Maritime accident prediction method and device based on interpretable ensemble machine learning

Abstract

The application discloses a maritime accident prediction method and device based on an interpretable integrated machine learning, relates to the technical field of maritime safety risk analysis, and comprises the following steps: acquiring accident investigation data and preprocessing, balancing the data through a ten-fold stratified oversampling method and training through cross-validation, determining a performance-optimal model by using a test set and performing interpretable analysis, and explaining the influence of features on accident prediction results. Through the construction of a closed-loop "data processing-model optimization-explanation output" process, the SMOTE oversampling, the ten-fold stratified cross-validation training and the heterogeneous base model integrated learning strategy are adopted, the processing capacity for the data imbalance problem of the accident category is improved, the data leakage problem of oversampling is avoided, and the possible bias of a single model is overcome. The interpretable analysis of the model prediction result can quantitatively show the contribution of each feature to the prediction from the global and local aspects, reveals the nonlinear relationship and interaction effect among the features, and provides a transparent explanation for the model decision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of maritime safety risk analysis, and in particular to a maritime accident prediction method and device based on an interpretable ensemble machine learning. BACKGROUND

[0002] Maritime accident analysis is an important work in the field of maritime safety management, and provides important data reference for cause analysis of maritime accidents, operation guidance to avoid maritime accidents, and even safety navigation specifications. However, maritime accidents are caused by the variability of the marine environment and the complexity of the causes. Different types and sizes of ships are prone to cause large differences in factors under different marine environments, for example, large ships have strong resistance to wind and waves, but have high operation risk, and vice versa, small ships have poor resistance to wind and waves, but have high operation flexibility, resulting in significant differences between the cause analysis directions of maritime accidents.

[0003] Maritime safety accident analysis currently mainly relies on qualitative analysis or quantitative methods based on classical statistical models, such as Bayesian networks, fault tree analysis, etc. However, traditional methods have difficulty in accurately capturing the complex mechanisms of accident causes when dealing with the interaction between multiple factors and complex nonlinear relationships, especially in the face of large differences in the causes of maritime accidents. Even relying on machine learning models also has the defects of insufficient generalization ability and robustness. Moreover, the "black box" characteristics of machine learning models make it difficult to provide interpretability of the decision-making process and prediction results, limiting their application in actual maritime safety management. SUMMARY

[0004] The embodiments of the present application provide a maritime accident prediction method and device based on an interpretable ensemble machine learning to solve the technical problems of data class imbalance and uninterpretability caused by large differences in causes in maritime accident prediction.

[0005] In a first aspect, the embodiments of the present application provide a maritime accident prediction method based on an interpretable ensemble machine learning, comprising:

[0006] S101, acquiring multi-source maritime accident investigation data and performing data preprocessing, and then dividing into a training set and a test set;

[0007] S102, extracting key cause features by fusing feature importance methods on the training set to form a key feature training set;

[0008] S103, ten-fold stratification is performed on the key feature training set to form ten training subsets, ten-fold cross-validation training is performed on the plurality of base models and at least one ensemble learning model composed of the plurality of base models by using the ten training subsets, and SMOTE oversampling method is embedded in the training process to eliminate data imbalance of the accident severity and the accident type;

[0009] S104, the trained plurality of base models and at least one ensemble learning model are respectively subjected to offshore accident prediction model performance evaluation by using a test set, and the optimal model is determined according to the evaluation result;

[0010] S105, the optimal model is subjected to interpretable analysis by using SHAP method, SHAP values of each feature in each sample of the test set are calculated, then mean values of SHAP absolute values of each feature on all samples are calculated and are sorted in descending order to form a global feature importance ranking, and a feature dependency graph, a feature interaction dependency graph and an individual prediction explanation graph are drawn to respectively analyze interaction coupling effects between any two features and influences of key cause features on the prediction result in a high-risk accident.

[0011] In a second aspect, an embodiment of the present application provides an offshore accident prediction device based on an interpretable ensemble machine learning, comprising:

[0012] A data preprocessing module is configured to acquire and preprocess multi-source maritime accident investigation data.

[0013] A data balancing module is configured to eliminate data imbalance of accident types by using ten-fold stratification oversampling method on a training set.

[0014] A model training module is configured to perform ten-fold cross-validation training on a plurality of base models and at least one ensemble learning model composed of the plurality of base models.

[0015] A model performance evaluation module is configured to perform offshore accident prediction model performance evaluation on the trained model by using a test set to determine an optimal model.

[0016] An interpretable analysis module is configured to perform interpretable analysis on the optimal model by using SHAP method, to calculate SHAP values of each feature in each sample of the test set, then to calculate mean values of SHAP absolute values of each feature on all samples and sort them in descending order to form a global feature importance ranking, and to draw a feature dependency graph and an individual prediction explanation graph.

[0017] In a third aspect, an embodiment of the present application provides an electronic device, comprising:

[0018] One or more processors;

[0019] a storage device storing one or more programs,

[0020] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned offshore accident prediction method based on interpretable ensemble machine learning.

[0021] In a fourth aspect, an embodiment of the present application provides a storage medium containing computer executable instructions for executing the above-mentioned offshore accident prediction method based on interpretable ensemble machine learning when executed by a computer processor.

[0022] The offshore accident prediction method and device based on interpretable ensemble machine learning provided by the embodiments of the present application construct a high-quality multi-source structured data set by preprocessing data and divide the training set and the test set. The ten-fold stratified oversampling method is used to eliminate the data imbalance of the accident type for the training set, and then used for the training of the heterogeneous base model and the ensemble learning model. After the model is trained, the performance of each heterogeneous base model and the ensemble learning model is tested by using the test set, and the performance optimal model is selected for interpretable analysis. By constructing a closed loop of "data preprocessing feature selection model training interpretable analysis" process, using the training scheme of SMOTE oversampling combined with ten-fold stratified cross-validation, and the heterogeneous base model ensemble learning strategy, the processing effect of the class imbalance problem is significantly improved. The ten-fold stratified oversampling method can also reduce the data leakage problem common in oversampling, overcome the possible bias of a single model, and has better performance than a single base model in the offshore accident severity and type prediction task, ensuring that the model performance does not produce false images due to oversampling, and the evaluation results truly reflect the performance of the model in the actual unbalanced environment. The SHAP method is used for interpretable analysis of the prediction results of the ensemble model, which can quantitatively show the contribution of each feature to the prediction from the global and local levels, reveal the nonlinear relationship and interaction effect between the features, provide transparent explanation of the model decision, enhance the credibility of the results, and provide clear risk decision basis for maritime safety management personnel. BRIEF DESCRIPTION OF DRAWINGS

[0023] The accompanying drawings, which form a part of the present application, are intended to provide further understanding of the present application and are incorporated herein for a purpose of explanations and are not intended as improper limitations on the present application. In the drawings:

[0024] Figure 1 A flowchart of an offshore accident prediction method based on interpretable ensemble machine learning according to the first embodiment of the present application;

[0025] Figure 2 A feature dependency graph according to the first embodiment of the present application;

[0026] Figure 3 a feature interaction dependency graph according to the embodiment one of the present application;

[0027] Figure 4 an individual explanation graph of the accident severity prediction task according to the embodiment one of the present application;

[0028] Figure 5 an individual explanation graph of the accident type prediction task according to the embodiment one of the present application;

[0029] Figure 6 a flow chart of a maritime accident prediction method based on explainable ensemble machine learning according to the embodiment two of the present application;

[0030] Figure 7 a structural schematic diagram of a maritime accident prediction device based on explainable ensemble machine learning according to the embodiment three of the present application;

[0031] Figure 8 a structural diagram of an electronic device according to the embodiment four of the present application. DETAILED DESCRIPTION

[0032] The present application will be further described below in conjunction with the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application. In addition, it should be noted that, for the convenience of description, only the parts related to the present application are shown in the drawings, but not all the structures.

[0033] Embodiment one

[0034] Figure 1 a flow chart of a maritime accident prediction method based on explainable ensemble machine learning according to the embodiment one of the present application, specifically comprising the following steps:

[0035] S101, acquiring multi-source maritime accident investigation data and performing data preprocessing, and then dividing into a training set and a test set.

[0036] The multi-source maritime accident investigation data is acquired from multi-national maritime safety management agencies, and the collected data is preprocessed by removing duplicates and integrity screening, etc. to eliminate data lacking key causation elements, and is arranged into a structured database, with the feature variables in the maritime accident investigation data as the fields (table headers) of the structured database. Exemplarily, the feature variables include total tonnage of the ship, wind power, visibility, operation, lookout, time of occurrence, sea conditions, risk assessment, etc., covering the feature description from the perspectives of accident information, ship factors, environmental factors, human factors and management factors, etc. Then, the preprocessed structured data is divided into a training set and a test set which can be used for machine learning model training and performance evaluation according to a proportion.

[0037] Specifically, multi-source maritime accident investigation data is acquired, and de-duplication and integrity screening is performed to obtain a structured data set.

[0038] Multi-source maritime accident investigation data is acquired from maritime accident databases of maritime safety management agencies in multiple countries, and the acquired data is de-duplicated. To address the problem of repeated records between different databases, a regularized de-duplication method based on key field matching is used for data de-duplication. For example, the core fields of accident occurrence time, ship identification information (such as ship name / tonnage), accident location, and accident type are compared for consistency. When highly similar records are found, only the most complete information is retained, thereby ensuring the uniqueness of the sample and eliminating duplicate data. In addition, data integrity screening is performed based on a missing rate threshold to eliminate data that lacks key causal elements (such as more than 30% of key fields such as accident type, severity, and environmental conditions missing). This ensures the uniqueness and completeness of the data, and the uniformity and completeness of the features of the sample during subsequent machine model analysis. Finally, a high-quality structured data set is formed.

[0039] Feature variable extraction is performed on the structured data set, and missing values are filled to form an original data set. The original data set is divided into a training set and a test set according to stratified sampling.

[0040] The feature variables are extracted from the structured data set, and exemplary feature variables include ship gross tonnage, wind, visibility, operation, lookout, occurrence time, sea state, risk assessment, and other original independent variables. Among them, the occurrence time (day / night), operation (yes / no), lookout (yes / no), weather, visibility, sea state, ship type, regulations, risk assessment, and other dimensions of the features are used as classification variables for classifying the accident type; the wind and ship gross tonnage are used as continuous numerical variables. When predicting using the feature variables, two classification tasks can be performed. One is to predict the severity of the accident, and the target variable output under this task is the accident severity level, which is divided into three levels (not serious, serious, and very serious) according to the International Maritime Organization (IMO) standard. The other is to classify the accident, and the target variable output under this task is the accident type, which is divided into six types (collision, contact, grounding, fire and explosion, sinking, and others) according to the multi-country maritime accident report. When filling in the missing values, for the classification variables in the feature variables, the mode is used to fill in the missing values, and One-Hot encoding is performed; for the continuous variables in the feature variables, the median is used to fill in the missing values, and Z-score standardization is performed to form the original data set. Then, according to the stratified sampling method, the data set is divided into a training set and a test set in a ratio of 8:2, and it is necessary to ensure that the proportion of each category after division is consistent with the original data set to maintain the representativeness of the data distribution. To ensure the complete repeatability of the experiment, the random state (Random State) can also be set to 42 to control all random processes involved in data division, subsequent model training, random initialization, Bootstrap sampling, etc., so that the results are completely consistent every time the experiment is run.

[0041] S102, extracting key cause features from the training set by fusing feature importance methods to form a key feature training set.

[0042] Before model training, key cause features can also be selected from all features to improve the expression ability of important features and the prediction performance of the model, while reducing the feature dimension and reducing computational consumption. Exemplarily, Pearson Correlation Coefficient, Maximum Information Coefficient (MIC), Random Forest Importance, Mean |SHAP Value|, and other indicators can be used to comprehensively filter, embed, and post-explain the features from four dimensions of linear correlation, nonlinear correlation, model intrinsic importance, and prediction contribution for comprehensive quantitative evaluation and screening. The features with high importance ranking are selected as key cause features and the remaining features are removed for training of the base model and the ensemble learning model to form a key feature training set.

[0043] S103, stratify the key feature training set into ten folds, form ten training subsets, use the ten training subsets to perform ten-fold cross-validation training on multiple base models and at least one ensemble learning model formed by the multiple base models, and embed SMOTE oversampling method in the training process to eliminate the imbalance of accident severity and accident type data.

[0044] The training set is divided into ten subsets of similar size according to the stratified sampling strategy, i.e. "folds", while ensuring that the proportion of each category (accident type or severity) in each subset is consistent with the overall original training set, obtaining ten stratified training subsets. Ten-fold stratified training subsets are mainly used for ten-fold cross-validation, that is, 9 folds are combined for training and the remaining one fold is used for validation, and 10 rounds of cross-validation are performed (i.e. each fold is used as a validation set for one round). According to the ten-fold cross-validation method, the base model and the ensemble learning model are trained respectively. In the training process of each fold, the SMOTE oversampling algorithm is applied to the "training subset" combined by the 9 folds used for training in the current round to eliminate the imbalance of the number of samples of accident severity and accident type. The base model is a plurality of independent machine learning models, which can be XGBoost, LightGBM, CatBoost, random forest, support vector machine, etc.

[0045] S104, use the test set to evaluate the performance of the trained multiple base models and at least one ensemble learning model for marine accident prediction model, and determine the optimal model according to the evaluation result.

[0046] After the base model and the ensemble learning model are trained, the test set divided in the initial stage is used to evaluate the performance of each model for marine accident prediction, and the evaluation indicators can include accuracy, precision, recall, F1 score, AUC-ROC value, etc. According to the evaluation result, the optimal model is selected for subsequent model deployment, use and interpretation analysis.

[0047] Specifically, the test set is used to test the trained multiple base models and at least one ensemble learning model, and the accuracy, precision, recall, F1 score and AUC-ROC curve parameters of each model are obtained.

[0048] The trained base model and ensemble model are tested respectively on the test set divided in the initial stage in an 8:2 ratio to maintain the original distribution, and the accuracy, precision, recall, F1 score and AUC-ROC curve parameters of each model are obtained. The test set is a data set that has not been involved in training and has not been subjected to SMOTE oversampling processing. All models use the same test set to ensure the objectivity and generalization ability of model evaluation, so that the model can both learn the characteristics of the minority class through balanced data and adapt to the reality of data imbalance in actual scenarios.

[0049] According to the obtained test results, the model with the best performance for predicting maritime accidents is selected as the optimal model.

[0050] The accuracy rate can provide an overview of the overall performance of the model. The precision rate, recall rate and F1 score are important for scenarios that focus on high-risk minority classes (such as "very serious" accidents). A high recall rate means that as many real high-risk events as possible can be captured. The AUC-ROC value measures the overall ability of the model to distinguish between positive and negative classes at different classification thresholds, and is robust to imbalanced data. Based on these indicators and previous experience (for example, when the recall rates are similar, prefer the model with a higher AUC), the optimal model is determined. For example, when performing one of the "accident severity prediction" tasks, the stacked ensemble learning model performs best, with an accuracy of 0.8970 and an AUC-ROC value of 0.8184, as shown in Table 1. When performing another "accident type prediction" task, the voting ensemble learning model performs best, with an F1 score of 0.8317 and an AUC-ROC value of 0.9076, as shown in Table 2.

[0051] Table 1 Performance of each model on the test set for the accident severity prediction task

[0052]

[0053] Table 2 Performance of each model on the test set for the accident type prediction task

[0054]

[0055] S105, using SHAP method to perform interpretable analysis on the optimal model, calculating the SHAP value of each feature in each sample of the test set, then calculating the mean of the absolute value of each feature on all samples and sorting in descending order to form a global feature importance ranking, and drawing feature dependency graphs, feature interaction dependency graphs and individual prediction explanation graphs to analyze the interaction coupling effect between any two features and the influence of key cause features on the prediction results in high-risk accidents.

[0056] To realize the transparency of model decision, after determining the trained optimal model, SHAP (Shapley Additive explanation) method is used for explainability analysis, the model prediction result is regarded as the income of "common game" of multiple features, and the income is fairly distributed to each feature, so as to quantify the marginal contribution of each feature to the prediction result. By forming global feature importance ranking, feature dependency graph, feature interaction dependency graph and individual prediction explanation graph, the overall importance of features and its influence on the prediction direction are analyzed at the global level, the nonlinear effect of single feature and the coupling mechanism between different features are revealed at the feature level, and the prediction basis of the model for a single accident sample is explained at the individual sample level.

[0057] First, for the optimal model obtained by training, SHAP (SHapley Additive exPlanations) algorithm is used to calculate the SHAP value of each feature in each sample of the test set, and the explainability of the model prediction result is analyzed. The value quantitatively represents the contribution of a single feature to the prediction result of a single sample; for multi-classification tasks, the SHAP value of each class is calculated. Then, based on the SHAP values of all samples, the mean of the absolute value of SHAP of each feature on all samples is calculated to measure the overall influence degree of each feature on all samples, and the global feature importance ranking is obtained by descending order to identify the key cause factors that have the greatest impact on the model prediction. For example, it is found that "total tonnage of ship" and "wind force" are the key features affecting the severity of the accident. According to the global feature importance ranking, the scatter plot of all samples is drawn with the feature name as the vertical coordinate and the SHAP value as the horizontal coordinate, forming a "feature dependency graph" (also known as SHAP summary plot), as shown in FIG. 2. Figure 2The SHAP values are shown to reflect the importance ranking of the features, the high and low changes of the feature values, and the positive and negative effects on the model prediction results: when the SHAP value is positive, it means that the feature has a positive driving effect on the prediction result of the target variable; when the SHAP value is negative, it means that it has an inhibitory effect on the prediction result. According to the feature dependence graph, the distribution density of the scatter points in different SHAP value intervals reflects the concentration degree of the samples at the corresponding feature contribution level. For example, “total tonnage” and “wind force” are among the top in the SHAP absolute value mean ranking, which are the key features affecting the prediction result of the accident severity. The scatter points of “total tonnage” are distributed to the negative value region as a whole, and the SHAP value corresponding to the low value sample is relatively high, indicating that the smaller the tonnage, the higher the SHAP value, that is, the low tonnage ship has a positive driving effect on the prediction of “accident severity” (more likely to cause serious accidents); while the “wind force” feature shows a relatively dispersed SHAP value distribution in different value intervals, indicating that its influence on the accident severity has certain complexity and uncertainty.

[0058] Then any two features can be combined in pairs to form multiple pairs of feature combinations, and a feature interaction dependence graph of each pair of two features can be drawn. By selecting feature A as the main feature and its original value as the horizontal axis, the vertical axis is still the SHAP value corresponding to feature A, and feature B is introduced as a color mapping variable to represent its value size (continuous feature) or category state (discrete feature) by color depth, where each scatter point represents a sample, as shown in Figure 3 The feature interaction dependence graph is drawn to study and analyze the coupling effect between features: according to color stratification and trend separation, if the scatter points or trend distribution of different colors (representing different levels of feature B) are obviously separated within the same feature A value interval, it means that feature B has an adjusting effect on feature A, that is, there is an interaction between the two; according to the coupling direction judgment, if the SHAP value corresponding to feature A significantly increases under adverse conditions (such as operation absence, strong wind, etc.), it means that the two features jointly promote the prediction result to develop towards high risk, which embodies the risk amplification effect; according to risk superposition and key scenario identification, by comparing the change trend of SHAP value under different feature combinations, the feature combination that leads to significant deterioration of model prediction can be identified, that is, the potential “risk coupling point”. For example, in the feature interaction dependence graph of “total tonnage – operation”, when the operation is in the absence state, the SHAP value corresponding to the low total tonnage is significantly higher than that in the normal operation scenario, indicating that human operation errors can increase the inherent risk of small tonnage ships, which embodies the typical “man-ship” coupling effect.

[0059] For the minority class samples (high-risk accidents such as very severe or sinking), individual prediction explanation graphs are drawn to visualize how each feature in the sample "pushes" the model's baseline prediction value to the final prediction result, i.e. the local explanation form of SHAP value (Force Plot or its equivalent waterfall visualization form), to explain the basis of the model's prediction for a single sample (such as how poor visibility and lack of lookout push the prediction result to the high-risk category), and further explain the reasons for the model's decision. For example, Figure 4 Figure 1 shows an individual explanation graph for the accident severity prediction task, and Figure 5 Figure 2 shows an individual explanation graph for the accident type prediction task. Red features in the figure represent positive pushing effects on the prediction result, usually corresponding to risk factors; blue features represent inhibitory effects on the prediction result, usually corresponding to protective factors; the length of the bar reflects the influence of the feature on the individual prediction result. The individual explanation graph is constructed based on the following prediction decomposition relationship:

[0060]

[0061] wherein, represents the final prediction output of the model for a given single sample , is the average prediction value of the model on the background data set (usually the training set or its subset), also known as the baseline prediction value, representing the expected output of the model without considering any specific feature information, is the number of input features; represents the marginal contribution value of the th feature to the prediction result of the sample, i.e. the incremental effect on the prediction result before and after considering the feature added to the model. The baseline prediction value represents the average prediction output of the model on the training samples; the SHAP values of the features are sorted by absolute value size, and are used as positive or negative "forces" to push the prediction result from the baseline value to the final output; then horizontal bars are used to represent the contribution direction and amplitude of each feature, with red indicating positive pushing of the prediction value and blue indicating negative pulling of the prediction value, to visualize the expression; the result of the superposition of all feature contributions is the final prediction value of the model for the sample. From Figure 4As can be seen, the final prediction value of sample #1 is significantly lower than the benchmark value, so the total tonnage of the ship makes a significant negative contribution in this sample, which is the dominant factor in reducing the prediction of accident severity, indicating that larger ship size has a significant buffering effect on the formation of accident consequences; the prediction result of sample #2 is significantly higher than the benchmark value, which belongs to high severity accident, and the wind characteristics become the main positive driving factor in the current context, while the lower total tonnage of the ship further amplifies the risk effect, the synergistic effect of the two significantly pushes up the prediction value, revealing how severe environmental conditions stimulate the inherent vulnerability of the ship, indicating a high-risk scenario of low tonnage and severe environment superposition; the prediction value of sample #3 is slightly higher than the benchmark value, which reflects the result of the game of multiple factors, among which the lack of lookout is the main risk driving factor, which makes a positive contribution to the prediction value, while wind and total tonnage do not behave as the main risk source in this specific context, and their SHAP contribution direction plays a certain inhibitory role in the increase of the prediction value, the joint action of positive and negative factors makes the model finally give a moderately high accident severity prediction, reflecting the effective characterization ability of the model to complex nonlinear interaction. Finally, maritime safety management personnel can understand the logic of model decision-making and identify the core risk coupling relationship, so as to develop more targeted risk prevention measures.

[0062] In this embodiment, high-quality multi-source structured data sets are constructed by preprocessing data, and training sets and test sets are divided. The ten-fold stratified oversampling method is used to eliminate the data imbalance of the accident type for the training set, and then used for the training of heterogeneous base models and ensemble learning models. After the model is trained, the performance of each heterogeneous base model and ensemble learning model is tested by using the test set, and the model with the best performance is selected for interpretable analysis. By constructing a closed-loop "data processing-model optimization-explanation output" process, using the training scheme of SMOTE oversampling combined with ten-fold stratified cross-validation, and the heterogeneous base model ensemble learning strategy, the processing effect of the class imbalance problem is significantly improved. The ten-fold stratified oversampling method can also prevent the data leakage problem caused by oversampling, and overcome the possible bias of a single model. In the task of predicting the severity and type of maritime accidents, the performance is better than that of a single base model, ensuring that the model performance does not produce a false image due to oversampling, and the evaluation results truly reflect the performance of the model in the actual unbalanced environment. The SHAP method is used to analyze the interpretability of the prediction results of the integrated model, which can quantitatively show the contribution of each feature to the prediction from the global and local levels, and reveal the nonlinear relationship and interaction effect between the features, providing transparent explanation of the model decision-making, enhancing the credibility of the results, and providing clear risk decision-making basis for maritime safety management personnel.

[0063] Embodiment Two

[0064] Figure 6The flowchart of the offshore accident prediction method based on the interpretable ensemble machine learning according to Embodiment Two of the present application is optimized based on the above-mentioned embodiment. In this embodiment, S103 is specifically optimized as follows:

[0065] The key feature training set is divided into ten subsets of similar size according to the stratified sampling strategy, to obtain ten training subsets.

[0066] The ten training subsets are used to perform ten-fold cross-validation training on the plurality of base models and at least one ensemble learning model composed of the plurality of base models.

[0067] During the ten-fold cross-validation training, the nine training subsets used in each round of training are processed by the SMOTE oversampling method to interpolate the feature space of the minority class samples to eliminate data imbalance, and the remaining one training subset is used as the validation set in each round and is not processed.

[0068] Correspondingly, the offshore accident prediction method based on the interpretable ensemble machine learning provided in this embodiment specifically includes:

[0069] S201, acquiring multi-source maritime accident investigation data and performing data preprocessing, and then dividing the data into a training set and a test set.

[0070] S202, extracting key cause features from the training set by using a fusion feature importance method to form a key feature training set.

[0071] Specifically, the training set is evaluated by using a fusion feature importance method, and key cause features are selected by using Pearson correlation coefficient, maximum information coefficient, random forest feature importance index, and SHAP value importance index, respectively, and the remaining features are removed to obtain the key feature training set.

[0072] By fusing feature importance evaluation, the key cause feature with relatively important influence on the prediction result is screened. For each candidate feature, four importance indicators, including Pearson correlation coefficient (measuring the linear correlation between the feature and the target variable, with a value range of [-1, 1], the greater the absolute value, the stronger the linear correlation, reflecting the linear direction and intensity), maximum information coefficient (capturing nonlinear correlation), random forest feature importance (calculating the average impurity reduction of feature splitting nodes in all decision trees to evaluate the built-in importance, which can automatically handle complex interactions and nonlinear patterns between features, reflecting the intrinsic contribution of features in the model construction process), and SHAP value importance (a post-explanation method based on game theory, which quantitatively measures the contribution of a single feature to the model's prediction output, and can fairly allocate the contribution value of each feature to a single prediction result), are used to comprehensively depict the feature importance and avoid misjudgment caused by a single method. According to the comprehensive importance of the features indicated by the four importance indicators, the top-ranked features are selected as the key cause features, and the remaining features are removed to form a key feature training set for subsequent training of the base model and the ensemble learning model.

[0073] In an optional implementation of the embodiment, the Pearson correlation coefficient, the maximum information coefficient, the random forest feature importance indicator, and the SHAP value importance indicator of each dimension feature are calculated according to the training set. The four indicators are in the form of mean value. Then, the mean value of the four indicators of each feature is taken to obtain the importance comprehensive score of each feature.

[0074] First, the four importance indicators of each feature are independently calculated. Then, the average value of the same indicator value of each feature is taken to obtain four stable indicator values of each feature, i.e. (feature dimension x 4) indicators, reflecting the importance stability of the feature in the training set, so as to more reliably estimate the real contribution. Then, the Min-Max method is used to standardize the four stable indicator values of each feature, which maps the indicator values with different dimensions to a unified [0, 1] interval, avoiding bias caused by value scale difference. The arithmetic mean of the four normalized indicators of each feature is calculated, i.e. the arithmetic mean of the four normalized indicators of each feature is taken to form an importance comprehensive score corresponding to each feature.

[0075] All features are sorted in descending order according to the importance comprehensive score, and the top-ranked features are selected as the key cause features.

[0076] According to the importance comprehensive score in descending order, the top-ranked features are selected, and the higher the comprehensive score, the stronger the importance of the feature under various evaluation methods. For example, the top 6 features are selected as the key cause features, and the other features except the 6 key cause features are removed to form ten key cause feature training subsets corresponding to the original ten-fold, which integrates multiple evaluation perspectives and significantly improves the robustness of feature selection. For example, for the accident severity prediction task, the selected key cause features include total tonnage of the ship, wind, visibility, operation, lookout, and occurrence time; for the accident type prediction task, the selected key cause features include wind, total tonnage, sea state, lookout, risk assessment, and occurrence time, which is highly consistent with the experience of maritime experts.

[0077] S203, dividing the key feature training set into ten subsets of similar size according to the stratified sampling strategy to obtain ten training subsets.

[0078] The training set is divided into ten subsets of similar size according to the stratified sampling strategy, while ensuring that the proportion of each category (accident type or severity) in each subset is consistent with the overall proportion of the original training set, to obtain ten stratified training subsets.

[0079] S204, using the training subsets to perform ten-fold cross-validation training on multiple base models and at least one ensemble learning model composed of multiple base models. Optionally, the ensemble learning model includes a voting ensemble model and a stacking ensemble model.

[0080] The ten key causal characteristics obtained by screening are used to train subsets of base models and ensemble learning models respectively through ten-fold cross-validation. For example, the plurality of heterogeneous base models include XGBoost, LightGBM and CatBoost models based on gradient boosting, random forest models based on Bagging, and support vector machine models based on kernel methods. The ensemble learning models include a voting ensemble model and a stacking ensemble model (both composed of the five base models) composed of all the above base models. The voting ensemble model uses a soft voting mechanism to average the class probabilities of the outputs of the plurality of heterogeneous base models, and finally selects the class with the highest average probability as the integrated prediction result, which is simple and efficient. The first layer of the stacking ensemble model is also the plurality of heterogeneous base models, which use the remaining nine-fold samples as training subsets to train each base model, and the trained base models are used to predict the single validation fold sample that does not participate in training, and output the corresponding class prediction probability value; the second layer uses the output results of the first layer as input features to form a meta-feature matrix (i.e. Out-of-Fold prediction probability) after being sorted by sample order, and uses a multinomial logistic regression model as a meta-model to train the meta-feature matrix, which can achieve a higher performance upper limit. In this embodiment, ten-fold stratification is used, and a higher number of folds means that the size of each sub-training fold is closer to the complete training set, so that the performance of the plurality of heterogeneous base models on each sub-validation fold is closer to their performance on the full set, thereby providing the meta-model with input features that are less biased and more reliable.

[0081] In S205, during the ten-fold cross-validation training process, the nine training subsets used in each round of training are processed by the SMOTE oversampling method to interpolate the feature space of the minority class samples to eliminate data imbalance, and the remaining one training subset is used as the validation set for each round without processing.

[0082] The maritime accident data naturally has a serious imbalance problem of accident types, for example, the number of "collision" accident samples is much more than that of "explosion" or "sinking" accident. If the model is directly trained on the data set with imbalance between accident severity and accident type, it will often lead to overfitting of the majority class by the model, and the model is more inclined to predict the class with more samples, while the identification ability of the high-risk minority class is poor, resulting in systematic bias in the prediction result. Therefore, during the ten-fold cross-validation training process, the SMOTE oversampling algorithm is applied to the nine training subsets used in this round of training every time, and new samples are synthesized for the minority class samples in the merged training subsets, so that the number of samples of each class reaches a basic balance, which is used for model training in this round of cross-validation. Specifically, for the minority class samples in the "training subset", new samples are synthesized by randomly selecting their neighboring samples in the feature space and performing linear interpolation, so that the number of samples of each class in the "training subset" reaches a basic balance, so as to eliminate the data imbalance of accident types. For example, during the subsequent ten-fold cross-validation, when performing one kind of "accident severity prediction" task, k-neighbors is set to 5; when performing another kind of "accident type prediction" task, k-neighbors is set to 3. While the one fold left for validation in the current round strictly maintains its original unbalanced distribution and does not perform any oversampling. This process is independently carried out for ten rounds, and finally the trained base model and ensemble learning model are obtained. It can effectively avoid the leakage of synthesized sample information from the training phase to the validation phase, ensure the authenticity of model evaluation, improve the identification ability of the model for the minority class accidents, and effectively alleviate the bias problem caused by class imbalance.

[0083] Optionally, during the model training phase, a random search strategy is used to optimize the hyperparameters in the preset hyperparameter space, the performance of different hyperparameter combinations is evaluated through cross-validation, and the hyperparameter combination with the best performance is selected for model training.

[0084] In the model training stage, the performance of different hyperparameter combinations can be evaluated by using ten-fold cross-validation training method. Hyperparameter optimization can be performed in the pre-set hyperparameter space by using a random search strategy (e.g., 50 rounds of random search using RandomizedSearchCV), and the hyperparameter combination with the best performance is selected for model training to obtain the target model parameter instance trained by the optimal hyperparameter combination. The pre-set hyperparameter space is determined by considering the characteristics of the algorithm itself, existing literature experience, and the size and complexity of the data set to determine its reasonable value range. The recommended default parameters of the library (such as scikit-learn, XGBoost, LightGBM) are taken as the center, and they are reasonably expanded by 1-2 orders of magnitude above and below. The data-driven constraints are: small sample size → control model depth and learning rate, moderate feature dimension → avoid excessive weak learner stacking. Random search can more efficiently explore high-dimensional hyperparameter space under limited computing budget. Under the same computing cost, random search is more likely to find hyperparameter combinations with performance close to the optimal, especially for models with high parameter dimension or large parameter sensitivity difference. A dual synergy strategy combining algorithm level and data level can also be used to solve the data imbalance problem of accident categories. At the algorithm level, for models that support class weight (such as LightGBM, random forest, SVM), the class_weight='balanced' parameter can be configured to modify the loss function and force the model to pay attention to the minority class during training. For the CatBoost model, the auto_class_weights='Balanced' option can be used. At the data level, SMOTE oversampling can be used in the above ten-fold cross-validation training process with k-neighbors=5 for accident severity prediction and k-neighbors=3 for accident type prediction to directly expand the decision boundary of the minority class in the feature space. Algorithm level and data level can synergistically address the class imbalance problem from different angles to achieve better generalization performance.

[0085] S206, using the test set to evaluate the performance of the trained multiple base models and at least one ensemble learning model for maritime accident prediction model, and determining the optimal model according to the evaluation results.

[0086] S207, using SHAP method to perform explainable analysis on the optimal model, calculating the SHAP value of each feature in each sample of the test set, then calculating the mean of the SHAP absolute value of each feature on all samples and sorting in descending order to form a global feature importance ranking, and drawing feature dependency graphs, feature interaction dependency graphs and individual prediction explanation graphs to analyze the interaction coupling effect between any two features and the influence of key cause features on the prediction results in high-risk accidents.

[0087] The embodiment filters the feature importance by the multi-index method fusion of Pearson coefficient, MIC, random forest importance and SHAP value, comprehensively filters the features from multiple dimensions of linearity, nonlinearity, model built-in and prediction contribution, avoids the limitation of single method, effectively identifies the key cause factors affecting different prediction tasks, and combines the heterogeneous ensemble learning strategy to show better performance than single base model in different prediction tasks. The training link of ten-fold stratified cross-validation adopts random search to optimize the hyperparameters, and the "double synergy" strategy of class weight configuration at the algorithm level and SMOTE oversampling at the data level is combined to process the class imbalance, forming a double guarantee mechanism from the optimization target to the input data and the learning ability of the model to the minority class samples, effectively alleviating the prediction bias problem of the model caused by the natural imbalance of the maritime accident data.

[0088] Embodiment three

[0089] Figure 7 A structure diagram of a maritime accident prediction device based on an interpretable ensemble machine learning according to Embodiment three of the present application, in the embodiment, the maritime accident prediction device based on the interpretable ensemble machine learning comprises:

[0090] A data preprocessing module 810 is configured to acquire and preprocess multi-source maritime accident investigation data.

[0091] A data balancing module 820 is configured to eliminate the data imbalance of the accident type by ten-fold stratified oversampling method on the training set.

[0092] A model training module 830 is configured to perform ten-fold cross-validation training on a plurality of base models and at least one ensemble learning model composed of the plurality of base models.

[0093] A model performance evaluation module 840 is configured to perform maritime accident prediction model performance evaluation on the trained model by using a test set to determine an optimal model.

[0094] An interpretable analysis module 850 is configured to perform interpretable analysis on the optimal model by using SHAP method, calculate the SHAP value of each feature in each sample of the test set, then calculate the mean of the SHAP absolute value of each feature on all samples and sort in descending order to form a global feature importance ranking, and draw a feature dependence graph and an individual prediction explanation graph.

[0095] The embodiment preprocesses the obtained multi-source maritime accident investigation data through a data preprocessing module, balances the data of the accident types through a data balancing module, trains the base model and the ensemble learning model through ten-fold cross-validation training of a model training module, evaluates the performance of the trained model through a test set by a model performance evaluation module, and performs explainable analysis on the optimal model through an explainable analysis module. By constructing a closed-loop "data processing-model optimization-explanation output" process, adopting a training scheme combining SMOTE oversampling and ten-fold stratified cross-validation, and a heterogeneous base model ensemble learning strategy, the processing effect of the class imbalance problem is significantly improved. The ten-fold stratified oversampling method can also eliminate the data leakage problem common in oversampling, overcome the possible bias of a single model, and outperform a single base model in the prediction task of maritime accident severity and type, ensuring that the model performance does not produce a false impression due to oversampling, and the evaluation results truly reflect the performance of the model in the actual unbalanced environment. The SHAP method is used to perform explainable analysis on the prediction results of the integrated model, which can quantitatively show the contribution of each feature to the prediction from the global and local levels, reveal the nonlinear relationship and interaction effect between the features, provide transparent explanation of the model decision, and enhance the credibility of the results, thereby providing clear risk decision basis for maritime safety management personnel.

[0096] The maritime accident prediction device based on the explainable ensemble machine learning provided by the embodiment of the present application can execute the maritime accident prediction method based on the explainable ensemble machine learning provided by any embodiment of the present application, and has the corresponding function modules and beneficial effects of the execution method.

[0097] Embodiment four

[0098] Figure 8 A structural diagram of an electronic device according to Embodiment Four of the present application, Figure 8 A block diagram of an exemplary electronic device 12 suitable for use in implementing embodiments of the present application is shown. Figure 8 The electronic device 12 shown is merely one example and should not be taken as limiting the functionality or use of embodiments of the present application.

[0099] As shown in Figure 8 The electronic device 12 is in the form of a general computing device. Components of the electronic device 12 can include, but are not limited to, one or more processors or processing units 16, system memory 28, and a bus 18 that connects the various system components, including the system memory 28 and the processing unit 16.

[0100] Bus 18 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration bus, a processor or local bus using any of a variety of bus architectures. By way of example, these architectures include Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, Enhanced ISA (EISA) bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus.

[0101] Electronic device 12 typically includes a variety of computer system readable media. These media can be any available media that is accessible by electronic device 12 and includes both volatile and non-volatile media, removable and non-removable media.

[0102] System memory 28 can include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache memory 32. Electronic device 12 can further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 34 can be provided for reading from and writing to a non-removable, non-volatile magnetic media (e.g., a "hard drive"). Figure 8 not shown, a magnetic hard disk drive for reading from and writing to non-removable, non-volatile magnetic media (e.g., a "hard drive"). Although not specifically shown, such Figure 8 In alternative embodiments, a magnetic hard disk drive, a solid state drive (SSD) which is a non- volatile storage medium, a floppy disk drive for reading from and writing to a removable, non-volatile magnetic disk (e.g., a "floppy disk"), and an optical disk drive for reading from or writing to a removable, non-volatile optical disk (e.g., a CD-ROM, DVD-ROM, etc.) can be provided. In such instances, each can be connected to bus 18 by one or more data media interfaces. System memory 28 can include at least one program product having a set (e.g., at least one) of program modules that are configured to carry out the functions of embodiments of the application.

[0103] Program / utility 40, having a set (at least one) of program modules 42, can be stored in system memory 28 by way of example, such as an operating system, one or more application programs, other program modules, and program data, each or some combination thereof, which may

[0104] The electronic device 12 can also be in communication with one or more external devices 14 such as a keyboard, a pointing device, a display 24, etc.; can also be in communication with one or more devices that enable a user to interact with the electronic device 12 / server / computer; and / or can be in communication with any devices (such as a network card, a modem, etc.) that enable the electronic device 12 to communicate with one or more other computing devices. Such communication can be facilitated by an Input / Output (I / O) interface 22. Further, the electronic device 12 can communicate with one or more networks (such as a local area network (LAN), a wide area network (WAN), and / or a public network such as the Internet) through a network adapter 20. As Figure 8 illustrated, the network adapter 20 is in communication with the other modules of the electronic device 12 through the bus 18. As will be appreciated, although not shown, other hardware and / or software modules could be used in conjunction with the electronic device 12. For example, a microcode implementation, firmware, a database, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), etc. can be utilized. Figure 8

[0105] The processing unit 16 executes various function applications and data processing by running programs stored in the system memory 28, such as implementing the offshore accident prediction method based on interpretable ensemble machine learning provided by the embodiments of the present application.

[0106] Embodiment Five

[0107] The embodiments of the present application also provide a storage medium containing computer executable instructions, which when executed by a computer processor, are used to perform the offshore accident prediction method based on interpretable ensemble machine learning as provided by the above embodiments.

[0108] The computer storage medium of the embodiments of the present application can adopt any combination of one or more computer readable mediums. The computer readable medium can be a computer readable signal medium or a computer readable storage medium. The computer readable storage medium, for example, can be, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination of the above. More specific examples (non-exhaustive list) of the computer readable storage medium include an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this document, the computer readable storage medium can be any tangible medium that contains or stores a program that can be used by / instruction execution system, device or apparatus. ​

[0109] A computer readable signal medium can include a propagated data signal with computer executable code embodied therein. The computer readable signal medium can be any medium that can be used to carry computer executable code for use by or in connection with an instruction execution system, apparatus, or device.

[0110] The computer readable medium can include any medium that can store or transfer information for use by or in connection with an instruction execution system, apparatus, or device.

[0111] The computer program code can be implemented in any of a variety of programming languages, including object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0112] Note that the above-mentioned embodiments illustrate rather than limit the application, since various changes and modifications are possible within the scope of the present application as defined by the appended claims.

Claims

1. A method for predicting maritime accidents based on interpretable ensemble machine learning, characterized in that, include: S101: Acquire multi-source maritime accident investigation data and perform data preprocessing, then divide the data into training set and test set; S102, extract key causal features from the training set using a fusion feature importance method to form a key feature training set; S103, perform ten-fold stratification on the key feature training set to form ten training subsets, use the ten training subsets to perform ten-fold cross-validation training on multiple base models and at least one ensemble learning model composed of multiple base models, and embed the SMOTE oversampling method during the training process to eliminate data imbalance between accident severity and accident type. The base models are multiple independent machine learning models. S104. The performance of the maritime accident prediction model is evaluated using the test set for multiple trained base models and at least one ensemble learning model, and the optimal model is determined based on the evaluation results. S105. The SHAP method is used to perform interpretability analysis on the optimal model. The SHAP value of each feature in each sample of the test set is calculated. Then, the mean of the absolute values ​​of the SHAP of each feature over all samples is calculated and sorted in descending order to form a global feature importance ranking. Feature dependency graph, feature interaction dependency graph and individual prediction interpretation graph are drawn to analyze the interaction coupling effect between any two features and the impact of key causal features in high-risk accidents on the prediction results.

2. The method according to claim 1, characterized in that, S101 includes: Obtain multi-source maritime accident investigation data, and perform deduplication and integrity filtering to obtain a structured dataset; Feature variables are extracted from the structured dataset, and missing values ​​are filled to form the original dataset. The original dataset is then divided into training and test sets using stratified sampling.

3. The method according to claim 1, characterized in that, S102 includes: The training set is evaluated for the importance of fused features. Key causal features are selected by using Pearson correlation coefficient, maximum information coefficient, random forest feature importance index, and SHAP value importance index, and the remaining features are removed to obtain the key feature training set.

4. The method according to claim 3, characterized in that, The key causative features for screening include: Based on the training set, calculate the Pearson correlation coefficient, maximum information coefficient, random forest feature importance index, and SHAP value importance index for each feature dimension. All four indices are expressed as mean values. Then, for each feature, take the mean of these four indices to obtain the comprehensive importance score for each feature. All features are sorted in descending order based on their overall importance score, and the top-ranked features are selected as key causative features.

5. The method according to claim 1, characterized in that, S103 includes: The key feature training set is divided into ten similarly sized subsets according to a stratified sampling strategy, resulting in ten training subsets. Ten-fold cross-validation training is performed on multiple base models and at least one ensemble learning model composed of multiple base models using ten training subsets; During the 10-fold cross-validation training process, the nine training subsets used in each round of training are processed by the SMOTE oversampling method to perform feature space interpolation on the minority class samples to eliminate data imbalance. The remaining training subset is used as the validation set for each round and is not processed.

6. The method according to claim 5, characterized in that, S103 further includes: the ensemble learning model includes a voting ensemble model and a stacked ensemble model; During the model training phase, a random search strategy is used to optimize hyperparameters in a predefined hyperparameter space. The performance of different hyperparameter combinations is evaluated through cross-validation, and the hyperparameter combination with the best performance is selected for model training.

7. The method according to claim 1, characterized in that, S104 includes: The trained base models and at least one ensemble learning model were tested using a test set to obtain the accuracy, precision, recall, F1 score and AUC-ROC curve parameters for each model. Based on the test results, the model with the best performance in predicting maritime accidents was selected and determined as the optimal model.

8. A marine accident prediction device based on interpretable ensemble machine learning, used to implement the marine accident prediction method based on interpretable ensemble machine learning as described in any one of claims 1-7, characterized in that, include: The data preprocessing module is used to acquire and preprocess multi-source maritime accident investigation data; The data balancing module is used to eliminate data imbalance of accident types on the training set using a 10-fold stratified oversampling method; The model training module is used to perform 10-fold cross-validation training on multiple base models and at least one ensemble learning model composed of multiple base models. The model performance evaluation module is used to evaluate the performance of the trained model for maritime accident prediction using a test set to determine the optimal model. The interpretable analysis module is used to perform interpretable analysis on the optimal model using the SHAP method. It calculates the SHAP value of each feature in each sample of the test set, then calculates the mean of the absolute SHAP values ​​of each feature across all samples and sorts them in descending order to form a global feature importance ranking, and draws feature dependency graphs and individual prediction interpretation graphs.

9. An electronic device, characterized in that, The electronic device includes: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the maritime accident prediction method based on interpretable ensemble machine learning as described in any one of claims 1-7.

10. A storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the maritime accident prediction method based on interpretable ensemble machine learning as described in any one of claims 1-7.

Images