Intelligent auxiliary decision-making method and system for ship in complex operation scene

Abstract

The invention relates to an intelligent auxiliary decision-making method and system for a ship in a complex operation scene, and the method comprises the steps: obtaining the navigation, ship body, cabin and energy efficiency data of the ship in historical operation, and forming a multi-source heterogeneous data set; state judgment is carried out based on a parameter preset range, and a structured data set with binary tags is generated; carrying out standardization processing on the structured data set, and carrying out category balance processing on the training set by adopting a mixed oversampling strategy; a CART decision-making tree is used as a base learning device, hyper-parameters are optimized in combination with a random search optimization and cross validation method, and a random forest aided decision-making model is trained and evaluated; and inputting data acquired in real time into the random forest auxiliary decision-making model after the performance reaches the standard, outputting auxiliary decision-making information including the category of the ship operation state, the safety risk level and the abnormal early warning information, and converting the auxiliary decision-making information into a visual graph or chart through a feature importance visualization technology. And real-time intelligent auxiliary decision making of the ship in a complex operation scene is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent auxiliary decision-making technology for ships, specifically to an intelligent auxiliary decision-making method and system for ships in complex operating scenarios. Background Technology

[0002] The efficiency and scientific rigor of ship decision-making in complex operational scenarios directly impact navigation safety, operating costs, and environmental benefits. Traditional manual decision-making suffers from issues such as slow response times, poor multi-objective coordination, and low data utilization, making it difficult to achieve a dynamic balance between safety, economy, and energy efficiency goals. With the International Maritime Organization (IMO) raising its energy efficiency standards, shipping companies urgently need intelligent decision-making solutions. Random forest algorithms offer significant advantages in handling high-dimensional data, mitigating overfitting, and assessing feature importance; however, existing ship decision support systems have not fully utilized the algorithm's characteristics and lack systematic solutions for multi-subsystem data fusion, real-time decision generation, and imbalanced data processing.

[0003] Current mainstream solutions mainly include theoretical modeling based on physical formulas, rule engines and threshold alarm systems, single machine learning model applications, and independent auxiliary systems for single business scenarios. These solutions have significant drawbacks: 1) Physical modeling methods rely on simplified assumptions and ignore historical data mining, making it difficult to characterize the dynamic interaction of multiple variables; 2) The response of threshold alarms combined with human decision-making is lagging, and rule engines cannot handle nonlinear risk coupling; 3) Single machine learning models are limited by the imbalance of data categories and the ability to capture feature interactions, resulting in a high rate of missed detection for rare risks and a tendency to overfit; 4) Independent systems suffer from severe data isolation and lack collaborative support across the entire operation process, making them unable to adapt to the IMO energy efficiency standard upgrade requirements and limited by the poor interpretability and high computational resource consumption of deep learning models.

[0004] Therefore, there is an urgent need for an intelligent auxiliary decision-making method that can integrate multi-source heterogeneous data on ship navigation, hull, engine room, and energy efficiency, effectively handle the problem of class imbalance, and have high classification accuracy and strong interpretability, so as to realize real-time classification of ship operating status and early warning of safety risk levels, and improve the scientific nature of decision-making and response timeliness in complex operating scenarios. Summary of the Invention

[0005] To address the issues of insufficient multi-dimensional data fusion and delayed real-time decision-making in complex ship operation scenarios, this invention provides an intelligent auxiliary decision-making method for ships in complex operating scenarios. This method achieves multi-objective collaborative decision-making through data-driven and ensemble learning, improving operational reliability and economy. This invention also provides an intelligent auxiliary decision-making system for ships in complex operating scenarios.

[0006] The technical solution of the present invention is as follows:

[0007] A method for intelligent auxiliary decision-making for ships in complex operational scenarios, characterized by the following steps:

[0008] Multi-dimensional data acquisition steps: Acquire navigation data, hull data, engine room data and energy efficiency data recorded by the ship during its historical operating cycle to form a multi-source heterogeneous dataset containing multiple dimensions;

[0009] Data labeling steps: Based on the preset normal value range of each data in the multi-source heterogeneous dataset, the state of each sample in the multi-source heterogeneous dataset is judged. If any parameter in a sample exceeds its preset normal value range, the sample is marked as an abnormal state; if all parameters in the sample are within the normal value range, the sample is marked as a normal state, and a structured dataset with binary labels is generated.

[0010] The class balancing process involves standardizing the structured dataset and dividing it into a training set and a test set according to a preset ratio. Then, a hybrid oversampling strategy, including random resampling and SMOTE oversampling, is used to augment the minority class samples in the training set whose sample count is less than a preset threshold, thereby achieving class balancing in the training set.

[0011] Model training and hyperparameter optimization steps: Using CART decision trees as base learners, a random forest auxiliary decision model with classification function is constructed using the Bagging strategy; the random forest auxiliary decision model is trained using the training set after class balancing, and the hyperparameters of the random forest auxiliary decision model are optimized by combining random search optimization method and cross-validation method during the training process to obtain the trained and optimized random forest auxiliary decision model; the performance of the trained and optimized random forest auxiliary decision model is evaluated by the test set to obtain the random forest auxiliary decision model that meets the performance evaluation criteria;

[0012] Intelligent Assisted Decision-Making Steps: Real-time collection of ship navigation data, hull data, engine room data, and energy efficiency data, followed by standardization processing similar to that in the category balancing step, and input into the random forest assisted decision-making model after performance evaluation has met the standards; The random forest assisted decision-making model determines the category to which the ship's current operating status belongs based on its built-in classification logic, and automatically generates and outputs assisted decision-making information including the category to which the ship's operating status belongs, the safety risk level, and abnormal warning information by combining the feature importance corresponding to the category. This information is then converted into visual graphics or charts through feature importance visualization technology, enabling real-time intelligent assisted decision-making for ships in complex operating scenarios.

[0013] Preferably, in the class balancing process, the use of a hybrid oversampling strategy to augment the minority class samples in the training set to achieve class balancing specifically includes: firstly, randomly resampling the minority class samples in the training set whose sample size is less than a preset threshold to augment their sample size; and then, based on the random resampling, using SMOTE oversampling on all augmented minority class samples to generate synthetic samples, thereby increasing sample diversity and reducing the risk of overfitting; if SMOTE oversampling fails due to insufficient nearest neighbor samples, it automatically switches to random resampling to achieve class balancing.

[0014] Preferably, the optimization of the hyperparameters of the random forest-aided decision-making model in the model training and hyperparameter optimization steps specifically includes:

[0015] The RandomizedSearchCV (Randomized SearchCV) optimization method is used to perform multiple random searches and randomly select multiple hyperparameter combinations. For each selected hyperparameter combination, a 5-fold cross-validation method is used, with the model prediction accuracy as the optimization objective, to select the optimal hyperparameter combination.

[0016] Preferably, the standardization process of the structured dataset in the category balancing step specifically includes: using the StandardScaler method to standardize the feature data in the structured dataset other than the binary labels, converting each feature data into a standard form with a mean of 0 and a standard deviation of 1, so as to eliminate the dimensional differences between different parameters.

[0017] Preferably, in the model training and hyperparameter optimization steps, the performance evaluation of the trained and optimized random forest-aided decision-making model using a test set specifically includes:

[0018] The performance of the trained and optimized random forest-assisted decision-making model is evaluated using a test set and based on multi-dimensional evaluation metrics, including classification report, confusion matrix, and area under the ROC curve. The classification report is used to analyze the classification accuracy of samples of each category, including precision, recall, and F1 score. The confusion matrix is ​​used to present the matching between the sample prediction results and the true labels. The area under the ROC curve is used to measure the generalization ability of the model. When all multi-dimensional evaluation metrics reach their respective preset performance thresholds, the random forest-assisted decision-making model that has passed the performance evaluation is obtained.

[0019] An intelligent auxiliary decision-making system for ships in complex operating scenarios is characterized by comprising, in sequence, a multi-dimensional data acquisition module, a data annotation module, a category balancing module, a model training and hyperparameter optimization module, and an intelligent auxiliary decision-making module.

[0020] The multi-dimensional data acquisition module acquires navigation data, hull data, engine room data and energy efficiency data recorded by the ship during its historical operating cycle, forming a multi-source heterogeneous dataset containing multiple dimensions.

[0021] The data labeling module performs a status judgment on each sample in the multi-source heterogeneous dataset based on the preset normal value range of each data in the multi-source heterogeneous dataset. If any parameter in a sample exceeds its preset normal value range, the sample is marked as an abnormal state; if all parameters in the sample are within the normal value range, the sample is marked as a normal state, and a structured dataset with binary labels is generated.

[0022] The class balancing module standardizes the structured dataset and divides the standardized structured dataset into a training set and a test set according to a preset ratio. Then, it uses a hybrid oversampling strategy, including random resampling and SMOTE oversampling, to expand the minority class samples in the training set whose sample count is less than a preset threshold, so as to achieve class balancing of the training set.

[0023] The model training and hyperparameter optimization module uses CART decision trees as base learners and employs a Bagging strategy to construct a random forest auxiliary decision model with classification capabilities. The random forest auxiliary decision model is trained using a training set after class balancing, and during training, the hyperparameters of the random forest auxiliary decision model are optimized using a random search optimization method and a cross-validation method to obtain a trained and optimized random forest auxiliary decision model. The performance of the trained and optimized random forest auxiliary decision model is then evaluated using a test set to obtain a random forest auxiliary decision model that meets the performance evaluation criteria.

[0024] The intelligent auxiliary decision-making module collects real-time navigation data, hull data, engine room data, and energy efficiency data of the ship. After undergoing the same standardized processing as in the category balancing module, the data is input into the random forest auxiliary decision-making model after the performance evaluation has met the standards. The random forest auxiliary decision-making model determines the category to which the ship's current operating status belongs based on its built-in classification logic. Combining the feature importance corresponding to the category, it automatically generates and outputs auxiliary decision-making information, including the category to which the ship's operating status belongs, the safety risk level, and abnormal warning information. This information is then converted into visual graphics or charts through feature importance visualization technology, enabling real-time intelligent auxiliary decision-making for ships in complex operating scenarios.

[0025] Preferably, in the class balancing processing module, the use of a hybrid oversampling strategy to expand the minority class samples in the training set to achieve class balancing specifically includes: firstly, randomly resampling the minority class samples in the training set whose sample size is less than a preset threshold to expand their sample size; and then, based on the random resampling, using SMOTE oversampling on all expanded minority class samples to generate synthetic samples to increase sample diversity and reduce the risk of overfitting; if SMOTE oversampling fails due to insufficient nearest neighbor samples, it automatically switches to random resampling to achieve class balancing.

[0026] Preferably, the optimization of the hyperparameters of the random forest-aided decision-making model in the model training and hyperparameter optimization module specifically includes:

[0027] A random search optimization method is used to perform multiple random searches and randomly select multiple hyperparameter combinations. For each selected hyperparameter combination, a cross-validation method is used with the model prediction accuracy as the optimization objective to select the optimal hyperparameter combination.

[0028] Preferably, in the category balancing processing module, the standardization processing of the structured dataset specifically includes: using the StandardScaler method to standardize the feature data in the structured dataset other than the binary labels, converting each feature data into a standard form with a mean of 0 and a standard deviation of 1, so as to eliminate the dimensional differences between different parameters.

[0029] Preferably, in the model training and hyperparameter optimization module, the performance evaluation of the trained and optimized random forest-aided decision-making model using a test set specifically includes:

[0030] The performance of the trained and optimized random forest-assisted decision-making model is evaluated using a test set and based on multi-dimensional evaluation metrics, including classification report, confusion matrix, and area under the ROC curve. The classification report is used to analyze the classification accuracy of samples of each category, including precision, recall, and F1 score. The confusion matrix is ​​used to present the matching between the sample prediction results and the true labels. The area under the ROC curve is used to measure the generalization ability of the model. When all multi-dimensional evaluation metrics reach their respective preset performance thresholds, the random forest-assisted decision-making model that has passed the performance evaluation is obtained.

[0031] The beneficial effects of this invention are as follows:

[0032] This invention provides an intelligent auxiliary decision-making method for ships in complex operating scenarios, also known as a ship auxiliary decision-making method based on the CART-Random Forest algorithm. First, it acquires navigation data, hull data, engine room data, and energy efficiency data recorded during the ship's historical operating cycle, forming a multi-source heterogeneous dataset containing multiple dimensions. By constructing an 82-dimensional multi-source heterogeneous data fusion framework covering four subsystems—intelligent navigation, hull, engine room, and energy efficiency—it overcomes the limitations of traditional single-dimensional monitoring, solves the data silo problem in traditional auxiliary decision-making systems, and achieves integrated data correlation analysis of the entire ship operation process, providing a data foundation for high-dimensional feature modeling. Furthermore, based on preset normal value ranges for each data point, it performs state judgment on each sample in the multi-source heterogeneous dataset, generating a structured dataset with binary labels. By transforming continuous operating parameters into clear "normal / abnormal" state labels, it transforms the complex multivariate regression problem into a more stable and intuitive classification problem, greatly reducing the learning difficulty of the model and improving the interpretability of the decision results. Then, the structured dataset was standardized to eliminate the influence of different dimensional parameters on the model. A hybrid oversampling strategy (random resampling and SMOTE oversampling) was used to expand the minority class samples in the training set to achieve class balance in the training set. This effectively solved the overfitting problem of traditional algorithms in small sample scenarios and the class imbalance problem caused by the scarcity of abnormal state samples in ship operation data. It fundamentally avoided the defect of machine learning models biased towards the majority class (normal state) and missed rare risks (abnormal state), and significantly improved the recall rate of the model for high-risk scenarios. Using CART decision trees as the base learner, a random forest-assisted decision model is constructed using the Bagging strategy. By employing the ensemble strategy of CART decision trees and Bagging, the advantages of the random forest-assisted decision model in handling high-dimensional features, capturing nonlinear relationships, and mitigating overfitting are fully utilized. The random forest-assisted decision model is trained using the training set, and during the training process, the hyperparameters of the random forest-assisted decision model are optimized by combining random search optimization and cross-validation methods. Hyperparameter optimization through random search and cross-validation avoids the blindness of manual parameter tuning and automatically finds the optimal configuration of the model with higher efficiency, thereby ensuring that the final random forest-assisted decision model has both high accuracy and high robustness.Finally, the real-time collected data is input into the random forest-assisted decision-making model after the performance evaluation meets the standards. The model outputs auxiliary decision-making information, including the category of the ship's operating status, the level of safety risk, and abnormal warning information. The auxiliary decision-making information and its associated risk factors are transformed into visual graphics or charts through feature importance visualization technology. This enables real-time intelligent auxiliary decision-making for ships in complex operating scenarios. By outputting intuitive risk levels and warning information, the data-driven algorithm decision-making logic is transformed into operational suggestions that crew members can directly understand. This breaks down the "black box" barrier of artificial intelligence and ultimately achieves the synergistic optimization goals of improving navigation safety, reducing operating costs, and meeting energy efficiency standards.

[0033] This invention provides a novel systematic solution to multi-objective decision-making problems in complex ship operation scenarios by constructing a data-driven, model-optimized, and collaborative intelligent auxiliary framework. Its overall technical effectiveness is reflected in the following four aspects: 1) By systematically accessing and integrating 82-dimensional feature parameters from four subsystems—intelligent navigation, intelligent hull, intelligent engine room, and intelligent energy efficiency—a unified data pool covering the entire ship operation process is built. This breaks down the data silos caused by traditional single subsystems or independent solutions, providing a global perspective for the model and fundamentally laying the data foundation for multi-objective (safety, economy, energy efficiency) collaborative optimization. 2) A hybrid oversampling (SMOTE + random oversampling) strategy is adopted to solve the problem of data class imbalance. Using CART decision trees as the base learner, a random forest auxiliary decision-making model is constructed using a Bagging ensemble strategy, and hyperparameter optimization is performed by combining random search and cross-validation. This significantly improves the recognition rate (high recall) of rare high-risk scenarios (such as sudden failures and extreme sea conditions) while ensuring high model accuracy and strong generalization ability, effectively avoiding overfitting, and making the model performance far exceed that of traditional single models. 3) By introducing feature importance analysis, the algorithm's decision-making logic is transformed into a quantitative and visualized risk cause report (e.g., showing visibility as the primary cause of collision warnings). This breaks down the "black box" barrier of artificial intelligence, making the decision-making process transparent and understandable to the crew, greatly enhancing their trust in the intelligent system, and realizing a shift from a "passive alarm" to a "proactive cognition and decision-making" human-machine collaborative model. In summary, this invention enables real-time, accurate, and explainable intelligent auxiliary decision-making for ships in complex operating scenarios, significantly improving navigation safety, decision-making scientificity, and operational efficiency management.

[0034] Furthermore, a hybrid oversampling strategy is employed to augment the minority class samples in the training set to achieve class balance. Specifically, this involves: firstly, randomly resampling the minority class samples in the training set whose sample size is less than a preset threshold to expand their sample size. This rapidly increases the sample cardinality of the very few classes by replicating existing abnormal samples, addressing the problem of the SMOTE algorithm failing to start due to insufficient samples (e.g., less than 5), and ensuring the executability of subsequent oversampling processes. Based on the random resampling, SMOTE oversampling is applied to all minority class samples to generate synthetic samples, increasing sample diversity and reducing overfitting risk. This avoids the model memory effect caused by simple replication, enhances the continuity of minority class distribution in the feature space, improves the learning quality of classification boundaries, and effectively reduces the risk of model overfitting. If SMOTE oversampling fails due to insufficient nearest neighbor samples, the system automatically switches to a random resampling strategy. Dynamic fallback is achieved by setting judgment conditions such as k-nearest neighbor search failure, ensuring that class balance can still be achieved even in cases of sparse or unevenly distributed local samples, thus improving the robustness and engineering applicability of the entire oversampling process.

[0035] Furthermore, the optimization of the hyperparameters of the random forest-assisted decision-making model specifically includes: employing the RandomizedSearchCV (Randomized SearchCV) method to perform multiple random searches, randomly selecting multiple sets of hyperparameter combinations. For each selected hyperparameter combination, a 5-fold cross-validation method is used, with the model's prediction accuracy as the optimization objective. This approach can accurately select the optimal hyperparameter combination while maintaining computational efficiency, significantly improving the model's classification performance. Through the hyperparameter optimization mechanism of random search + cross-validation, automatic optimization is achieved under reasonable computational resource consumption, avoiding the reliance on experience and blind spots of manual parameter tuning. This effectively improves the adaptability of the random forest-assisted decision-making model on multi-source ship data, enabling the model to achieve high performance on the test set and ensuring the high accuracy and deployability of the intelligent assisted decision-making results. Simultaneously, the cross-validation mechanism enhances the robustness of model evaluation, providing a reliable basis for subsequent performance benchmark assessments.

[0036] Furthermore, the standardization process for the structured dataset specifically includes: using the StandardScaler method to standardize the feature data in the structured dataset, excluding the binary labels, and converting each feature data into a standard form with a mean of 0 and a standard deviation of 1, so as to eliminate the dimensional differences between different parameters and effectively solve the modeling bias problem caused by differences in units and orders of magnitude in multi-source ship data.

[0037] Furthermore, the performance of the random forest classifier is evaluated using a test set and based on multi-dimensional evaluation metrics, including classification reports (precision, recall, F1 score), confusion matrix (such as the number of true positives, true negatives, false positives, and false negatives), and area under the ROC curve (ROC-AUC). When the ROC-AUC and other multi-dimensional evaluation metrics reach their respective preset performance thresholds, a random forest-assisted decision-making model that meets the performance evaluation criteria is obtained. This ensures that the model not only has high classification accuracy but also achieves a balance between abnormal sample identification rate and false alarm control, significantly improving the reliability of model decisions and the overall security of the system.

[0038] This invention also relates to an intelligent auxiliary decision-making system for ships in complex operating scenarios. This system corresponds to the aforementioned intelligent auxiliary decision-making method for ships in complex operating scenarios and can be understood as a system that implements the aforementioned intelligent auxiliary decision-making method for ships in complex operating scenarios. It includes a multi-dimensional data acquisition module, a data annotation module, a class balancing processing module, a model training and hyperparameter optimization module, and an intelligent auxiliary decision-making module, connected sequentially. These modules work collaboratively, fusing heterogeneous data from multiple sources such as navigation, hull, engine room, and energy efficiency to construct a high-dimensional feature space (e.g., 82-dimensional input features). This achieves comprehensive perception and systematic modeling of the ship's overall operational status, effectively solving problems such as data silos, response lag, and poor multi-objective coordination in traditional auxiliary decision-making systems. Through standardization processing, the dimensional differences between different physical quantities such as speed, pressure, and temperature are eliminated, ensuring fair contribution of each feature in model training. Furthermore, a hybrid oversampling strategy (random resampling + SMOTE) is introduced to prioritize expanding the sample base of the few minority classes and generate diverse synthetic samples, significantly alleviating the class imbalance problem caused by scarce abnormal ship data and improving the model's ability to identify rare faults. Using CART decision trees as the base learner, a random forest-assisted decision-making model is constructed using a Bagging ensemble strategy, effectively reducing variance and suppressing overfitting. Combining random search optimization and cross-validation methods, the optimal hyperparameter combination is automatically searched with accuracy as the optimization objective, improving model tuning efficiency and generalization performance. By collecting multi-source ship data in real time and inputting it into a performance-evaluated random forest classifier, the system outputs the current operational status classification result, generating interpretable safety risk levels or anomaly warning information. This supports crew members in quickly identifying the causes of risks, upgrading from "post-event alarms" to "pre-event warnings + intelligent decision-making suggestions." This invention features low computational overhead, strong interpretability, and does not rely on deep learning frameworks or GPU resources. It is suitable for deployment in ship embedded systems, possessing good engineering practicality and promotional value. It significantly improves the scientific nature of ship decision-making, response timeliness, and navigation safety in complex operational scenarios, providing reliable technical support for the construction of intelligent shipping systems. Attached Figure Description

[0039] Figure 1 This is a flowchart of the intelligent auxiliary decision-making method for ships in complex operating scenarios according to the present invention.

[0040] Figure 2 This is a preferred flowchart of the intelligent auxiliary decision-making method for ships in complex operating scenarios according to the present invention. Detailed Implementation

[0041] The present invention will now be described with reference to the accompanying drawings.

[0042] This invention relates to an intelligent auxiliary decision-making method for ships in complex operating scenarios, the flowchart of which is shown below. Figure 1 As shown, the steps are as follows:

[0043] I. Multi-dimensional data acquisition steps: Acquire navigation data, hull data, engine room data and energy efficiency data recorded by the ship during its historical operating cycle to form a multi-source heterogeneous dataset containing multiple dimensions.

[0044] Specifically, such as Figure 2 As shown, the first step is to load a multi-dimensional ship operation data scenario: through a sensor network deployed throughout the ship, multi-dimensional operational data from four subsystems—intelligent navigation, intelligent hull, intelligent engine room, and intelligent energy efficiency—are systematically collected. The intelligent navigation subsystem collects 24 parameters related to navigation, collision avoidance, and motion status, such as speed, load, distance between two ships (DCPA), time to closest encounter (TCPA), longitude, latitude, heading, course, relative course, and visibility, to comprehensively assess positional anomalies, motion status, and collision risk. The intelligent hull subsystem collects 20 parameters related to the safety and stability of the hull structure, such as water depth, roll angle, ship stability height (GM value), pitch angle, draft, freeboard, acceleration, and hull stress, to comprehensively assess the risks of grounding, capsizing, and sinking. The intelligent engine room subsystem covers seven major modules: main engine, auxiliary engines, shafting, fuel, lubricating oil, cooling water, and electrical systems. It collects 20 fault warning indicators, such as fuel line pressure, shafting vibration amplitude, main engine speed, main engine exhaust temperature, scavenging air pressure, lubricating oil pressure and temperature, cylinder liner water temperature, and generator voltage and frequency. The intelligent energy efficiency subsystem collects 18 economic indicators related to ship fuel economy and emissions, such as optimal speed, daily fuel consumption (SFOC), ship trim, main engine power, propulsion efficiency, auxiliary engine power, and carbon intensity index (CII). The system systematically collects 2000 sets of the above data, forming a multi-source heterogeneous raw dataset with 82 dimensions of features. This breaks through the limitations of traditional single-dimensional monitoring and lays a comprehensive data foundation for subsequent multi-objective collaborative decision-making.

[0045] II. Data Labeling Steps: Based on the preset normal value range of each data in the multi-source heterogeneous dataset, the state of each sample in the multi-source heterogeneous dataset is judged. If any parameter in a sample exceeds its preset normal value range, the sample is marked as an abnormal state; if all parameters in the sample are within the normal value range, the sample is marked as a normal state, and a structured dataset with binary labels is generated.

[0046] Specifically, firstly, a safe normal value range is pre-defined for each data item in the multi-source heterogeneous dataset based on prior knowledge. For example, the normal range for "roll angle" is set to -15° to +15°, and the normal range for "fuel line pressure" is set to 0.5MPa to 0.9MPa. Then, the state of each sample in the multi-source heterogeneous dataset (each sample contains all 82 parameters) is assessed. If the value of any parameter in a sample exceeds its own pre-defined normal value range, the overall state of that sample is marked as abnormal (label 1); only when all 82 parameters in a sample are within their respective normal value ranges is the sample marked as normal (label 0). Finally, a structured dataset containing feature vectors (X) and binary labels (y) is generated, covering both continuous and discrete features, to support subsequent machine learning model training.

[0047] III. Class Balance Processing Steps: The structured dataset is standardized, and the standardized structured dataset is divided into training set and test set according to a preset ratio. Then, a hybrid oversampling strategy is used to expand the minority class samples in the training set whose sample number is less than a preset threshold, so as to achieve class balance processing of the training set. The hybrid oversampling strategy includes random resampling and SMOTE oversampling.

[0048] Specifically, the StandardScaler method is first used to standardize the feature data (i.e., feature vector X) in the structured dataset, excluding the binary labels (i.e., as shown below). Figure 2The preprocessing shown transforms each feature data into a standard form with a mean of 0 and a standard deviation of 1 to eliminate the influence of different dimensions and orders of magnitude among the 82-dimensional features, ensuring that the data conforms to a standard normal distribution. Subsequently, the standardized structured dataset is randomly divided into training and test sets in a 7:3 ratio to ensure that the test set represents the true data distribution. To address the class imbalance problem, a hybrid oversampling strategy (hierarchical strategy) is employed to augment the minority class samples in the training set, achieving class balance (data balance). First, the number of "abnormal" and "normal" samples in the training set is counted, identifying the "abnormal" class (e.g., <50 samples) as the minority class. Then, a hybrid oversampling strategy is used to randomly resample (i.e., copy existing samples), rapidly expanding the sample size to a preset base (e.g., 100) to reduce overfitting risk. Based on this random resampling (also known as random oversampling), the SMOTE oversampling algorithm is applied to all expanded minority class samples to generate synthetic samples. The SMOTE oversampling algorithm synthesizes new, diverse minority class samples in the feature space through linear interpolation (e.g., synthesizing a new, reasonable pressure value sample based on two similar "low fuel pressure" samples), ensuring data distribution balance. Accurate oversampling is achieved by dynamically adjusting the number of nearest neighbors, reducing overfitting risk and laying a solid data foundation for subsequent model training. If the SMOTE algorithm fails to find effective nearest neighbors due to an insufficient number of samples in a certain category (less than 5), it automatically switches back to a pure random resampling strategy to ensure the process completes. This step fundamentally solves the problem of data class imbalance and significantly improves the sensitivity and recall rate for identifying high-risk, low-occurrence anomalies.

[0049] IV. Model Training and Hyperparameter Optimization Steps: Using CART decision trees as base learners, a random forest auxiliary decision model with classification capabilities is constructed using the Bagging strategy. The random forest auxiliary decision model is trained using a training set after class balancing. During the training process, the hyperparameters of the random forest auxiliary decision model are optimized by combining random search optimization and cross-validation methods to obtain the trained and optimized random forest auxiliary decision model. The performance of the trained and optimized random forest auxiliary decision model is evaluated using a test set to obtain a random forest auxiliary decision model that meets the performance evaluation criteria.

[0050] Specifically, firstly, a CART decision tree is used as the base learner, with the Gini coefficient as the splitting criterion; then, a Bagging strategy is employed to construct a random forest-assisted decision model (a type of random forest model, or a random forest classification model, or an ensemble learning model based on CART-random forest, also corresponding to...). Figure 2The CART-Random Forest model shown is used to improve overfitting resistance and feature importance analysis capabilities. The random forest-aided decision-making model is then trained using a class-balanced training set. During training, to achieve optimal performance, a hyperparameter grid for the random forest-aided decision-making model is defined (including the number of trees n_estimators: [100, 200, 300], maximum depth max_depth: [10, 20, None], etc.). The RandomizedSearchCV method is used for 200 random search samplings to randomly select multiple hyperparameter combinations. For each selected hyperparameter combination, a 5-fold cross-validation method is used, and the optimal hyperparameter combination is found with the model's prediction accuracy as the optimization objective. The optimal hyperparameters are then used for training to obtain the final trained and optimized random forest-aided decision-making model. Finally, the performance of the trained and optimized random forest-assisted decision-making model is evaluated using a test set and based on evaluation metrics. This involves comprehensively measuring the model's generalization ability by generating classification reports (precision, recall, F1 score), confusion matrices (e.g., number of true positives, true negatives, false positives, and false negatives), and calculating the area under the ROC curve (ROC-AUC), thus validating model performance and uncovering feature value. Simultaneously, feature importance analysis visually presents the contribution of each feature, providing a quantitative basis for feature selection and model optimization. For example, the analysis found that "visibility" and "distance between two ships" have the highest contribution weight to collision risk prediction, while "daily fuel consumption" contributes the most to energy efficiency anomalies. This analysis result can be used to guide subsequent feature engineering, eliminating redundant features and simplifying the model structure. When multiple evaluation metrics, such as ROC-AUC and F1 score, reach their respective preset performance thresholds (e.g., ROC-AUC ≥ 0.95 and F1 score ≥ 0.90), the model performance is confirmed to be up to standard, resulting in a qualified random forest classifier, which is saved as a .pkl file, indicating its performance status. If ROC-AUC < 0.95 or F1 score < 0.90, the hyperparameter range is expanded, and the model is re-evaluated. Furthermore, Docker containerization technology is used to package the performance-evaluated random forest classifier and its runtime environment, deploying them on a shipboard embedded industrial control computer. This achieves lightweight deployment, enabling the model to run stably on the embedded shipboard industrial control computer, enhancing the system's environmental adaptability and portability. It achieves second-level decision-making closed-loop in an 8GB memory hardware environment, reducing deployment costs compared to traditional solutions, ensuring compatibility with the shipbuilding industrial environment, meeting IMO energy efficiency standards, and achieving multi-objective collaborative optimization of safety, efficiency, and energy efficiency.

[0051] V. Intelligent Assisted Decision-Making Steps: Real-time collection of ship navigation data, hull data, engine room data, and energy efficiency data, after being standardized in the same way as in the category balancing process, is input into the random forest assisted decision-making model after the performance evaluation has met the standards. The random forest assisted decision-making model determines the category to which the ship's current operating status belongs based on its built-in classification logic, and automatically generates and outputs assisted decision-making information including the category to which the ship's operating status belongs, the safety risk level, and abnormal warning information by combining the feature importance corresponding to the category. This information is then converted into visual graphics or charts through feature importance visualization technology, realizing real-time intelligent assisted decision-making for ships in complex operating scenarios.

[0052] Specifically, during the real-time operation of the ship, the deployed sensors collect 82 dimensions of data from four major subsystems: navigation, hull, engine room, and energy efficiency. After undergoing the same standardization process as during training in the category balancing step, the real-time 82-dimensional data is input into the random forest auxiliary decision-making model after the performance evaluation has met the standards. The model first outputs a binary classification result (i.e., "1-abnormal" or "0-normal") based on its built-in classification logic. Then, it combines the feature importance of the corresponding category of the classification result (e.g., when it is judged as "1-abnormal", the model simultaneously outputs the importance ranking of key influencing parameters such as "distance between the two ships (contribution weight 45%)" and "fuel pipeline pressure (contribution weight 30%)") to automatically generate specific auxiliary decision-making information. This includes the category to which the ship's operating status belongs ("abnormal category - collision risk association" or "normal category - energy efficiency compliance"), the safety risk level (e.g., "Level 1 warning" corresponds to a feature importance sum ≥ 70%, "Level 2 warning" corresponds to a feature importance sum 50%-70%), and abnormal warning information (e.g., "high collision risk, core influencing factor: distance between the two ships < 2 nautical miles", "abnormal fuel pressure, core influencing factor: pressure value exceeds the normal range by 15%). Furthermore, by using feature importance visualization technology, auxiliary decision-making information and the ranking of the importance of key features are transformed into intuitive and visual graphics or charts (such as feature contribution weight bar charts and risk level-influencing factor correlation heatmaps), which are then presented to the crew in real time through a visualization interface. This enables real-time and interpretable intelligent auxiliary decision-making for ships in complex operating scenarios, taking into account safety, economy and energy efficiency goals.

[0053] This invention also relates to an intelligent auxiliary decision-making system for ships in complex operating scenarios. This intelligent auxiliary decision-making system for ships in complex operating scenarios corresponds to the aforementioned intelligent auxiliary decision-making method for ships in complex operating scenarios, and can also be understood as a system that implements the aforementioned intelligent auxiliary decision-making method for ships in complex operating scenarios. The system includes, in sequence, a multi-dimensional data acquisition module, a data annotation module, a category balancing processing module, a model training and hyperparameter optimization module, and an intelligent auxiliary decision-making module. Specifically,

[0054] The multi-dimensional data acquisition module acquires navigation data, hull data, engine room data and energy efficiency data recorded by the ship during its historical operating cycle, forming a multi-source heterogeneous dataset containing multiple dimensions.

[0055] The data labeling module performs a status judgment on each sample in the multi-source heterogeneous dataset based on the preset normal value range of each data in the multi-source heterogeneous dataset. If any parameter in a sample exceeds its preset normal value range, the sample is marked as an abnormal state; if all parameters in the sample are within the normal value range, the sample is marked as a normal state, and a structured dataset with binary labels is generated.

[0056] The class balancing module standardizes the structured dataset and divides the standardized structured dataset into a training set and a test set according to a preset ratio. Then, it uses a hybrid oversampling strategy, including random resampling and SMOTE oversampling, to expand the minority class samples in the training set whose sample count is less than a preset threshold, so as to achieve class balancing of the training set.

[0057] The model training and hyperparameter optimization module uses CART decision trees as base learners and employs a Bagging strategy to construct a random forest auxiliary decision model with classification capabilities. The random forest auxiliary decision model is trained using a training set after class balancing, and during training, the hyperparameters of the random forest auxiliary decision model are optimized using a random search optimization method and a cross-validation method to obtain a trained and optimized random forest auxiliary decision model. The performance of the trained and optimized random forest auxiliary decision model is then evaluated using a test set to obtain a random forest auxiliary decision model that meets the performance evaluation criteria.

[0058] The intelligent auxiliary decision-making module collects real-time navigation data, hull data, engine room data, and energy efficiency data of the ship. After undergoing the same standardized processing as in the category balancing module, the data is input into the random forest auxiliary decision-making model after the performance evaluation has met the standards. The random forest auxiliary decision-making model determines the category to which the ship's current operating status belongs based on its built-in classification logic. Combining the feature importance corresponding to the category, it automatically generates and outputs auxiliary decision-making information, including the category to which the ship's operating status belongs, the safety risk level, and abnormal warning information. This information is then converted into visual graphics or charts through feature importance visualization technology, enabling real-time intelligent auxiliary decision-making for ships in complex operating scenarios.

[0059] Preferably, in the class balancing processing module, the mixed oversampling strategy is used to expand the minority class samples in the training set to achieve class balancing. Specifically, this includes: firstly, randomly resampling the minority class samples in the training set whose sample size is less than a preset threshold to expand their sample size; and then, based on the random resampling, using SMOTE oversampling to generate synthetic samples from all expanded minority class samples to increase sample diversity and reduce the risk of overfitting; if SMOTE oversampling fails due to insufficient neighbor samples, it automatically switches to random resampling to achieve class balancing.

[0060] Preferably, in the model training and hyperparameter optimization module, optimizing the hyperparameters of the random forest-assisted decision-making model specifically includes: performing multiple random searches using a random search optimization method to randomly select multiple sets of hyperparameter combinations; for each selected set of hyperparameter combinations, using a cross-validation method with the model prediction accuracy as the optimization objective, to select the optimal hyperparameter combination.

[0061] Preferably, in the category balancing processing module, the standardization processing of the structured dataset specifically includes: using the StandardScaler method to standardize the feature data in the structured dataset other than the binary labels, converting each feature data into a standard form with a mean of 0 and a standard deviation of 1, so as to eliminate the dimensional differences between different parameters.

[0062] Preferably, in the model training and hyperparameter optimization module, the performance evaluation of the trained and optimized random forest auxiliary decision-making model using a test set specifically includes: evaluating the performance of the trained and optimized random forest auxiliary decision-making model using a test set and based on multi-dimensional evaluation metrics, including classification reports, confusion matrices, and the area under the ROC curve; wherein, the classification report is used to analyze the classification accuracy of samples of each category, including precision, recall, and F1 score; the confusion matrix is ​​used to present the matching situation between the sample prediction results and the true labels; and the area under the ROC curve is used to measure the generalization ability of the model. When all multi-dimensional evaluation metrics reach their respective preset performance thresholds, a random forest auxiliary decision-making model that has passed the performance evaluation is obtained.

[0063] This invention provides an objective and scientific intelligent auxiliary decision-making method and system for ships in complex operating scenarios. By constructing an intelligent auxiliary framework that is data-driven, model-optimized, and collaborative in decision-making, it provides a brand-new systematic solution to the multi-objective decision-making problem in complex ship operating scenarios. It realizes real-time, accurate, and interpretable intelligent auxiliary decision-making for ships in complex operating scenarios, and significantly improves navigation safety, decision-making scientificity, and operational energy efficiency management.

[0064] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail with reference to the accompanying drawings and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention. In short, all technical solutions and improvements that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention patent.

Claims

1. A method for intelligent auxiliary decision-making for ships in complex operational scenarios, characterized in that, Includes the following steps: Multi-dimensional data acquisition steps: Acquire navigation data, hull data, engine room data and energy efficiency data recorded by the ship during its historical operating cycle to form a multi-source heterogeneous dataset containing multiple dimensions; Data labeling steps: Based on the preset normal value range of each data in the multi-source heterogeneous dataset, the state of each sample in the multi-source heterogeneous dataset is judged. If any parameter in a sample exceeds its preset normal value range, the sample is marked as an abnormal state; if all parameters in the sample are within the normal value range, the sample is marked as a normal state, and a structured dataset with binary labels is generated. The class balancing process involves standardizing the structured dataset and dividing it into a training set and a test set according to a preset ratio. Then, a hybrid oversampling strategy, including random resampling and SMOTE oversampling, is used to augment the minority class samples in the training set whose sample count is less than a preset threshold, thereby achieving class balancing in the training set. Model training and hyperparameter optimization steps: Using CART decision trees as base learners, a random forest auxiliary decision model with classification function is constructed using the Bagging strategy; the random forest auxiliary decision model is trained using the training set after class balancing, and the hyperparameters of the random forest auxiliary decision model are optimized by combining random search optimization method and cross-validation method during the training process to obtain the trained and optimized random forest auxiliary decision model; the performance of the trained and optimized random forest auxiliary decision model is evaluated by the test set to obtain the random forest auxiliary decision model that meets the performance evaluation criteria; Intelligent Assisted Decision-Making Steps: Real-time collection of ship navigation data, hull data, engine room data, and energy efficiency data, followed by standardization processing similar to that in the category balancing step, and input into the random forest assisted decision-making model after performance evaluation has met the standards; The random forest assisted decision-making model determines the category to which the ship's current operating status belongs based on its built-in classification logic, and automatically generates and outputs assisted decision-making information including the category to which the ship's operating status belongs, the safety risk level, and abnormal warning information by combining the feature importance corresponding to the category. This information is then converted into visual graphics or charts through feature importance visualization technology, enabling real-time intelligent assisted decision-making for ships in complex operating scenarios.

2. The intelligent auxiliary decision-making method for ships in complex operating scenarios according to claim 1, characterized in that, In the class balancing process, a hybrid oversampling strategy is used to expand the minority class samples in the training set to achieve class balancing. Specifically, this includes: firstly, randomly resampling the minority class samples in the training set whose sample size is less than a preset threshold to expand their sample size; then, based on the random resampling, using SMOTE oversampling on all expanded minority class samples to generate synthetic samples, thereby increasing sample diversity and reducing the risk of overfitting; if SMOTE oversampling fails due to insufficient nearest neighbor samples, it automatically switches to random resampling to achieve class balancing.

3. The intelligent auxiliary decision-making method for ships in complex operating scenarios according to claim 1 or 2, characterized in that, The optimization of hyperparameters in the model training and hyperparameter optimization steps specifically includes: The RandomizedSearchCV (Randomized SearchCV) optimization method is used to perform multiple random searches and randomly select multiple hyperparameter combinations. For each selected hyperparameter combination, a 5-fold cross-validation method is used, with the model prediction accuracy as the optimization objective, to select the optimal hyperparameter combination.

4. The intelligent auxiliary decision-making method for ships in complex operating scenarios according to claim 1 or 2, characterized in that, The category balancing process includes standardizing the structured dataset by using the StandardScaler method to standardize the feature data in the structured dataset, excluding the binary labels, and converting each feature data into a standard form with a mean of 0 and a standard deviation of 1, so as to eliminate the dimensional differences between different parameters.

5. The intelligent auxiliary decision-making method for ships in complex operating scenarios according to claim 3, characterized in that, The model training and hyperparameter optimization steps specifically include evaluating the performance of the trained and optimized random forest-aided decision-making model using a test set, including: The performance of the trained and optimized random forest-assisted decision-making model is evaluated using a test set and based on multi-dimensional evaluation metrics, including classification report, confusion matrix, and area under the ROC curve. The classification report analyzes the classification accuracy of each class of samples, including precision, recall, and F1 score. The confusion matrix presents the matching between the predicted results and the true labels. The area under the ROC curve measures the model's generalization ability. When all multi-dimensional evaluation metrics reach their respective preset performance thresholds, a random forest-assisted decision-making model that meets the performance evaluation criteria is obtained.

6. An intelligent auxiliary decision-making system for ships in complex operating scenarios, characterized in that, It includes a multi-dimensional data acquisition module, a data annotation module, a class balancing module, a model training and hyperparameter optimization module, and an intelligent decision-making assistance module, which are connected in sequence. The multi-dimensional data acquisition module acquires navigation data, hull data, engine room data, and energy efficiency data recorded by the ship during its historical operating cycle, forming a multi-source heterogeneous dataset containing multiple dimensions. The data labeling module performs state judgment on each sample in the multi-source heterogeneous dataset according to the preset normal value range of each data in the multi-source heterogeneous dataset. If any parameter in a sample exceeds its preset normal value range, the sample is marked as an abnormal state; if all parameters in the sample are within the normal value range, the sample is marked as a normal state, and a structured dataset with binary labels is generated. The class balancing module standardizes the structured dataset and divides the standardized structured dataset into a training set and a test set according to a preset ratio. Then, it uses a hybrid oversampling strategy, including random resampling and SMOTE oversampling, to expand the minority class samples in the training set whose sample count is less than a preset threshold, so as to achieve class balancing of the training set. The model training and hyperparameter optimization module uses CART decision trees as base learners and employs a Bagging strategy to construct a random forest auxiliary decision model with classification capabilities. The random forest auxiliary decision model is trained using a training set after class balancing, and during training, the hyperparameters of the random forest auxiliary decision model are optimized using a random search optimization method and a cross-validation method to obtain a trained and optimized random forest auxiliary decision model. The performance of the trained and optimized random forest auxiliary decision model is then evaluated using a test set to obtain a random forest auxiliary decision model that meets the performance evaluation criteria. The intelligent auxiliary decision-making module collects real-time navigation data, hull data, engine room data, and energy efficiency data of the ship. After undergoing the same standardized processing as in the category balancing module, the data is input into the random forest auxiliary decision-making model after the performance evaluation has met the standards. The random forest auxiliary decision-making model determines the category to which the ship's current operating status belongs based on its built-in classification logic. Combining the feature importance corresponding to the category, it automatically generates and outputs auxiliary decision-making information, including the category to which the ship's operating status belongs, the safety risk level, and abnormal warning information. This information is then converted into visual graphics or charts through feature importance visualization technology, enabling real-time intelligent auxiliary decision-making for ships in complex operating scenarios.

7. The intelligent auxiliary decision-making system for ships in complex operating scenarios according to claim 6, characterized in that, In the class balancing module, a hybrid oversampling strategy is used to expand the minority class samples in the training set to achieve class balancing. Specifically, this includes: firstly, randomly resampling the minority class samples in the training set whose sample size is less than a preset threshold to expand their sample size; then, based on the random resampling, using SMOTE oversampling on all expanded minority class samples to generate synthetic samples, thereby increasing sample diversity and reducing the risk of overfitting; if SMOTE oversampling fails due to insufficient nearest neighbor samples, it automatically switches to random resampling to achieve class balancing.

8. The intelligent auxiliary decision-making system for ships in complex operating scenarios according to claim 6 or 7, characterized in that, The optimization of hyperparameters in the model training and hyperparameter optimization module specifically includes: A random search optimization method is used to perform multiple random searches and randomly select multiple hyperparameter combinations. For each selected hyperparameter combination, a cross-validation method is used with the model prediction accuracy as the optimization objective to select the optimal hyperparameter combination.

9. The intelligent auxiliary decision-making system for ships in complex operating scenarios according to claim 6 or 7, characterized in that, The standardization process for the structured dataset in the category balancing module specifically includes: using the StandardScaler method to standardize the feature data in the structured dataset, excluding the binary labels, and converting each feature data into a standard form with a mean of 0 and a standard deviation of 1, so as to eliminate the dimensional differences between different parameters.

10. The intelligent auxiliary decision-making system for ships in complex operating scenarios according to claim 8, characterized in that, The model training and hyperparameter optimization module specifically includes performance evaluation of the trained and optimized random forest-aided decision-making model using a test set, including: The performance of the trained and optimized random forest-assisted decision-making model is evaluated using a test set and based on multi-dimensional evaluation metrics, including classification report, confusion matrix, and area under the ROC curve. The classification report analyzes the classification accuracy of each class of samples, including precision, recall, and F1 score. The confusion matrix presents the matching between the predicted results and the true labels. The area under the ROC curve measures the model's generalization ability. When all multi-dimensional evaluation metrics reach their respective preset performance thresholds, a random forest-assisted decision-making model that meets the performance evaluation criteria is obtained.

Images