively 20 feature data from the meteorological data, 5 feature data from the AIS data, 11 feature data from the static data of the ship, 4 feature data from the industry knowledge data and 1 label data from the AIS data. The 1 label data contained in the training set and the test set is the label data from the AIS data;

[0117] In the meteorological data, the 20 feature data from the meteorological data are further subdivided into the meteorological data at the current time and the meteorological data at a certain time in the same voyage.

[0118] Specifically, the 10 meteorological data are further subdivided into the meteorological data at the current time and the meteorological data at a certain time (for example, time A) in the same voyage, so that a total of 20 meteorological-related data features are used at the input end of the model.

[0119] In the AIS data, the current speed is set as the label data, and the speed at a certain time in the same voyage is used as a feature data from the AIS; at the same time, the ship's heading is also subdivided into the heading at the current time and the heading at a certain time in the same voyage as two feature data from the AIS; in addition, the current ship's latitude and longitude position data are also used as two feature data from the AIS.

[0120] Specifically, in the AIS data, the current time speed is set as the label data, and the speed at a previous time (e.g., time A) in the same voyage is taken as one of the feature data. Similarly, the heading of the ship is also divided into the current time heading and the heading at a previous time (e.g., time A) in the same voyage, both of which are taken as the feature data. In addition, the latitude and longitude position data of the current ship are also included in the category of the feature data. In this embodiment, the AIS information at the previous time (the "some" represents not fixed) is taken as the feature of the later time, and the mapping relationship is "AIS information at the previous time + weather information corresponding to this point + weather information at point A = actual speed at point A". This construction scheme no longer needs the still water speed, can be applied to the real air navigation business, effectively prolongs the prediction time, and can ensure accurate prediction for a long time.

[0121] Specifically, the training set is specially used for the training process of the model, and the test set is used for verifying and evaluating the performance of the model. In this scheme, the random division method is used to split the sample set to ensure that there is no repeated sample in the training set and the test set. 90% of the samples in the preprocessed sample set are randomly selected as the training set, and the remaining 10% of the samples are taken as the test set. At the same time, in order to ensure the effectiveness of the model, the samples in the training set and the test set need to maintain high correlation, so as to accurately reflect the performance of the model in the actual application scenario.

[0122] Thirdly, the training set and the test set are input into the learning model to train the learning model until the optimal parameters are found, and the optimal parameters are saved.

[0123] Specifically, as shown in Figure 2 , first, the training set and the test set are input into the learning model, and the training set and the test set jointly constitute the basis for training the machine learning model. The model includes but is not limited to LightGBM, XGBoost, random forest and BP neural network. In the training stage, the model parameters are continuously optimized by minimizing the loss function until the optimal parameters are found, and these optimal parameters are saved. The optimal parameters are verified by using the test set, so as to ensure the correctness of the optimal parameters.

[0124] Fourthly, the new test set containing historical data and industry knowledge data is input into the prediction model to realize the prediction of the ship speed.

[0125] Specific implementation case

[0126] For the small sample problem of shipping business, this scheme proposes an innovative solution.

[0127] Based on this scheme, the combinations in Table 8 are selected for experiments, as shown in Figure 9 .

[0128] The data used in the experiment of the application covers eight ship types including dry bulk carriers, open hatch cargo ships, general cargo ships and ore carriers.

[0129] As shown in Figure 3 , 100 voyage data were selected for testing and verification in the experiment. The test results show that compared with the root mean square error (RMSE) of 2.313 obtained by the empirical statistical method, the RMSE of the "data + knowledge" dual-drive training scheme proposed by the application is significantly reduced to 1.009, and the overall error is reduced by as much as 56.37%.

[0130] The innovative scheme of the application has been applied in actual shipping business, and has won good feedback due to its high accuracy.

[0131] Specifically, the planned route from the Port of Recalada to the Port of Singapore passes through the Cape of Good Hope region with poor weather conditions. On August 7, 2024, when passing through the Cape of Good Hope region, the speed error obtained by the scheme of the application is significantly lower than that of the empirical statistical method, indicating that the predicted value is closer to the true sailing speed. In particular, at 12 o'clock on August 7, the gap between the predicted value and the true value is reduced to only 0.06 knots.

[0132] In summary, after the shipping company adopts the model trained by the scheme of the application, the most safe route avoiding the poor weather can be accurately deduced and planned, which not only effectively reduces the economic loss, but also ensures the safety of the crew.

[0133] The scheme innovatively adopts a "data + knowledge" hybrid driving training mode, which combines meteorological data, AIS data, ship static data and industry knowledge data, significantly improves the accuracy of sailing speed prediction, effectively addresses the problem of data scarcity in shipping business, and promotes the intelligentization process of shipping.

Claims

1. A data- and knowledge-driven method for predicting ship speed, characterized in that, Includes the following steps: The first step is to collect raw data and preprocess it to obtain standardized data. The raw data includes meteorological data, AIS data, ship static data, and industry knowledge data. The second step is to collect all the data. It is divided into multiple training sets and test sets. Both the training sets and test sets contain feature data used to input the learning model and label data used to verify the prediction results of the learning model. The training set and the test set each contain 40 feature data points and 1 label data point. The 40 feature data points in both sets consist of 20 feature data points from meteorological data, 5 feature data points from AIS data, 11 feature data points from ship static data, 4 feature data points from industry knowledge data, and 1 label data point from AIS data. The 1 label data point in both sets is a label data point from AIS data. Among them, the meteorological data includes 20 feature data from the meteorological data, which are formed by further subdividing 10 meteorological data into meteorological data at the current moment and meteorological data at a previous moment in the same voyage; In AIS data, the current speed is set as the tag data, while the speed at a previous moment in the same voyage is used as one feature data from AIS; at the same time, the ship's heading is also subdivided into the heading at the current moment and the heading at a previous moment in the same voyage as two feature data from AIS; in addition, the current latitude and longitude position data of the ship is also used as two feature data from AIS. The third step is to input the training set and test set into the learning model to train the learning model until the optimal parameters are found and then save the optimal parameters. The fourth step is to input a new test set containing historical data and industry knowledge data into the prediction model to achieve the prediction of ship speed.

2. The data and knowledge-driven ship speed prediction method as described in claim 1, characterized in that, The industry knowledge data is obtained by calculation using empirical formulas; these empirical formulas include formulas from the Russian Central Maritime Research Institute, the Qingdao Meteorological Navigation Research Consortium, the National Meteorological Center of the China Meteorological Administration, and / or semi-empirical formulas; wherein, the semi-empirical formulas are: in, This represents the ship's actual speed in the waves; The speed of the ship in still water; To follow the direction of wind and waves and wind level The changing stall direction factor; The factors include the block coefficient, loading status, and Froude number. The stall correction factor changes accordingly; In order to match the ship type and wind class and Ship displacement in units Relevant ship type factors.

3. The data and knowledge-driven ship speed prediction method as described in claim 1, characterized in that, The first step involves preprocessing the raw data as follows: S1: Remove outlier data; the specific rules are as follows: (1) If all the meteorological data in a certain data point is displayed as 0, the data point shall be removed; (2) If the speed is below 4 knots or above 20 knots, the data will be discarded; (3) Wind level If the speed of a ship during normal navigation falls between 4 and 5 knots, the data will be discarded. (4) If the speed of the charter party is 0, the data should be discarded; (5) Wind level When the value is 0, data with a difference of more than 8 between the ship's speed and the charter speed will be removed. S2: Encoding Processing The ship's loading status is switched using 0-1 coding: in, Unloaded The code is 0 if the vessel is empty and 1 if the vessel is fully loaded. S3: Normalization The data is standardized using a normalization method, transforming it to the range [0, 1]. .

4. The data and knowledge-driven ship speed prediction method as described in claim 3, characterized in that, The normalization is a Min-Max normalization method, which transforms the data to the [0, 1] interval using the following formula: in, It is the raw data; , These are the maximum and minimum values ​​in the original data, respectively.

5. The data and knowledge-driven ship speed prediction method as described in claim 1, characterized in that, The sample set is split randomly to ensure that there are no duplicate samples in the training set and the test set. 90% of the samples are randomly selected from the preprocessed sample set as the training set data, and the remaining 10% of the samples are used as the test set data.

6. The data and knowledge-driven ship speed prediction method as described in claim 1, characterized in that, The learning models include LightGBM, XGBoost, Random Forest, and BP neural network.

7. The data and knowledge-driven ship speed prediction method as described in claim 1, characterized in that, The learning model continuously optimizes its parameters by minimizing the loss function until the optimal parameters are found.

8. The data and knowledge-driven ship speed prediction method as described in claim 7, characterized in that, The minimized loss function is: in: The weights are based on the loss function of the data. That is, the weights of the knowledge-based loss function; and It can be a variety of error measurement methods, including root mean square error (RMSE), mean square error (MSE), and mean absolute error (MAE).

Images