CN118333397A - Prediction method for severity of marine traffic accident - Google Patents

Abstract

The present invention provides a method for predicting the severity of marine traffic accidents, comprising: using marine accident investigation reports to construct a data set of factors affecting marine accident risks; based on the constructed data set of factors affecting marine accident risks, using a feature selection method to train the accuracy of a machine learning model and the interpretability of feature selection; using a three-stage performance evaluation method of stability evaluation, prediction performance evaluation, comprehensive evaluation and statistical test to evaluate the performance of the feature selection method; using six machine learning models for comparison to measure the performance of different prediction factors, and using the machine learning model with the best prediction performance of the severity of ship accidents as a benchmark model; using the selected model with the highest prediction performance and the best features to predict the severity of the accident, and performing benefit evaluation, and counterfactually analyzing the effects of risk control measures from a quantitative perspective. The method of the present invention can effectively analyze and predict the severity of marine accidents.

Description

A method for predicting the severity of marine traffic accidents

Technical Field

The present invention relates to the technical field of marine accident risk prediction, and in particular to a method for predicting the severity of marine traffic accidents.

Background technique

In recent years, with the development of artificial intelligence technology, machine learning (ML) has been widely used to analyze maritime accidents and improve navigation safety and efficiency. Its application is indispensable for preventing accidents and establishing a safe marine environment. Accident prediction is the basis for scientific decision-making and is crucial for accident prevention. Machine learning is effective in identifying risk influencing factors (RIFs) and predicting accidents. However, its current application is more limited to accident frequency analysis and risk influencing factor identification, and its ability to predict the severity of maritime accidents has yet to be developed.

Machine learning technology can learn and explore the patterns of accidents of different severity, which is essential for targeted accident prevention, post-accident identification and simulation. However, few studies have attempted to estimate or predict the severity of maritime accidents. Some studies have proposed an overlay model to identify accident-prone areas and predict the severity of accidents based on the spatial distribution of maritime accidents. By analyzing the characteristics of accident clustering and spatial correlation, the pattern of accident severity is determined, and with the help of the overlay model, traffic characteristics that are closely related to the severity of the accident are determined, and predictions are made based on these characteristics. There is also the use of data-driven methods to predict the severity of maritime casualties, which is to determine the key aspects of predicting the severity of ship collision accidents by analyzing the relationship between the risk influencing factors of such accidents. Although the above studies provide valuable information and insights for maritime accident analysis, there are still some research gaps that need to be filled:

1. Most studies on the prediction of the severity of marine accidents have only examined a few risk factors and have not yet conducted a comprehensive analysis from the perspective of safety system engineering.

2. There are relatively few studies focusing on the prediction of the severity of maritime accidents, among which there are deficiencies in the accuracy of data processing methods or prediction models, such as data leakage and high recall rate.

3. Existing studies rarely present both the severity of accidents and related risk control measures, and even fewer involve quantitative analysis of the effectiveness of risk control measures.

The above-mentioned background technology is intended to assist in understanding the inventive concept and technical solution of the present invention. It does not necessarily belong to the prior art of this patent application. In the absence of clear evidence that the above-mentioned content has been disclosed before the filing date of this patent application, the above-mentioned background technology should not be used to evaluate the novelty of the technical solution of this application.

Summary of the invention

According to the above-mentioned current situation of insufficient research on the prediction of the severity of marine accidents, a method for predicting the severity of marine traffic accidents is provided with the goal of preventing serious marine accidents and providing a safe marine environment. The present invention first analyzes relevant marine accident investigation reports from the perspective of safety system engineering to identify risk influencing factors and establish accident severity labels. Secondly, three feature selection (FS) methods are developed and used to determine key risk influencing factors and evaluate their performance. Then, 6 state-of-the-art machine learning models are used to train a data set containing key risk influencing factors and predict the severity of the accident. Finally, the key risk influencing factors are interpreted and analyzed to understand the significant benefits of controlling specific risk influencing factors.

In order to solve at least one of the technical problems mentioned in the above background technology, the present invention aims to provide a method for predicting the severity of maritime traffic accidents, which mainly involves five aspects: establishing a database from the perspective of safety system engineering; developing a feature selection method; performance evaluation of the feature selection method; accident severity prediction evaluation; and effectiveness evaluation of risk control measures based on counterfactual analysis. The present invention can not only provide a basis for safety assessment and risk prevention and control research, but also effectively avoid the problems of data leakage in previous studies, filling the research gap in the field of maritime accident severity prediction.

The technical means adopted by the present invention are as follows:

A method for predicting the severity of a marine traffic accident, comprising:

S1. Using marine accident investigation reports, construct a dataset of factors affecting marine accident risks;

S2. Based on the constructed dataset of factors affecting marine accident risks, feature selection methods are used to train the accuracy of machine learning models and the interpretability of feature selection;

S3, using a three-stage performance evaluation method of stability evaluation, prediction performance evaluation, comprehensive evaluation and statistical test to evaluate the performance of the feature selection method;

S4. Six machine learning models were used for comparison to measure the performance of different predictors, and the machine learning model with the best performance in predicting the severity of ship accidents was used as the benchmark model;

S5. Use the selected model with the highest prediction performance and the best features to predict the severity of the accident, conduct benefit evaluation, and analyze the effectiveness of risk control measures from a quantitative perspective.

Furthermore, step S1 specifically includes:

S11. Extract human, ship, environment, management factors and basic accident information from the accident investigation report as standard-level features of the marine accident risk influencing factor dataset;

S12, classifying standard-level features into 68 index-level features during the data processing stage;

S13. Convert continuous data and character data into discrete categorical data.

Furthermore, step S2 specifically includes:

S21. Feature selection method based on association rule mining and three-weight ranking algorithm to mine the interactions between accident features affecting marine accidents and rank them;

S22. Adopting the feature selection method based on the new mutual information entropy (NMIE) guided by accident consequences, the influencing factors are ranked by using the information entropy of each influencing factor and the severity of the accident;

S23. Using the feature selection method adopted in step S21, the relationship between each feature is mined and the features are sorted. Then, using the feature selection method adopted in step S22, the mutual information entropy between the features not mined by the feature selection method adopted in step S22 and the target is calculated, and the mutual information entropy between the features and the target is sorted. The two sorting results are combined to obtain the final feature sorting result.

Furthermore, step S21 specifically includes:

S211. Use association rule mining technology to find the association, co-occurrence or causal relationship between accident features: In order to accurately and effectively find the relationship between accident features, the FP-Growth algorithm is used to mine frequent item sets, and the association rules are generated by constructing an FP tree to screen out important association rules; if an association rule meets the minimum support and confidence, and the lift is greater than 1, then the association rule is considered valid. Valid association rules are used The calculation formula is as follows:

Among them, Supp(X) represents the support of transaction X; express confidence level; express The degree of improvement; N _X represents the number of occurrences of X transaction in the project set; N represents the number of transactions in the project set;

S212. Use graph theory methods to map the association rules between different accident characteristics to the complex influencing factor interaction network; the nodes in the complex influencing factor interaction network represent accident characteristics, and the edges represent the associations between them; the weights of the edges correspond to the confidence of the association rules; the complex influencing factor interaction network reflects the interaction of various accident characteristics in the development process of marine accidents; use the feature selection method based on association rule mining and three-weight ranking algorithm to determine the importance of each accident feature in the complex influencing factor interaction network, and rank the accident features accordingly.

Furthermore, the feature selection method based on association rule mining and three-weight ranking algorithm adopted in step S21 includes two optimization improvements, which are as follows:

The first improvement: The concept of a weighted transition probability matrix is proposed, and confidence is used as the transition probability between adjacent risk influencing factors to replace the definition in the original algorithm. The calculation formula is as follows:

Among them, PM represents the transition probability matrix; n represents the number of influencing factors mined by the FP-Growth algorithm; g represents the Ground node added to the original network; δ _ij represents the transition probability from node _ni to node _nj ; i→j represents the existence of an edge from node _ni to node _nj ; represents the in-degree of node n _j ; e _ji represents the association between influencing factor n _j and influencing factor n _i ; t represents the number of iterations; represents the score obtained by node n _i at the tth iteration; (matrix) _ij represents an n+1-order square matrix with only the i-th row and j-th column elements being 1 and the rest being 0; (zeros-one) _j represents a 1×(n+1)-order matrix with only the j-th column elements being 1 and the rest being 0; (ones) represents an (n+1)×1-order matrix with all elements being 1; * represents the Hadamard product; · represents the matrix product;

The second improvement: Based on the theory of node external strength and node centrality coefficient in graph theory, a new comprehensive weight is proposed. After the algorithm iteration is completed, the comprehensive weight is used to obtain additional scores from the final score of the Ground node. The calculation formula is as follows:

in, represents the output intensity of influencing factor n _i ; CC _i represents the betweenness central coefficient of influencing factor n _i ; σ _ij (v) represents the number of shortest paths from influencing factor n _i to influencing factor n _j through influencing factor n _v ; σ _ij represents the number of shortest paths from influencing factor n _i to influencing factor n _j ; μ _i represents the comprehensive weight proposed in this study; t _end represents the number of iterations at the end of the algorithm iteration.

Furthermore, in step S22, the feature selection method based on the new mutual information entropy (NMIE) guided by accident consequences is calculated by the following formula:

Among them, NMIE(X,Y) represents the mutual information entropy between feature X and label Y; m represents the number of states of the label; n represents the number of states of feature X; I( _xi , _yk ) represents the mutual information between state i of feature X and state k of the label.

Furthermore, step S3 specifically includes:

S31. Stability evaluation: The stability of the feature selection method is measured by considering the influence of randomness and changes in the size of the training set. Methods similar to k-folding, random numbers, and repeated experiments are used to determine the influence of randomness and changes in the size of the training set on the feature selection process. Spearman's rho is used to measure the stability of the ranking results. The calculation formula is as follows:

Where Sprr ^R1,R2 represents the correlation between ranking R1 and ranking R2; R1 is used as the reference ranking; R2 is used as the experimental ranking; N represents the number of main features; M represents all features included in the main features of R1; if feature i is not in the main features of R2, then R2 _i = N+1;

S32. Prediction performance evaluation: Train the prediction model through the risk influencing factors selected by features, measure the accuracy of the prediction model, find the subset of risk influencing factors that achieves the highest accuracy with the least risk influencing factors, and use the corresponding indicators to evaluate the prediction performance of the model;

S33. Comprehensive evaluation and statistical test: The stability of the feature selection process is evaluated by determining whether the feature selection method has an acceptable level of stability; if the best stability of the feature selection method is lower than 0.6, or the difference with the feature selection method with the best stability is greater than 0.3, the stability of the feature selection method is considered to be unacceptable; unacceptable stability indicates that the feature selection method is invalid; based on the prediction performance evaluation information, the prediction performance of the feature selection method with acceptable stability is analyzed; if the prediction performance indicators of different feature selection methods are not similar, then the method with better prediction performance indicators is considered to be the better performing method; if the prediction performance indicators are the same, a 5-fold cross-validation is performed to test the difference in the prediction performance of the machine learning model; if the test result is significant greater than 0.05, there is no significant difference in the prediction performance of the two methods; if there is similarity in the prediction performance of different feature selection methods, the feature selection method with better stability is considered to have better performance.

Furthermore, step S4 specifically includes:

S41, introduce six machine learning models: support vector machine, naive Bayes classifier, random forest, AdaBoost, XGBoost, and LightGBM;

S42. Accuracy, precision, recall, F1 value and area under the curve were used to evaluate the prediction performance of the machine learning model, and the optical gradient enhancement machine with the best prediction performance for ship accident severity prediction was taken as the benchmark model.

Compared with the prior art, the present invention has the following advantages:

1. The method for predicting the severity of marine traffic accidents provided by the present invention can effectively analyze and predict the severity of marine accidents, and provide a theoretical basis and application example for applying artificial intelligence technology to safety assessment and accident prevention research. Shipping companies and competent authorities can use the framework and improved machine learning model proposed by the present invention to develop a marine accident severity prediction system and establish a more effective risk control system. They can also use the feature selection method proposed in this study to study the interaction relationship between risk influencing factors in the marine accident system (i.e. mutual drive and dependence, mutual causality, etc.) and the correlation between risk influencing factors and accident severity, so as to find out the risk influencing factors that are crucial to marine accidents, and use the feature selection method evaluation framework proposed by the present invention to evaluate the effectiveness of the key risk influencing factors identified.

2. The method for predicting the severity of marine traffic accidents provided by the present invention has established a new research framework for predicting the severity of marine accidents. The framework mainly includes five aspects: establishing a database from the perspective of safety system engineering; developing a feature selection method; performance evaluation of the feature selection method; accident severity prediction evaluation and effectiveness evaluation of risk control measures based on counterfactual analysis.

3. The method for predicting the severity of maritime traffic accidents provided by the present invention has developed three feature selection methods for maritime accidents. Among them, the first method mainly uses association rule mining, improved weighted LeaderRank algorithm and complex network technology to mine the correlation between risk influencing factors; the second method mainly uses the improved mutual information entropy algorithm to analyze the correlation between risk influencing factors and the severity of accidents; the third method combines the advantages of the first method and the second method, and comprehensively considers the interaction between risk influencing factors and the relationship between different states of risk influencing factors and the severity of accidents.

4. The marine traffic accident severity prediction method provided by the present invention has developed a new evaluation method to comprehensively measure the performance of the proposed feature selection method from multiple aspects such as stability, performance improvement and statistical test. In the stability evaluation process, a new convergence and stability algorithm is proposed to improve the performance of stability evaluation.

5. The method for predicting the severity of marine traffic accidents provided by the present invention conducts a comparative analysis of the main classic machine learning models, compares their prediction performance, provides a baseline model for developing a marine accident severity prediction model, and uses statistical methods to verify the effectiveness of the feature selection method in improving the performance of the machine learning model.

Based on the above reasons, the present invention can be widely promoted in the fields of marine accident risk prediction and the like.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

FIG1 is a flow chart of the method of the present invention.

FIG. 2 is a source distribution diagram of marine traffic accident reports provided by an embodiment of the present invention.

FIG3 is a feature selection method based on association rule mining and 3-WLR provided by an embodiment of the present invention.

FIG. 4 is a frequency distribution tree diagram of all risk influencing factors provided by an embodiment of the present invention.

FIG. 5 is a distribution diagram of accident severity provided by an embodiment of the present invention.

FIG6 is a stability comparison diagram of the feature selection method provided by an embodiment of the present invention.

FIG. 7 shows the prediction accuracy based on different feature selection methods provided by an embodiment of the present invention.

FIG8 is a generalization performance diagram of six advanced machine learning models based on risk influencing factors selected by three feature selection methods provided in an embodiment of the present invention.

FIG. 9 shows the win, draw and loss percentages between different machine learning models provided by an embodiment of the present invention.

FIG. 10 is a diagram showing the benefits of a single risk control measure provided by an embodiment of the present invention.

Detailed ways

In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

As shown in FIG1 , the present invention provides a method for predicting the severity of a marine traffic accident, comprising:

S1. Using marine accident investigation reports, construct a dataset of factors affecting marine accident risks;

S2. Based on the constructed dataset of factors affecting marine accident risks, a feature selection method is used to train the accuracy of the machine learning model and the interpretability of feature selection. In this embodiment, the three feature selection methods proposed in the present invention and four traditional machine learning feature selection methods such as RFLV, SVM, RF and GBDT as benchmark methods are all trained on the training set data of step S1.

S3, using a three-stage performance evaluation method of stability evaluation, prediction performance evaluation, comprehensive evaluation and statistical test to evaluate the performance of the feature selection method;

S4. Six machine learning (ML) models were used for comparison to measure the performance of different predictors, and the machine learning model with the best performance in predicting the severity of ship accidents was used as the benchmark model;

S5. Use the selected model with the highest prediction performance and the best features to predict the severity of the accident, conduct benefit evaluation, and analyze the effectiveness of risk control measures from a quantitative perspective.

In specific implementation, as a preferred embodiment of the present invention, the marine accident risk influencing factor dataset constructed in step S1 is composed of the risk influencing factors (i.e., accident characteristics) and accident severity labels (i.e., target categories) of each accident. The construction process of the marine accident risk influencing factor dataset includes three stages: marine accident dataset extraction, accident severity labeling, and accident feature extraction. Step S1 specifically includes:

S11. Extract human, ship, environment, management factors and basic accident information from the accident investigation report as standard-level features of the marine accident risk influencing factor dataset;

S12, classifying standard-level features into 68 index-level features during the data processing stage;

S13. Convert continuous data and character data into discrete categorical data.

In this embodiment, the present invention uses the data from 2000 to 2019 in the databases of seven maritime investigation agencies, including the China Maritime Safety Administration (MSA), the United States Federal Maritime Accident Investigation Bureau (BSU), the United States National Transportation Safety Board (NTSB), the Japan Transport Safety Board (JTSB), the Australian Transport Safety Board (ATSB), the Canadian Transportation Safety Board (TSB) and the Marine Accident Investigation Bureau (MAIB) as the main data source, and screens the accident reports to finally obtain 1294 maritime accident report data. Based on these 1294 maritime accident investigation reports, 68 risk influencing factors were identified and divided into five categories: human factors, ship factors, environmental factors, management factors and accident information. Among them, the source distribution of the accident report is shown in Figure 2. The specific description of the features is shown in Table A in Appendix A. Figure 4 shows the frequency distribution of all risk influencing factors. Figure 4 shows that among the five types of risk influencing factors, human factors and management factors appear more frequently. Among the human factors, operational errors (H5) and violations of rules and regulations (H6) appear most frequently. Among the management factors, safety management system defects (M3) occur more frequently. Figure 4 shows that collision (A1) and grounding/stranding (A2) are the main types of marine accidents. Bulk carriers (ST1) and fishing vessels (ST7) are the main types of ships involved in accidents. When most accidents occurred, the traffic conditions in the waterway were generally busy (ET). At the same time, Figure 5 shows the distribution of accident severity under different accident time, accident type, ship type and ship age conditions in a bar chart. It can be seen from Figure 5(a) that serious accidents mostly occurred between 0 and 4 o'clock (T6), while Figure 5(b) shows that the most frequent serious accident is capsizing/sinking accident (A5). It can be seen from Figure 5(c) that the severity of accidents varies greatly depending on the type of ship. Fishing vessels (ST7) are most likely to have serious accidents. Figure 5(d) shows that the proportion of ships with serious accidents increases with the increase of ship age.

In specific implementation, as a preferred embodiment of the present invention, step S2 specifically includes:

S21, a feature selection method based on association rule mining and a three-weight ranking algorithm (3-WLR) to mine the interactions between accident features that affect marine accidents and rank them; step S21 specifically includes:

S211. Use association rule mining technology to find the association, co-occurrence or causal relationship between accident features: This step provides valuable information between accident features. In order to accurately and effectively find the relationship between accident features, the FP-Growth algorithm is used to mine frequent item sets, and association rules are generated by constructing FP trees. Three commonly used evaluation indicators, namely support, confidence and lift, are used to screen out important association rules; if an association rule meets the minimum support and confidence, and the lift is greater than 1, then the association rule is considered valid. Valid association rules are used The calculation formula is as follows:

Among them, Supp(X) represents the support of transaction X; express confidence level; express The degree of improvement; N _X represents the number of occurrences of X transaction in the project set; N represents the number of transactions in the project set;

S212. Use graph theory methods to map the association rules between different accident characteristics to the complex influencing factor interaction network; the nodes in the complex influencing factor interaction network represent accident characteristics, and the edges represent the associations between them; the weights of the edges correspond to the confidence of the association rules; the complex influencing factor interaction network reflects the interaction of various accident characteristics in the development process of marine accidents; use the feature selection method based on association rule mining and three-weight ranking algorithm to determine the importance of each accident feature in the complex influencing factor interaction network, and rank the accident features accordingly.

The feature selection method based on association rule mining and three-weight ranking algorithm (3-WLR) adopted in step S21 includes two optimization improvements, which are as follows:

The first improvement: The concept of a weighted transition probability matrix is proposed, and confidence is used as the transition probability between adjacent risk influencing factors to replace the definition in the original algorithm. The calculation formula is as follows:

Among them, PM represents the transition probability matrix; n represents the number of influencing factors mined by the FP-Growth algorithm; g represents the Ground node added to the original network; δ _ij represents the transition probability from node _ni to node _nj ; i→j represents the existence of an edge from node _ni to node _nj ; represents the in-degree of node n _j ; e _ji represents the association between influencing factor n _j and influencing factor n _i ; t represents the number of iterations; represents the score obtained by node n _i at the tth iteration; (matrix) _ij represents an n+1-order square matrix with only the i-th row and j-th column elements being 1 and the rest being 0; (zeros-one) _j represents a 1×(n+1)-order matrix with only the j-th column elements being 1 and the rest being 0; (ones) represents an (n+1)×1-order matrix with all elements being 1; * represents the Hadamard product; · represents the matrix product;

The second improvement: Based on the theory of node external strength and node centrality coefficient in graph theory, a new comprehensive weight is proposed. After the algorithm iteration is completed, the comprehensive weight is used to obtain additional scores from the final score of the Ground node. The calculation formula is as follows:

in, represents the output intensity of influencing factor n _i ; CC _i represents the betweenness central coefficient of influencing factor n _i ; σ _ij (v) represents the number of shortest paths from influencing factor n _i to influencing factor n _j through influencing factor n _v ; σ _ij represents the number of shortest paths from influencing factor n _i to influencing factor n _j ; μ _i represents the comprehensive weight proposed in this study; t _end represents the number of iterations at the end of the algorithm iteration.

S22, using a feature selection method based on a new mutual information entropy (NMIE) guided by accident consequences, and using the information entropy of each influencing factor (feature) and the severity of the accident to sort the influencing factors (features); in step S22, the feature selection method based on a new mutual information entropy (NMIE) guided by accident consequences, the calculation formula of the mutual information entropy is as follows:

Among them, NMIE(X,Y) represents the mutual information entropy between feature X and label Y; m represents the number of states of the label; n represents the number of states of feature X; I( _xi , _yk ) represents the mutual information between state i of feature X and state k of the label.

S23. Using the feature selection method adopted in step S21, the relationship between each feature is mined and the features are sorted. Then, using the feature selection method adopted in step S22, the mutual information entropy between the features not mined by the feature selection method adopted in step S22 and the target is calculated, and the mutual information entropy between the features and the target is sorted. The two sorting results are combined to obtain the final feature sorting result.

In specific implementation, as a preferred embodiment of the present invention, step S3 specifically includes:

S31. Stability evaluation: The stability of the feature selection method is measured by considering the influence of randomness and changes in the size of the training set. Methods similar to k-folding, random numbers, and repeated experiments are used to determine the influence of randomness and changes in the size of the training set on the feature selection process. Spearman's rho is used to measure the stability of the ranking results. The calculation formula is as follows:

Where Sprr ^R1,R2 represents the correlation between ranking R1 and ranking R2; R1 is used as the reference ranking; R2 is used as the experimental ranking; N represents the number of main features; M represents all features included in the main features of R1; if feature i is not in the main features of R2, then R2 _i = N+1;

In this embodiment, the metric used in the stability evaluation method has both an upper limit and a lower limit. First, since the Spearman's rho used in the stability evaluation method is less than or equal to 1, the metric has an upper limit. Second, That is, Spearman's rho is greater than or equal to -1, so this indicator has a lower limit. In addition, if the ranking results are more similar, Spearman's rho will become more significant. This means that if the feature selection method is more stable, Spearman's rho will be closer to 1, and if it is less stable, Spearman's rho will be closer to -1. Finally, this indicator is monotonically increasing and can be determined by the above formula.

S32. Prediction performance evaluation: train the prediction model through the risk influencing factors selected by features, measure the accuracy of the prediction model, find out the subset of risk influencing factors that achieves the highest accuracy with the least risk influencing factors (in this embodiment, the marine accident risk influencing factor data set is divided into 80% training set and 20% test set by random sampling without playback. At the same time, the unbalanced data is processed using the support vector machine-synthetic minority oversampling technique (SVM-SMOTE), which specializes in processing class imbalanced data.), and use the corresponding indicators to evaluate the prediction performance of the model;

S33. Comprehensive evaluation and statistical test: The stability of the feature selection process is evaluated by determining whether the feature selection method has an acceptable level of stability; if the best stability of the feature selection method is lower than 0.6, or the difference with the feature selection method with the best stability is greater than 0.3, the stability of the feature selection method is considered to be unacceptable; unacceptable stability indicates that the feature selection method is invalid; based on the prediction performance evaluation information, the prediction performance of the feature selection method with acceptable stability is analyzed; if the prediction performance indicators of different feature selection methods are not similar, then the method with better prediction performance indicators is considered to be the better performing method; if the prediction performance indicators are the same, a 5-fold cross-validation is performed to test the difference in the prediction performance of the machine learning model; if the test result is significant greater than 0.05, there is no significant difference in the prediction performance of the two methods; if there is similarity in the prediction performance of different feature selection methods, the feature selection method with better stability is considered to have better performance.

In this embodiment, the stability of each feature selection method is measured using the ranking-based stability algorithm proposed above. As shown in Figure 6(a), the 3-WLR algorithm proposed in the present invention shows stronger robustness in terms of feature selection method stability. When the number of risk influencing factors selected exceeds 21, the feature selection method stability of the 3-WLR algorithm remains above 0.9. In contrast, the highest stability values of the PageRank algorithm and the WLR algorithm are 0.8461 and 0.8541, respectively. According to Figure 6(b), when tested on the marine accident risk influencing factor data set, the traditional embedded feature selection techniques based on GBDT and SVM did not show sufficient stability. This may be due to the lack of robustness and versatility of the classifiers, which prevents them from showing satisfactory performance in terms of stability. In contrast, the RFLV-based filtering method and the random forest-based embedded feature selection method have better stability. This may be because the variance test of RFLV is not affected by changes in the size of the same distribution data, while the random forest applies the Bagging framework to reduce the classifier prediction error and enhance stability. Compared with traditional feature selection methods (RF, RFLV, GBDT and SVM, etc.) of machine learning, the three feature selection methods proposed in the present invention all show higher stability in feature selection. The stability of the 3-WLR algorithm and the feature selection method used in step S23 is significantly stronger than that of the NMIE algorithm. This may be because the association rule mining technology (FP-Growth algorithm) is insensitive to data perturbations and is more robust. Comparing the 3-WLR algorithm and the feature selection method used in step S23, when the number of risk influencing factors is small, the stability of the feature selection method used in step S23 is higher than that of the 3-WLR algorithm. As the number of risk influencing factors increases (more than 30), the stability of the 3-WLR algorithm gradually becomes higher than that of the feature selection method used in step S23.

In specific implementation, as a preferred embodiment of the present invention, step S4 specifically includes:

S41, introduce six machine learning models: support vector machine (SVM), naive Bayes classifier (NB), random forest (RF), AdaBoost, XGBoost, and LightGBM;

Support Vector Machine (SVM): The main model is a linear classifier defined on a feature space with maximum margins; the basic idea is to solve a separating hyperplane that correctly divides the training data set and has a large geometric gap. SVM can also use kernel functions to solve nonlinear classification problems;

Naive Bayesian Classifier (NB): NB is a prediction model based on the assumption that each feature and sample data is independent. It has stable classification efficiency and performs well when predicting a small amount of data. This study uses a simple Bayesian model based on Bernoulli distribution to predict accident severity.

Random Forest (RF): RF is a standard ensemble learning model. It mainly adopts a bagging framework, which uses sampling techniques to select sample data and train multiple independent decision trees. Finally, it uses voting to get the final prediction result.

AdaBoost: AdaBoost is an ensemble learning method based on a boosting framework. It uses the technique of changing sample weights to focus on samples with poor training results, thereby improving training accuracy. Finally, the weighted summation method is used to output the results of all decision trees in the model to obtain the final prediction result.

XGBoost: XGBoost is also an ensemble learning method based on the boosting framework, which mainly uses the gradient boosting method based on the second-order Taylor expansion. This technology mainly targets under-trained samples and continuously adjusts the labels of actual samples. XGBoost has the characteristics of high operating efficiency and high prediction accuracy, and can also automatically handle missing data.

LightGBM: LightGBM and XGBoost are prediction models based on gradient boosting and second-order Taylor expansion. However, LightGBM uses techniques such as histogram optimization algorithm and gradient-based unilateral sampling to improve the computational efficiency and generalization ability of the model. At the same time, it adopts a maximum depth limit strategy to reduce the amount of calculation and prevent overfitting.

S42. Accuracy, precision, recall, F1 value and area under the curve (AUC) were used to evaluate the prediction performance of the machine learning model, and the optical gradient enhancement machine with the best prediction performance for ship accident severity prediction was taken as the benchmark model.

In this embodiment, accuracy measures the proportion of correctly identified samples to the total number of samples, while precision measures the proportion of samples that are actually positive among the samples identified as positive classes. Recall measures the proportion of correctly identified positive samples to all positive samples. Accuracy and recall are contradictory indicators. A high precision leads to a low recall, and vice versa. The F1 value is the weighted average of accuracy and recall. The area under the curve represents the area under the receiver operating characteristic curve. The closer the area under the curve is to 1, the more accurate the model result is.

In this embodiment, taking the LightGBM model as an example, a 5-fold cross validation is performed on the training data after feature selection. The experimental results are shown in Table 1 and Figure 7. The prediction performance of different feature selection methods for other machine learning models is shown in Appendix B. By analyzing Figure 7 and Table 1, it can be found that the 3-WLR algorithm selects a small number of risk influencing factors, and has better prediction performance than PageRank using the WLR algorithm. This means that the improved 3-WLR algorithm is indeed effective in the present invention. The results shown in Figure 7 and Table 1 also show that the LightGBM predictor trained on the risk influencing factors selected by the three feature selection methods proposed in the present invention has better machine learning performance than the four traditional feature selection methods. In particular, the feature selection method used in step S23 selects the least number of risk influencing factors and has the best prediction performance.

Table 1 Prediction performance of the number of risk influencing factor features selected based on different methods

The present invention aims to accurately predict the severity of marine accidents and develop a benchmark model for accident severity prediction. To this end, the generalization performance of six advanced machine learning models was compared. First, three different feature selection methods were used to select risk influencing factors from the training data. Secondly, the training data was balanced using the SVM-SMOTE technique. Then, six advanced machine learning models were trained on the balanced training data. Finally, the generalization performance of the models was evaluated using test data. Figure 8 shows the generalization performance of the six machine learning models. Figure 8 shows that the generalization performance of the NB model is poor, while the generalization performance of the ensemble learning models of the three Boosting frameworks, LightGBM, AdaBoost and XGBoost, is more prominent.

In order to more intuitively compare the generalization performance of each machine learning model, this section compares the winners and losers of each machine learning model in pairs according to the maximum values of each performance indicator shown in Figure 8 (as shown in Figure 9). From Figure 9, it can be seen that the LightGBM model has the best generalization performance, followed by the AdaBoost and XGBoost models; compared with the SVM model, the RF model has slightly better generalization performance; in particular, the area under the curve of the RF model is larger than that of the SVM model. This shows that the authenticity of the RF model results is higher than that of the SVM model, while the NB model has the worst generalization performance.

According to the results shown in Figure 9, the generalization performance of the LightGBM model after feature selection is optimal. The improvement of the generalization performance of the LightGBM model by the feature selection method of the present invention is shown in Table 2. It can be found from Table 2 that the number of features selected by the 3-WLR algorithm and the feature selection method used in step S23 is the smallest, that is, compared with the NMIE algorithm, these two feature selection methods can remove more redundant information. At the same time, the improvement of the generalization performance of the LightGBM model by the NMIE algorithm is also small. It can be found from Table 2 that the feature selection method used in step S23 has the greatest improvement on the overall generalization performance of the LightGBM model.

Table 2 Improving the generalization performance of the predictor (Light GBM) through feature selection methods

In specific implementation, as a preferred embodiment of the present invention, in step S5, the model with the highest prediction performance and the optimal features are used to evaluate the benefits of preventing serious accidents that can be achieved by controlling certain human factors and management factors. First, the present invention takes the feature selection method used in step S23 as an example to explain the selection of key features. When the minimum support is 0.1 and the minimum confidence is 0.3, the feature sorting results of the feature selection method used in step S23 are shown in Table 3. The order of the first 30 features in Table 3 is first mined by association rules and then sorted by the 3-WLR algorithm, and the order of the last 38 features is obtained by the mutual information entropy algorithm. The model is trained by selecting the LightGBM model and the first 41 risk influencing factors selected in the feature selection method used in step S23 to obtain an accuracy and reliability of 82.63% and 81.98% respectively. Secondly, in order to minimize the impact of model performance on the evaluation results, the present invention selects 914 serious accident data correctly predicted by the model as evaluation data. Then, new evaluation data is generated by "do-operator" (changing the state of certain factors in the evaluation data). Finally, the model is used to predict the severity status of new assessment data, and the proportion of non-serious accidents is used to evaluate the effectiveness of each factor.

Table 3 Feature ranking based on the feature selection method used in step S23

Note: min_sup = 0.1; min_con = 0.3

The human factors and management factors in the first 41 risk influencing factors of the feature selection method adopted in step S23 are controlled, and the specific benefits after control are obtained, as shown in Figure 10. Figure 10 shows that the shipping company obtains the highest benefit by hiring experienced crew members, that is, controlling H3, which helps prevent 10.12% of serious accidents. The second most effective strategy is to correct safety problems in a timely manner (control M4), which can reduce 7.00% of serious accidents. However, controlling H2 or M3 is the least effective strategy. Because Table 3 shows that M3 is the most important risk influencing factor, but in Figure 10, controlling M3 is one of the least effective strategies. The main reason why controlling M3 is not effective in reducing serious accidents is that M3 promotes the occurrence of non-serious accidents and inhibits the occurrence of serious accidents. Regarding the most important results of M3 in prediction, first, it is the most important from the perspective of the entire accident system, because the entire accident system includes both serious accidents and non-serious accidents. Secondly, improving the accuracy of prediction mainly depends on whether the risk influencing factors can distinguish between serious accidents and non-serious accidents, rather than on whether controlling the risk influencing factors can effectively reduce the severity of accidents.

In summary, the present invention provides a research framework for predicting the severity of maritime traffic accidents based on machine learning. The framework can effectively study maritime traffic accidents from four aspects: data set construction, feature selection, feature selection performance evaluation, and accident severity prediction. It not only comprehensively analyzes accident influencing factors from the perspective of safety system engineering, but also incorporates these risk influencing factors as accident characteristics into the maritime accident risk influencing factor data set. The present invention not only provides a more advanced methodological inspiration for maritime accident prediction research, but also has certain application value in real life, that is: using the framework of the present invention, it is possible to effectively analyze and predict the severity of maritime traffic accidents, providing a basis for safety assessment and risk prevention and control research.

Appendix A. Characteristic description of the dataset of factors affecting marine accident risk

Table A. Description of 68 risk influencing factors

Appendix B. Prediction accuracy based on different feature selection methods Table B1. Prediction accuracy based on RF.

Table B2. Prediction accuracy based on SVM.

Table B3. Prediction accuracy based on NB.

Table B4. Prediction accuracy based on AdaBoost.

Table B5. Prediction accuracy based on XGBoost.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims (8)

1. A method for predicting the severity of a marine traffic accident, comprising: S1. Using marine accident investigation reports, construct a dataset of factors affecting marine accident risks; S2. Based on the constructed dataset of factors affecting marine accident risks, feature selection methods are used to train the accuracy of machine learning models and the interpretability of feature selection; S3, using a three-stage performance evaluation method of stability evaluation, prediction performance evaluation, comprehensive evaluation and statistical test to evaluate the performance of the feature selection method; S4. Six machine learning models were used for comparison to measure the performance of different predictors, and the machine learning model with the best performance in predicting the severity of ship accidents was used as the benchmark model; S5. Use the selected model with the highest prediction performance and the best features to predict the severity of the accident, conduct benefit evaluation, and analyze the effectiveness of risk control measures from a quantitative perspective. 2. A method for predicting the severity of a marine traffic accident according to claim 1, characterized in that step S1 specifically comprises: S11. Extract human, ship, environment, management factors and basic accident information from the accident investigation report as standard-level features of the marine accident risk influencing factor dataset; S12, classifying standard-level features into 68 index-level features during the data processing stage; S13. Convert continuous data and character data into discrete categorical data. 3. The method for predicting the severity of a marine traffic accident according to claim 1, wherein step S2 specifically comprises: S21. Feature selection method based on association rule mining and three-weight ranking algorithm to mine the interactions between accident features affecting marine accidents and rank them; S22. Adopting the feature selection method based on the new mutual information entropy (NMIE) guided by accident consequences, the influencing factors are ranked by using the information entropy of each influencing factor and the severity of the accident; S23. Using the feature selection method adopted in step S21, the relationship between each feature is mined and the features are sorted. Then, using the feature selection method adopted in step S22, the mutual information entropy between the features not mined by the feature selection method adopted in step S22 and the target is calculated, and the mutual information entropy between the features and the target is sorted. The two sorting results are combined to obtain the final feature sorting result. 4. The method for predicting the severity of a marine traffic accident according to claim 3, wherein step S21 specifically comprises: S211. Use association rule mining technology to find the association, co-occurrence or causal relationship between accident features: In order to accurately and effectively find the relationship between accident features, the FP-Growth algorithm is used to mine frequent item sets, and the association rules are generated by constructing an FP tree to screen out important association rules; if an association rule meets the minimum support and confidence, and the lift is greater than 1, then the association rule is considered valid. Valid association rules are used The calculation formula is as follows: Among them, Supp(X) represents the support of transaction X; express confidence level; express The degree of improvement; N _X represents the number of occurrences of X transaction in the project set; N represents the number of transactions in the project set; S212. Use graph theory methods to map the association rules between different accident characteristics to the complex influencing factor interaction network; the nodes in the complex influencing factor interaction network represent accident characteristics, and the edges represent the associations between them; the weights of the edges correspond to the confidence of the association rules; the complex influencing factor interaction network reflects the interaction of various accident characteristics in the development process of marine accidents; use the feature selection method based on association rule mining and three-weight ranking algorithm to determine the importance of each accident feature in the complex influencing factor interaction network, and rank the accident features accordingly. 5. A method for predicting the severity of marine traffic accidents according to claim 3, characterized in that the feature selection method based on association rule mining and three-weight ranking algorithm adopted in step S21 includes two optimization improvements, which are as follows: The first improvement: The concept of a weighted transition probability matrix is proposed, and confidence is used as the transition probability between adjacent risk influencing factors to replace the definition in the original algorithm. The calculation formula is as follows: Among them, PM represents the transition probability matrix; n represents the number of influencing factors mined by the FP-Growth algorithm; g represents the Ground node added to the original network; δ _ij represents the transition probability from node _ni to node _nj ; i→j represents the existence of an edge from node _ni to node _nj ; represents the in-degree of node n _j ; e _ji represents the association between influencing factor n _j and influencing factor n _i ; t represents the number of iterations; represents the score obtained by node n _i at the tth iteration; (matrix) _ij represents an n+1-order square matrix with only the i-th row and j-th column elements being 1 and the rest being 0; (zeros-one) _j represents a 1×(n+1)-order matrix with only the j-th column elements being 1 and the rest being 0; (ones) represents an (n+1)×1-order matrix with all elements being 1; * represents the Hadamard product; · represents the matrix product; The second improvement: Based on the theory of node external strength and node centrality coefficient in graph theory, a new comprehensive weight is proposed. After the algorithm iteration is completed, the comprehensive weight is used to obtain additional scores from the final score of the Ground node. The calculation formula is as follows: in, represents the output intensity of influencing factor n _i ; CC _i represents the betweenness central coefficient of influencing factor n _i ; σ _ij (v) represents the number of shortest paths from influencing factor n _i to influencing factor n _j through influencing factor n _v ; σ _ij represents the number of shortest paths from influencing factor n _i to influencing factor n _j ; μ _i represents the comprehensive weight proposed in this study; t _end represents the number of iterations at the end of the algorithm iteration. 6. A method for predicting the severity of a marine traffic accident according to claim 3, characterized in that, in step S22, a feature selection method based on a new mutual information entropy (NMIE) guided by accident consequences is used, and the calculation formula of the mutual information entropy is as follows: Among them, NMIE(X,Y) represents the mutual information entropy between feature X and label Y; m represents the number of states of the label; n represents the number of states of feature X; I( _xi , _yk ) represents the mutual information between state i of feature X and state k of the label. 7. A method for predicting the severity of a marine traffic accident according to claim 1, characterized in that step S3 specifically comprises: S31. Stability evaluation: The stability of the feature selection method is measured by considering the influence of randomness and changes in the size of the training set. Methods similar to k-folding, random numbers, and repeated experiments are used to determine the influence of randomness and changes in the size of the training set on the feature selection process. Spearman's rho is used to measure the stability of the ranking results. The calculation formula is as follows: Where Sprr ^R1,R2 represents the correlation between ranking R1 and ranking R2; R1 is used as the reference ranking; R2 is used as the experimental ranking; N represents the number of main features; M represents all features included in the main features of R1; if feature i is not in the main features of R2, then R2 _i = N+1; S32. Prediction performance evaluation: Train the prediction model through the risk influencing factors selected by features, measure the accuracy of the prediction model, find the subset of risk influencing factors that achieves the highest accuracy with the least risk influencing factors, and use the corresponding indicators to evaluate the prediction performance of the model; S33. Comprehensive evaluation and statistical test: The stability of the feature selection process is evaluated by determining whether the feature selection method has an acceptable level of stability; if the best stability of the feature selection method is lower than 0.6, or the difference with the feature selection method with the best stability is greater than 0.3, the stability of the feature selection method is considered to be unacceptable; unacceptable stability indicates that the feature selection method is invalid; based on the prediction performance evaluation information, the prediction performance of the feature selection method with acceptable stability is analyzed; if the prediction performance indicators of different feature selection methods are not similar, then the method with better prediction performance indicators is considered to be the better performing method; if the prediction performance indicators are the same, a 5-fold cross-validation is performed to test the difference in the prediction performance of the machine learning model; if the test result is significant greater than 0.05, there is no significant difference in the prediction performance of the two methods; if there is similarity in the prediction performance of different feature selection methods, the feature selection method with better stability is considered to have better performance. 8. The method for predicting the severity of a marine traffic accident according to claim 1, wherein step S4 specifically comprises: S41, introduce six machine learning models: support vector machine, naive Bayes classifier, random forest, AdaBoost, XGBoost, and LightGBM; S42. Accuracy, precision, recall, F1 value and area under the curve (AUC) were used to evaluate the prediction performance of the machine learning model, and the optical gradient enhancement machine with the best prediction performance for ship accident severity prediction was taken as the benchmark model.