Ship classification method based on navigation big data

Abstract

The invention discloses a ship classification method based on navigation big data, and the method comprises the following steps: fusing AIS data and ARPA radar data, taking the ARPA radar data as a modeling feature, taking the ship type information of the AIS data as a label, and obtaining a fused data set; constructing a LightGBM ship multi-classification identification model, and setting a model objective function; dividing the fusion data set into a training set and a test set, and training the model; constructing a Focal Loss loss function, and integrating the Focal Loss loss function into the model; carrying out interpretive analysis on a model result; and obtaining trajectory data of a certain ship observed by a radar, and inputting the trajectory data into the trained model for ship type identification and prediction. According to the method, under the condition that AIS information is lacked, the ships can still be accurately classified, the marine traffic condition can be known more comprehensively, and the monitoring and coping capacity for abnormal behaviors is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of ship classification and identification, and in particular to a ship classification method based on navigation big data. Background Art

[0002] The Automatic Identification System (AIS) is a global ship tracking system for automatic identification and position reporting of ships. The AIS system installs transmitters on ships to transmit key information in real time, including ship name, call sign, country of registry, navigation status, speed, heading, etc. These data are received through satellites and ground base stations to form a global maritime traffic network.

[0003] With the continuous development of ship technology, there are many types of ships sailing on the sea. In practical applications, some ship data is often incomplete or missing. Among them, some ships may actively turn off the AIS (Automatic Identification System) or lose the signal, and can only obtain some basic data through radar, such as longitude and latitude, speed, heading, target length and other information, which makes it a challenging task to accurately classify and identify these ships.

[0004] In order to realize ship classification and identification, the existing technology mainly adopts the following technical means:

[0005] For example, the random forest ship target classification method based on AIS data feature optimization with patent publication number CN114882293A can efficiently utilize the rich information contained in AIS data, reduce the complexity of the model, and better distinguish different types of ships. The classification efficiency is better than the random forest based on the original features, and can meet the needs of accurate and efficient classification and identification of ship targets. This algorithm is relatively simple and easy to implement, but due to the simple model structure used, the accuracy is poor.

[0006] For example, the patent publication number CN113569921A is a GNN-based ship classification and identification method and device. This method uses ship trajectories to effectively extract spatial features for machine learning and improve the accuracy of ship trajectory classification and identification. However, this algorithm has many parameters, a long training time, and poor interpretability. Summary of the invention

[0007] In order to overcome the defects and shortcomings of the prior art, the present invention provides a ship classification method based on navigation big data. The present invention models the fused AIS and ARPA radar data, uses the category information in the AIS data as a label, fuses the AIS and ARPA radar data, and performs supervised learning training on the fused data (hereinafter referred to as fused data) to obtain a prediction model, thereby achieving accurate prediction of the ship type in the radar data, so that the ship can still be accurately classified in the absence of AIS information. By predicting the ship type, the maritime traffic situation can be more comprehensively understood, the monitoring and response capabilities for abnormal behaviors can be improved, and maritime safety management can be further strengthened.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] The present invention provides a ship classification method based on navigation big data, comprising the following steps:

[0010] The AIS data is fused with the ARPA radar data. The ARPA radar data is used as the modeling feature and the ship category information in the AIS data is used as the label to obtain the fused data set, which specifically includes:

[0011] Get the real trajectory data, aggregate different trajectory points by fusing the target ID, and then arrange them in the order of timestamps to form a complete record of each ship, so that the trajectory of the same ship is aggregated under the same ship;

[0012] Construct ship characteristics based on aggregation functions;

[0013] Construct a mapping table between label codes and conventional ship types, and perform label conversion based on the mapping table;

[0014] Filter out abnormal data based on outlier detection and finally obtain a fused data set;

[0015] Construct the LightGBM ship multi-classification recognition model and set the objective function of the LightGBM ship multi-classification recognition model;

[0016] The fused data set is divided into a training set and a test set, and the LightGBM ship multi-classification recognition model is trained;

[0017] Construct the Focal Loss loss function and integrate it into the LightGBM ship multi-classification recognition model;

[0018] Perform interpretive analysis on the model results and obtain a model interpretability report;

[0019] The trajectory data of a ship observed by radar is obtained and input into the trained LightGBM ship multi-classification recognition model for ship type recognition and prediction.

[0020] As a preferred technical solution, the ship characteristics are constructed based on the aggregation function, specifically aggregating the longitude coordinates, latitude coordinates, heading over the ground, speed over the ground, and time, compressing the continuous trajectory data into a piece of data representing the overall information of the ship, and adding the ship's own information to form multi-dimensional characteristics.

[0021] As a preferred technical solution, the aggregation methods of the aggregation function include maximum value, minimum value, mean value, number of times, as well as ship length and whether the ship is stationary.

[0022] As a preferred technical solution, abnormal data is screened out based on outlier detection, and specifically, an isolation forest algorithm is used for outlier screening.

[0023] As a preferred technical solution, the LightGBM ship multi-classification recognition model is expressed as:

[0024]

[0025] Among them, f t ∈F,F={f(x)=wq(x)}, each f t The learner corresponding to an independent tree structure q and leaf weight w is used to fit the label. The result of each prediction is the residual of the sum of the learner prediction results and the true value y of the ship data classification. i represents the score of the i-th ship data sample set, and wq(x) is the score of the predicted ship data sample x.

[0026] As a preferred technical solution, the objective function of LightGBM is constructed, and the second-order Taylor expansion of the objective function is as follows:

[0027]

[0028] Among them, l represents the loss function, y i represents the true value of the ship classification, Represents the sum of the prediction results of multiple learners that have been trained in the LightGBM ship multi-classification recognition model. is the actual output of the model, f t (x i ) is the predicted value of the ship classification by the currently trained learner, Ω(f t ) is used to control complexity;

[0029] The objective function of LightGBM is derived and expanded by Taylor formula as follows:

[0030]

[0031] Among them, g i is the first-order derivative of the loss function, h i is the second-order derivative of the loss function;

[0032] The tree training of LightGBM is as follows:

[0033]

[0034] Among them, T is the number of leaf nodes, w j is the weight of the leaf node. Each leaf node represents the data division process. Each division gives different data different weights. γ and λ are regularization coefficients.

[0035] As a preferred technical solution, the construction of the Focal Loss function is specifically expressed as:

[0036] FL(p t,k )=-α t,k ·(1-p t,k ) γ ·log(p t,k )

[0037] Among them, p t,k is the probability of the kth category predicted by the model, α t,k is the balancing factor, and γ is a parameter used to adjust the weights of samples that are easy to classify.

[0038] As a preferred technical solution, an explanatory analysis of the model results is conducted, including:

[0039] The LightGBM ship multi-classification recognition model is used to take the number of feature selections in the process of leaf node split selection and the split gain obtained when this feature is selected as a reference for feature importance;

[0040] Use the Shapley value of game theory to analyze the degree to which the characteristics of different samples affect the final results;

[0041] Select and remove some features based on their importance.

[0042] As a preferred technical solution, the game theory Shapley value is defined as follows:

[0043]

[0044] Among them, φ i(f) is the Shapley value of feature i, S is a subset in the feature set, f(S) is the ship classification prediction result of the model under the feature set S, f(S∪{i}) is the ship classification prediction result after adding feature i to the subset S, and |S| is the size of the subset S.

[0045] As a preferred technical solution, the trajectory data of a ship observed by radar is obtained and input into the trained LightGBM ship multi-classification recognition model for ship type recognition and prediction, which specifically includes:

[0046] Obtain the trajectory data of the ship, filter out the trajectory point information belonging to the ship according to the ship's unique identification ID, aggregate the trajectory point information, construct the ship features based on the aggregation function, and input the trained LightGBM ship multi-classification recognition model to obtain the prediction results of the LightGBM ship multi-classification recognition model and output them.

[0047] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0048] (1) To address the problem of too many types of ships in the original data, the present invention adopts a technical solution of converting labels according to the commonness of the ship categories and the number of ships under specific categories, thereby achieving the technical effect of avoiding too many categories and avoiding a decrease in model training accuracy due to too few ships under some categories.

[0049] (2) In order to solve the problem of few feature dimensions in the original data, the present invention adopts a technical solution of aggregating the longitude coordinates, latitude coordinates, heading over the ground, speed over the ground, and time, thereby achieving the technical effect of mining more features representing the ship and expanding the data dimension.

[0050] (3) The present invention uses the ship category information in the AIS data as a label, and then uses the radar data to model the ship type prediction. However, no AIS data is required during the prediction, and the ship type can be judged only by the radar data.

[0051] (4) The LightGBM model of the present invention is based on the gradient boosting tree. During the model construction process, it is necessary to select features as the basis for sample division. The importance of different features is analyzed according to the number of selections of different features during the LightGBM feature construction process and the degree of reduction in the entropy value of the feature, so that the model prediction process is no longer a black box, which facilitates the analysis of the model prediction recognition basis and the optimization of the model, thereby improving the model interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flow chart of the ship classification method based on navigation big data of the present invention;

[0053] Figure 2 It is a schematic diagram of the architecture of the LightGBM algorithm model of the present invention;

[0054] Figure 3 Schematic diagram of the curve of the Focal Loss function of the present invention changing with γ

[0055] Figure 4 This is a schematic diagram of the loss function change curve of the LightGBM training process of the present invention;

[0056] Figure 5 This is a schematic diagram of the LightGBM feature importance of the present invention;

[0057] Figure 6 This is a schematic diagram of the calculation of the LightGBM feature Shapley value of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0059] Example

[0060] like Figure 1 As shown, this embodiment provides a ship classification method based on navigation big data, comprising the following steps:

[0061] S1: Fuse AIS data with ARPA radar data, perform feature engineering on the fused data, and improve and optimize data features, including:

[0062] S11: Fuse AIS data with ARPA radar data, use ARPA radar data as modeling features, and use ship category information in AIS data as labels to obtain a fused data set. Compare the data quality of different fused data sources, including feature coverage, total number of samples, and completeness of data description.

[0063] S12: This embodiment uses a real marine navigation data set, aggregates different trajectory points by fusing target IDs, and then arranges them in order according to timestamps to form a complete record of each ship, so that the trajectories of the same ship are aggregated under the same ship;

[0064] As shown in Table 1 below, because the basic data of the ship itself is redundant in the trajectory information and is not easy to construct into features in the time series model, the longitude coordinates, latitude coordinates, ground heading, ground speed, and time are aggregated separately. The aggregation methods include maximum value, minimum value, mean value, and number of times. In addition, the length of the ship, whether the ship is stationary, and other constructed data are aggregated to obtain a total of 23 dimensions of features. The continuous trajectory data is compressed into a piece of data representing the overall information of the ship, and the ship's own information is added. Compared with the original trajectory information, the data is compressed and the data dimension is expanded. Compressing data can reduce the consumption of model training resources.

[0065] Table 1. Training data feature set

[0066]

[0067]

[0068] S13: constructing features for the aggregated data formed in step S12, that is, constructing ship features such as mean longitude, maximum longitude, maximum latitude, maximum speed, etc. through an aggregation function;

[0069] S14: Label conversion is performed on the data aggregated in step S12. The conversion is based on the mapping relationship between the AIS ship type digital code and the actual ship type. AIS data has a set of label codes for ship types, but there are too many ship types in the label codes. To address this problem, the labels are converted into common ship types through a mapping table of label meanings. The final classification is "yacht", "fishing boat", "passenger ship", "cargo ship", "tanker", "other" and other categories. Label conversion based on the commonness of ship categories and the number of ships in specific categories can avoid too many categories and too few ships in some categories, which can cause the model training accuracy to decrease.

[0070] S15: Perform outlier detection on the data of each ship generated in step S14, automatically filter out outliers using the isolation forest algorithm, and retain normal samples for subsequent steps;

[0071] S2: Establish a LightGBM ship multi-classification recognition model, use the fused data after feature engineering, and adjust the parameters of the LightGBM ship multi-classification recognition model, including:

[0072] S21: Build the LightGBM ship multi-classification recognition model. The LightGBM ship multi-classification recognition model is an efficient gradient boosting decision tree (GBDT) algorithm model. Through the histogram-based algorithm and leaf growth strategy, it greatly improves the training speed and memory efficiency of the model while ensuring the prediction accuracy;

[0073] like Figure 2 As shown, the LightGBM ship multi-classification recognition model can be expressed as:

[0074]

[0075] Among them, f t ∈F,F={f(x)=wq(x)}, each f t The learner corresponding to an independent tree structure q and leaf weight w is used to fit the label. Each prediction result is the residual of the sum of the previous learner prediction results and the true value y of the ship data classification, w i represents the score of the ith node (the set of ship data samples after the abstracted data division), wq(x) is the score of the predicted ship data sample x, that is, the model's predicted value for ship classification;

[0076] The objective function of LightGBM consists of two parts: the loss function and the regularization term. The second-order Taylor expansion of the objective function is as follows:

[0077]

[0078] Among them, l represents the loss function, such as MSE, cross entropy, etc., y i represents the true value of the ship classification, Represents the sum of the prediction results of multiple learners that LightGBM has trained. is the actual output of the model, f t (x i ) is the predicted value of the ship classification by the currently trained learner, and then Ω(f t ) is used to control the complexity, f t The number of leaf nodes is generally used to represent the complexity of the learner. The more leaf nodes there are, the more data division processes there are and the more complex the model is. This is similar to the l1 and l2 regularizations, which can prevent overfitting.

[0079] The objective function of LightGBM is derived and expanded by Taylor formula as follows:

[0080]

[0081] Among them, g i is the first-order derivative of the loss function, h i is the second-order derivative of the loss function. The above process simplifies the constant term that does not work in the optimization process of the loss function.

[0082] The tree training of LightGBM is as follows:

[0083]

[0084] Among them, T is the number of leaf nodes, w j is the weight of the leaf node. Each leaf node represents the data division process. Each division gives different data different weights. γ and λ are regularization coefficients.

[0085] In this embodiment, the LightGBM classification model sets the learning rate to 0.1, the feature sampling rate to 0.9, the sample sampling rate to 0.9, the iteration rounds to 300, the early stopping rounds to 10, and the minimum number of split samples to 50.

[0086] S22: Adjust hyperparameters, loss function, and optimizer to optimize the prediction model. Before model training, model-related hyperparameters need to be set; the model optimizer uses the Adam optimizer, and the loss function uses the MSE loss function; the accuracy is used as the model evaluation function; the training epoch is set to 20, and the batch_size is set to 32;

[0087] S23: Divide the data into a training set and a test set by using a hierarchical division method. The hierarchical division will keep the categories of the training set and the test set close to the entire data set during the data set segmentation process. In this embodiment, the ratio of the training set to the test set is 4:1.

[0088] S3: Optimize and adjust the loss function of the model and construct the Focal Loss loss function so that the model can perform well in the recognition of each category in the category imbalance problem. Based on the adjusted parameters and loss function, the training and verification data results of the model are obtained, including:

[0089] S31: Construct the Focal Loss loss function. The multi-classification Focal Loss is defined as follows:

[0090] FL(p t,k )=-α t,k ·(1-p t,k ) γ ·log(p t,k )

[0091] Among them, p t,k is the probability of the kth category predicted by the model, α t,k is a balance factor (optional), used to adjust the weights of positive and negative samples, usually set to the inverse frequency of each category, and γ is a parameter used to adjust the weights of samples that are easy to classify, usually with a smaller value, and the default value is 1. Figure 3 As shown in the figure, when α is set to 1, the change curve of the Focal Loss loss function is obtained. When γ is larger, the loss ratio of samples that are easy to be classified is smaller;

[0092] S32: Replace the loss function in the model, integrate the Focal Loss loss function into the LightGBM model, and set the α parameter according to the proportion of positive and negative samples;

[0093] S33: Train the model, record the intermediate process of model training, dynamically adjust parameters during training, or use random search and other methods to automatically find the most suitable parameter combination, such as Figure 4 As shown, the change of the loss function value of the model training is obtained;

[0094] S34: Draw a heat map of the model on the training data and test data, calculate the precision and recall of each category, and calculate the macro average and weighted average of each category as a reference for the final training effect of the model.

[0095] S4: Perform interpretive analysis on the model results and obtain a model interpretability report, such as which features are more important to the model's prediction and have a greater impact on the model's final output results;

[0096] In this embodiment, the model explanatory analysis specifically includes:

[0097] S41: The number of feature selections in the process of leaf node split selection using the LightGBM ship multi-classification recognition model and the split gain obtained when selecting this feature are used as a reference for feature importance, such as Figure 5 As shown, according to the number of times the feature is used when splitting the model, the ship length is the most important feature;

[0098] S42: Use the Shapley value of game theory to analyze the degree to which the characteristics of different samples affect the final results. Shapley value is a method derived from game theory, which is used to fairly distribute the contribution of each participant to the overall benefits in a cooperative game. Figure 6 As shown, according to the analysis of Shapley values, the ship length is the most important feature.

[0099] In this example, assuming there is a feature set N and a model f, the goal is to predict an output y, and the Shapley value is calculated by calculating the feature i in all possible feature subsets The marginal contribution in is used to measure the impact of feature i on the model output.

[0100] The Shapley value is defined as follows:

[0101]

[0102] Among them, φ i(f) is the Shapley value of feature i, S is a subset in the feature set that does not include feature i. Assuming that feature i is the longitude mean, S can be a feature set that does not include the longitude mean, such as {dimensional mean, ship speed mean}, {maximum ship speed}, etc., f(S) is the ship classification prediction result of the model under the feature set S, f(S∪{i}) is the ship classification prediction result after adding feature i to the subset S, and |S| is the size of the subset S.

[0103] The Shapley value considers the marginal contribution of feature i in different feature combinations and takes a weighted average of all possible combinations in a combinatorial way to derive the fair contribution of feature i to the overall model prediction.

[0104] S43: Based on the importance of the features analyzed in step S41 and step S42, some unimportant features can be eliminated, such as whether the ship is stationary and the average moving speed of the ship. This can reduce feature redundancy and avoid model overfitting. The maximum and minimum values ​​of the longitude and latitude of important features can also be diverged to discover new features and make full use of feature information, thereby improving the accuracy of model classification prediction.

[0105] S5: The trajectory data of a ship observed by radar is used as the research object and input into the model for ship type identification and prediction, including:

[0106] S51: Process the input sample according to the feature engineering process in step S1, filter out the track point information belonging to the ship according to the ship's unique identification ID, and then aggregate the track point information to obtain the model input features;

[0107] S52: Input the 23 features listed in Table 1 in the radar data obtained in step S51 into the model, obtain the model prediction result and output it.

[0108] The present invention builds a model based on AIS and ARPA radar fusion data (ARPA radar data is used as features and category information in AIS data is used as labels). Then, the ship type can be predicted only by radar data. Two model interpretability methods, namely feature usage times and Shapley value, are provided to improve the accuracy and interpretability of the algorithm.

[0109] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A ship classification method based on navigation big data, characterized in that: The steps include: The AIS data is fused with the ARPA radar data. The ARPA radar data is used as the modeling feature and the ship category information in the AIS data is used as the label to obtain the fused data set, which specifically includes: Get the real trajectory data, aggregate different trajectory points by fusing the target ID, and then arrange them in the order of timestamps to form a complete record of each ship, so that the trajectory of the same ship is aggregated under the same ship; Construct ship characteristics based on aggregation functions; Construct a mapping table between label codes and conventional ship types, and perform label conversion based on the mapping table; Filter out abnormal data based on outlier detection and finally obtain a fused data set; Construct the LightGBM ship multi-classification recognition model and set the objective function of the LightGBM ship multi-classification recognition model; The fused data set is divided into a training set and a test set, and the LightGBM ship multi-classification recognition model is trained; Construct the Focal Loss loss function and integrate it into the LightGBM ship multi-classification recognition model; Perform interpretive analysis on the model results and obtain a model interpretability report; The trajectory data of a ship observed by radar is obtained and input into the trained LightGBM ship multi-classification recognition model for ship type recognition and prediction.

2. The ship classification method based on navigation big data according to claim 1 is characterized in that: The ship characteristics are constructed based on the aggregation function, specifically aggregating the longitude coordinates, latitude coordinates, ground heading, ground speed, and time, compressing the continuous trajectory data into a piece of data representing the overall information of the ship, and adding the ship's own information to form a multi-dimensional feature.

3. The ship classification method based on navigation big data according to claim 2 is characterized in that: The aggregation methods of the aggregation function include maximum value, minimum value, mean value, number of times, length of the ship, and whether the ship is stationary.

4. The ship classification method based on navigation big data according to claim 1 is characterized in that: Abnormal data is filtered out based on outlier detection, and the isolation forest algorithm is used to filter out outliers.

5. The ship classification method based on navigation big data according to claim 1 is characterized in that: The LightGBM ship multi-classification recognition model is expressed as: Among them, f t ∈F,F={f(x)=wq(x)}, each f t The learner corresponding to an independent tree structure q and leaf weight w is used to fit the label. Each prediction result is the residual of the sum of the learner prediction results and the true value y of the ship data classification, w i represents the score of the i-th ship data sample set, and wq(x) is the score of the predicted ship data sample x.

6. The ship classification method based on navigation big data according to claim 5 is characterized in that: Construct the objective function of LightGBM. The second-order Taylor expansion of the objective function is as follows: Among them, l represents the loss function, y i represents the true value of the ship classification, Represents the sum of the prediction results of multiple learners that have been trained in the LightGBM ship multi-classification recognition model. is the actual output of the model, f t (x i ) is the predicted value of the ship classification by the currently trained learner, Ω(f t ) is used to control complexity; The objective function of LightGBM is derived and expanded by Taylor formula as follows: Among them, g i is the first-order derivative of the loss function, h i is the second-order derivative of the loss function; The tree training of LightGBM is as follows: Among them, T is the number of leaf nodes, w j is the weight of the leaf node. Each leaf node represents the data division process. Each division gives different data different weights. γ and λ are regularization coefficients.

7. The ship classification method based on navigation big data according to claim 1 is characterized in that: The construction of the FocalLoss loss function is specifically expressed as: FL(p t,k )=-a t,k ·(1-p t,k ) γ ·log(p t,k ) Among them, p t,k is the probability of the kth category predicted by the model, α t,k is the balancing factor, and γ is a parameter used to adjust the weights of samples that are easy to classify.

8. The ship classification method based on navigation big data according to claim 1 is characterized in that: Perform interpretive analysis on the model results, including: The LightGBM ship multi-classification recognition model is used to take the number of feature selections in the process of leaf node split selection and the split gain obtained when this feature is selected as a reference for feature importance; Use the game theory Shapley value to analyze the degree to which the characteristics of different samples affect the final results; Select and remove some features based on their importance.

9. The ship classification method based on navigation big data according to claim 8 is characterized in that: The game theory Shapley value is defined as follows: Among them, φ i (f) is the Shapley value of feature i, S is a subset in the feature set, f(S) is the ship classification prediction result of the model under the feature set S, f(S∪{i}) is the ship classification prediction result after adding feature i to the subset S, and |S| is the size of the subset S.

10. The ship classification method based on navigation big data according to claim 1 is characterized in that: Obtain the trajectory data of a ship observed by radar and input it into the trained LightGBM ship multi-classification recognition model to perform ship type recognition and prediction, including: Obtain the trajectory data of the ship, filter out the trajectory point information belonging to the ship according to the ship's unique identification ID, aggregate the trajectory point information, construct the ship features based on the aggregation function, and input the trained LightGBM ship multi-classification recognition model to obtain the prediction results of the LightGBM ship multi-classification recognition model and output them.

Images