Machine learning-based fall-incident drift prediction and search feedback correction method

Abstract

The application belongs to the technical field of maritime search and rescue, and provides a method for predicting and searching for a drifting accident based on machine learning. According to the Leeway drift theory, the application re-expresses the drift process as a residual learning problem about displacement components, and adopts a lightweight gradient boosting machine to capture the nonlinear interaction between wind, ocean current and space-time factors by using multi-source heterogeneous data. In addition, the application also develops a search and rescue event triggered trajectory correction method based on an exponential weighted moving average. By using discrete detection events in search and rescue operations as feedback, the predicted trajectory is dynamically adjusted. A closed-loop "prediction-search-correction" framework is established, which effectively reduces the error accumulation over time.

Description

Technical Field

[0001] This invention belongs to the field of maritime search and rescue technology, and relates to the prediction of drift and search feedback correction for drowning accidents, specifically to a machine learning-based method for the prediction of drift and search feedback correction for drowning accidents. Background Technology

[0002] In recent decades, the rapid expansion of global maritime transport, offshore fishing, and marine energy development has led to a continuous increase in people falling overboard (MOB) accidents at sea. During maritime search and rescue (SAR) operations, drifting targets are constantly displaced due to the combined effects of wind and ocean currents. Their trajectories are further influenced by nonlinear wind-current interactions, the spatiotemporal heterogeneity of the marine environment, and random disturbances. These factors significantly increase the difficulty of accurate location and timely rescue. Given the critical rescue window of approximately 24 to 96 hours, reliable drift trajectory prediction is crucial for the success of search and rescue operations. For many years, drift trajectory prediction has been a core research topic in the fields of maritime search and rescue, navigation safety, and marine environmental monitoring. From a practical application perspective, prediction models not only need high spatial accuracy but also need to maintain robustness and computational efficiency in complex and ever-changing marine environments. The main challenge lies in predicting the future trajectory of a target based on a finite initial or last known position, while effectively considering the influence of uncertain environmental factors.

[0003] Existing research on drift prediction typically employs one of two methodological approaches: physics-based models and data-driven machine learning methods.

[0004] Physics-based models can clearly characterize the physical mechanisms controlling target motion, thus possessing strong interpretability. For example, the Leeway model proposed by Allen et al. is widely used; this model quantifies the impact of wind-induced drift relative to background currents. Building upon this model, Zhu et al. further demonstrated that decomposing wind-driven drift into downwind and crosswind components not only improves the model's interpretability but also facilitates parameter calibration through field experiments. Subsequent studies systematically examined the drift characteristics of various targets, including motorboats, lifeboats, life rafts, and small vessels.

[0005] However, physics-based models often rely on empirical parameters and simplifying assumptions, making them highly sensitive to the accuracy of environmental forecasts and limiting their ability to capture nonlinear dynamics in rapidly changing marine environments. Consequently, prediction errors tend to accumulate as the forecast timeframe extends; furthermore, the transferability of parameters across different regions and target types remains limited.

[0006] In recent years, researchers have increasingly explored machine learning methods to overcome the limitations of traditional approaches by learning the complex relationships between environmental forcing factors and drift trajectories from historical data. Several important studies have employed recurrent neural networks and convolutional neural networks for sea surface drift prediction, achieving higher prediction accuracy in data-rich environments. Furthermore, multimodal architectures and attention-based models have improved the representation of spatiotemporal features, thereby enhancing the performance of drift forecasting and synthetic aperture radar (SAR) applications.

[0007] While purely data-driven models possess powerful predictive capabilities, they often lack explicit physical constraints and exhibit limited robustness in situations with sparse data or those requiring extrapolation. Their performance is highly dependent on the quality and coverage of the training data, while factors such as regional characteristics, seasonal variations, and target-specific behaviors are often not adequately reflected. In such cases, predictions may contradict physical laws, thereby reducing the reliability of practical applications.

[0008] To bridge the gap between physical interpretability and data-driven flexibility, researchers have proposed hybrid modeling and correction strategies. Some studies focus on post-hoc correction of numerical model outputs or offline bias mitigation. For example, Zhou et al. identified key variables influencing drift behavior through sensitivity analysis and proposed a multi-feature coupled drift model whose parameters are optimized using a constrained metaheuristic algorithm. Dynamic correction methods based on neural networks or statistical filtering have also been explored, including NN-GA-ARMA-based drift correction and FCNN-embedded dynamic adjustment suitable for unmanned surface vessels. While these methods can reduce systematic bias to some extent, they typically rely on high-quality training samples and are primarily applicable to offline analysis. Furthermore, uncertainty quantification and Monte Carlo-based drift modeling have been used to characterize stochastic variability in the drift process. However, most existing studies employ a static prediction paradigm, where predicted trajectories are primarily used for initial search area delineation and are rarely updated. In continuous search and rescue operations based on real-time observations or search results, the integration of prediction models with search execution remains low, leading to accumulated errors that can negatively impact search path generation and resource allocation.

[0009] To address these limitations, this invention proposes a physically constrained gradient boosting framework that combines tolerance-based drift modeling with machine learning to predict the trajectory of moving targets. The drift process is reformulated as a displacement residual learning problem and modeled using an LGBM architecture to capture the nonlinear interactions between wind, ocean currents, and spatiotemporal factors. Furthermore, a search event-triggered exponentially weighted moving average (SET-EWMA) dynamic search and correction algorithm is developed. This algorithm iteratively updates SAR detection feedback to correct the predicted trajectory, thus constructing a closed-loop prediction-search-correction framework that effectively reduces trajectory deviation and improves SAR detection performance. Summary of the Invention

[0010] The purpose of this invention is to provide a machine learning-based method for predicting drift and correcting search feedback in cases of drowning accidents, applying the prediction model and feedback method to maritime search and rescue, and providing a new solution for predicting maritime rescue routes.

[0011] To achieve the above objectives, the technical solution of the present invention is as follows:

[0012] This invention provides a machine learning-based method for predicting drift in water accidents, the prediction method specifically including the following steps:

[0013] S1. Model Building: Build and train a physically constrained drift prediction model for a fall-in-water accident. This model is used to predict the future drift trajectory of the target based on its initial position and environmental data.

[0014] S2. Search path generation: Using the prediction information output by the drift prediction model, a search path is generated for the target at sea;

[0015] S3. Feedback Correction: During the search process, discrete events detected in the search are used as feedback to dynamically correct the subsequent predicted trajectory of the drift prediction model;

[0016] S4. Replanning: Based on the corrected predicted trajectory, replan and execute the search path to form a closed loop of "prediction-search-correction-replanning".

[0017] Preferably, the model construction specifically includes:

[0018] S1.1. Dataset Construction and Preprocessing: Integrate multi-source heterogeneous data, including actual or simulated drift trajectory data of targets falling into the water, as well as corresponding wind and ocean current environmental field data, extract spatiotemporal features, and construct a feature matrix containing environmental variables, spatial location, regional type and seasonal information;

[0019] S1.2. Based on Leeway drift theory, a physical baseline model is established, and the observed displacement of the target drift is decomposed into physically interpretable displacement components based on wind and current and a residual term.

[0020] S1.3. A machine learning algorithm is used to train the model, with the aforementioned feature matrix as input and the aforementioned residual term as the learning target, in order to predict the drift bias that the physical baseline model fails to capture.

[0021] S1.4. Combine the output of the physical baseline model with the residual terms predicted by machine learning to obtain the final drift displacement prediction.

[0022] Preferably, the machine learning algorithm is a lightweight gradient booster.

[0023] Preferably, the search path generation includes:

[0024] S2.1. Based on the output of the above drift prediction model, the Monte Carlo method is used to sample the predicted position to generate a set of particles representing the uncertainty of the target position;

[0025] S2.2. Perform kernel density estimation on the particle set to generate a probability density field for the existence of the target;

[0026] S2.3. Perform spatial clustering on the particle set, identify high-probability regions, and use the cluster centers as candidate search nodes;

[0027] S2.4. Based on the probability of the search node and its distance from the current search position, a series of nodes are selected using a greedy algorithm to form an executable search path.

[0028] Preferably, the correction method is dynamic correction, which is a search event-triggered exponentially weighted moving average trajectory correction method, including:

[0029] S3.1. During the search time step When a target is detected, record the observation position. With model predicted location Calculate the detection residual ;

[0030] S3.2. Using the exponentially weighted moving average formula Update system bias estimate ,in This is an estimate of the prior bias before the update. For smoothing coefficients;

[0031] S3.3. In subsequent time steps In this case, the estimated deviation remains unchanged, i.e. ;

[0032] S3.4. Perform real-time correction on subsequent model-predicted locations: ,in This is for correcting the gain coefficient.

[0033] Preferably, the replanning specifically includes: based on the corrected drift trajectory, re-execute step S2, update the probability density field and candidate nodes of the search path, and drive the search platform to search along the new path. If the target is found again, the dynamic correction process of step S3 is triggered again, and so on until the task ends.

[0034] The present invention also provides the application of the above-described prediction method in predicting and searching for people who have fallen into the water or other passively drifting targets.

[0035] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the program to implement the steps of the prediction method described above.

[0036] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the prediction method described above.

[0037] The beneficial effects of this invention are:

[0038] 1. By integrating Leeway-based physical mechanisms into the LGBM learning framework, a physically constrained gradient boosting drift prediction model is developed, which can achieve stable and reliable drift prediction in complex marine environments.

[0039] 2. The study incorporated environmental classification features (including prediction area and season) to enhance the model's generalization ability across different sea areas; at the same time, the Shap analysis method was used to improve the model's interpretability.

[0040] 3. This paper proposes a dynamic search correction algorithm based on EWMA and triggered by search events, which can update the predicted trajectory in real time using search feedback. This algorithm forms a closed-loop prediction-search-correction mechanism, effectively improving the reliability of prediction and the success rate of subsequent searches. Attached Figure Description

[0041] Figure 1 This is a flowchart of the water drift prediction and search feedback correction method of the present invention;

[0042] Figure 2 This is a structural diagram of the Hybrid-LightGBM drift prediction model of the present invention;

[0043] Figure 3This is a schematic diagram illustrating the probability search region and candidate node generation of the present invention;

[0044] Figure 4 This is a schematic diagram comparing the drift trajectory prediction results of the present invention;

[0045] Figure 5 This is a schematic diagram of the search feedback trajectory correction of the present invention. Detailed Implementation

[0046] Unless otherwise specified, the experimental methods used in the following examples are conventional methods.

[0047] Unless otherwise specified, all materials and reagents used in the following examples are commercially available.

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0049] Example: Construction of Drift Trajectory Prediction Model

[0050] This paper proposes a physically constrained gradient boosting framework that combines tolerance-based drift modeling with machine learning to predict the trajectory of moving targets. The overall process is as follows: Figure 1 As shown.

[0051] 1. Method

[0052] 1.1 Drift Trajectory Prediction Model

[0053] This drift trajectory prediction model combines a physics-based base model with a machine learning residual learner. This integrated architecture helps predict potential drift locations in future time periods. Training data includes human model drift trajectories obtained from field experiments, actual human drift data from accident reports, and corresponding spatiotemporal environmental forcing fields (such as wind fields and ocean currents).

[0054] 1.1.1 Dataset Construction and Preprocessing

[0055] To construct the key variables required for model training, this embodiment integrates multi-source heterogeneous data (including meteorological observation data, accident report data, and human drift experiment trajectories) to build a training dataset for drift prediction. The human drift trajectory data comes from the public dataset of the 2021 "Digital China Innovation Competition" (https: / / www.datafountain.cn).

[0056] The Marine Forecasting Center provides meteorological observation data, covering core environmental variables such as wind fields. To generate samples that highly simulate actual SAR (Marine Radar) operating conditions, this embodiment also uses simulation samples from the Maritime Safety Administration of the People's Republic of China.

[0057] Given the significant differences between trajectory data and the wind / flow NetCDF product in terms of spatiotemporal resolution, sampling frequency, and coordinate conventions, systematic preprocessing and feature engineering are necessary to transform heterogeneous data into a feature matrix suitable for regression analysis. This embodiment unifies the data format and resolution, extracts environmental features and experiment-related variables (such as region type and season), and constructs a scalable feature matrix for model training.

[0058] The data processing flow is summarized as follows:

[0059] 1) Spatiotemporal interpolation of environmental forcing factors:

[0060] Using the spatiotemporal coordinates of each trajectory point as an index, wind force and current data within a ±2-hour time window surrounding each data point are extracted. A two-dimensional spatial interpolation scheme is employed to achieve continuous environmental characteristics, ensuring spatiotemporal consistency between the trajectory data and the environmental field.

[0061] 2) Eliminate terrain-restricted tracks:

[0062] The system identifies and excludes tracks originating from predefined constrained areas (such as land, nearshore zones, narrow waterways, and areas surrounding small islands). These tracks are significantly influenced by topography and fail to fully reflect the drift dynamics driven by wind and ocean currents.

[0063] 3) Displacement label construction and smoothing:

[0064] The longitude and latitude displacements between trajectory points are calculated at 5-minute intervals and used as learning targets. A rolling time window is used to smooth the wind speed, flow velocity, and displacement signals to reduce high-frequency noise and highlight stable drift patterns.

[0065] 4) Scale correction and feature selection:

[0066] Longitudinal displacement is adjusted by latitude-dependent scaling to ensure physical consistency. The zonal and meridional components of the wind field and ocean current at the current time step are selected as core features, supplemented by auxiliary variables for model training. After preprocessing, the generated feature matrix is ​​input into the drift prediction framework. In addition to basic environmental variables, categorical variables such as region type and season are included to distinguish differences between different drift mechanisms. Under different environmental conditions, a unified feature matrix is ​​ultimately formed to predict future drift locations. The overall technical path is as follows: Figure 2 As shown.

[0067] 1.2 Application of Physics-Based Modeling Techniques in the Leeway Framework

[0068] In maritime search and rescue operations, people who have fallen overboard are considered passively drifting targets lacking autonomous propulsion. Their motion is primarily influenced by the combined effects of background surface ocean currents and wind-induced drift. To maintain the interpretability of the physical meaning and avoid the limitations of purely data-driven methods, this embodiment develops a physically constrained drift prediction model based on the Leeway framework and further optimizes the model using residual learning techniques. According to Leeway theory, at a given time step t, the instantaneous velocity of a drifting target can be expressed as the sum of a current-driven component and another component, namely, the lateral drift component caused by wind. For small targets (such as people who have fallen overboard), the lateral drift component caused by crosswinds is usually assumed to be statistically symmetric, with a mean close to zero at the overall level. Therefore, only the contribution from the downwind direction is retained, resulting in the following basic physical model:

[0069] (1) (2) (3)

[0070] Equations (1)-(3) establish the physical drift baseline. Among them, Indicates the velocity of surface ocean currents. This represents a wind speed of 10 meters per second, while and These are empirical offset coefficients. These parameters are empirically calibrated using drift observation data. The process involves subtracting the surface ocean current velocity from the observed velocity to obtain the offset velocity, followed by linear regression analysis of the obtained velocity against the 10-meter wind speed in both leeward and crosswind coordinate systems. Constrained and unconstrained regression methods are used to determine the slope, intercept, and standard error. Based on the baseline velocity, the time interval... The physical displacement within can be expressed as:

[0071] (4) (5) (6)

[0072] Equations (4)-(6) decompose the observed displacement into two parts: a physically interpretable component and a residual term. This residual term reflects nonlinear effects and environmental variability that the Leeway model fails to capture. This machine learning model does not directly estimate the total displacement but is specifically designed to predict the residual displacement. This ensures physical consistency. During residual displacement modeling, the following feature vector is constructed at each time step:

[0073] (7) (8)

[0074] in and These represent the zonal and meridional components of wind speed and current speed, respectively. It includes auxiliary contextual features such as spatial location, season, and region classification; and For regression trees; For the number of trees, This represents the learning rate. The predicted displacement is calculated by integrating the residuals obtained from the learning along the longitude and latitude directions over the physical baseline:

[0075] (9)

[0076] This method ensures that the machine learning model operates as a physically constrained residual corrector, rather than as an unconstrained black-box predictor. Multiple machine learning algorithms are trained and evaluated to determine the most effective model for the drift prediction module.

[0077] 1.3 Machine Learning Training and Prediction

[0078] During the training phase, this embodiment evaluated four candidate algorithms: Support Vector Machine (LGBM), Support Vector Machine (SVM), Random Forest (RF), and XGBoost (XGB). All models were trained using cross-validation and their parameters were tuned via grid search to ensure stable and reproducible performance evaluation results. Ultimately, the model with the best prediction performance was selected as the core predictor for the drift prediction module.

[0079] To quantitatively evaluate the accuracy of drift trajectory prediction, three performance metrics were used: root mean square error (RMSE), mean absolute error (MAE), and normalized cumulative separation distance (NCSD). These metrics are used together to assess point-by-point position accuracy and overall trajectory consistency. RMSE and MAE use Havers distance to determine the spatial difference between the predicted and observed positions, taking into account the effect of spatial curvature. Conversely, NCSD assesses the cumulative deviation between the predicted and observed trajectories. The relevant metrics are defined as follows:

[0080] (10) (11) (12)

[0081] In these equations, and These represent the observed target position and the predicted target position at time step t, respectively. This represents the total number of prediction steps. RMSE is more sensitive to larger prediction errors, while MAE reflects the average absolute value of the position error. NCSD measures the ratio of the cumulative separation distance between the predicted and observed trajectories to the total length of the observed trajectories, thus focusing more on long-term trajectory consistency than point-by-point accuracy. The Haversey distance function is defined as follows:

[0082] (13)

[0083] In this equation, and These represent the latitudes of the observed and predicted locations, respectively. and These represent the corresponding longitudes, while r represents the Earth's radius.

[0084] 1.4 Error Correction Based on Search Feedback

[0085] 1.4.1 Search Path Generation

[0086] In maritime search and rescue operations, search path planning based on drift uncertainty is crucial within the "prediction-search-correction" closed-loop framework. To accurately simulate the spatial uncertainty of predicting drift positions, this embodiment employs a Monte Carlo particle sampling strategy at each prediction step. This strategy includes probability field construction, cluster analysis, and generating search paths using a greedy algorithm. Figure 3 ).

[0087] To quantify the spatial uncertainty of the predicted position at each time step t, this embodiment generates N particles by applying a random perturbation to the average predicted position. These particles represent samples of the target's potential spatial distribution. A continuous probability field is then constructed using kernel density estimation (KDE), where the density at spatial position x = (x, y) is defined by the following equation:

[0088] (14)

[0089] Here, h is a bandwidth parameter that controls the smoothness of the final density field: a smaller h leads to more concentrated peaks, while a larger h produces a smoother distribution. The generated KDE field ft(x, y) serves as a probability basis to guide the generation of search nodes in subsequent stages.

[0090] To translate the continuous probability density field into an executable search decision, the particle cloud needs to be spatially clustered to identify multiple high-probability regions. The centroids of these clusters are designated as search nodes:

[0091] (15)

[0092] Here, This represents a subset of particles belonging to cluster j, while This represents the corresponding probability weight. These search nodes and their weights are then passed to the path planning module.

[0093] At time step t, the search platform selects a set of search nodes to construct the path:

[0094] (16)

[0095] in, This represents the selection score of candidate node j. d represents the current search position, d is the Euclidean distance, and Y > 1 is a penalty exponent used to impose penalties on nodes that are farther away.

[0096] To meet the requirements for real-time responsiveness and stability in dynamic SAR scenarios, this embodiment employs a greedy heuristic algorithm: in each iteration, selects... The highest-scoring unvisited node. The observations are then used to update the distribution model in Example 1.4.2.

[0097] 1.4.2 Trajectory Correction Based on Search Feedback

[0098] To address long-term prediction errors caused by uncertain environmental factors and modeling inaccuracies, this embodiment proposes a trajectory correction mechanism utilizing dynamic search observation feedback. At each time step t, this mechanism adjusts the observed target position according to the research assumption. With predicted location There is an additive bias term. and residual noise Impact:

[0099] During the search process, time When a target is detected, the model records its observed position xotbk s. Let xtk represent the corresponding model-predicted position. The final detection residual is defined as:

[0100] (17)

[0101] Here, This is a potential deviation state Indirect observations indicate that this state is affected by location noise. The disturbance.

[0102] To reduce the noise sensitivity caused by a single residual and to make full use of the time correlation in the evolution of the bias, the exponentially weighted moving average (EWMA) method was adopted.

[0103] This method employs a recursive strategy to estimate the offset. It has been proven effective in online bias correction and time series forecasting.

[0104] (18) (19)

[0105] in This represents the prior estimate before the update. The bias estimate remains constant over time steps between adjacent detection events. Finally, the predicted location is corrected in real-time before being used for subsequent planning. In all cases, the adjusted predicted location is:

[0106] (20)

[0107] The gain β ∈ [0,1] determines the strength of the correction. Specifically, to consider the dependence of the search feedback, two different cases are defined: if a target is detected, the offset is updated; if no target is detected, the original offset is maintained and the correction step is skipped.

[0108] (21)

[0109] This update strategy employs limited search feedback to correct deviations at the trajectory level, effectively reducing error accumulation while maintaining spatial consistency of the predicted path. It is important to emphasize that this correction mechanism is not for post-hoc trajectory fitting, but is essentially an online adjustment strategy based on search feedback, designed to enhance the practical applicability of drift prediction in dynamic path planning and selective search decision-making. This forms a tightly coupled closed-loop process of prediction-search-correction-replanning.

[0110] Application Example Model Validation

[0111] The proposed MOB drift prediction framework was comprehensively validated through a multi-dimensional experimental design. In the trajectory prediction stage, standard statistical indicators were used to quantitatively evaluate the spatiotemporal accuracy of the predicted trajectory by comparing it with different prediction models. In the trajectory correction stage, various scenarios were designed to evaluate the changes in prediction accuracy before and after correction. Finally, the effectiveness of the model was evaluated based on real drift trajectories and corresponding search results, thus providing an application-oriented evaluation basis for the proposed method.

[0112] 1. Experimental Setup

[0113] The MOB drift prediction model developed in this embodiment uses a publicly available dataset from the "Digital China Innovation Competition" (Smart Ocean Track). This dataset contains drift trajectory data of 35 instrumented virtual targets deployed in the East China Sea and South China Sea during the four seasons of 2019-2020. The position data of each target was recorded at minute intervals, resulting in 82,699 data points for 35 complete drift trajectories. Following the data preprocessing procedure detailed in Embodiment 1.1.1, the data granularity was standardized, and a feature matrix was constructed. To ensure the objectivity of the model evaluation, the dataset was divided into a training set (70%) and a test set (30%) in chronological order. All candidate algorithms underwent hyperparameter optimization and were evaluated using cross-validation to ensure the robustness and repeatability of the models.

[0114] According to China's marine meteorological zoning operational standards, coastal waters are divided into 18 marine regions. This embodiment selects four representative regions for case verification: the East China Sea, the South China Sea, the Yellow Sea, and the Bohai Sea. Of the 35 available drift tracks, 8 representative tracks were selected for detailed analysis.

[0115] In the actual search and rescue correction simulation experiment, this embodiment selected real accident cases from the accident report.

[0116] The simulation experiment set the search and rescue operation window from 18:30 to 20:30. The search and rescue platform parameters were configured according to the actual equipment type.

[0117] The performance specifications adopted by the East China Sea Search and Rescue Bureau ensured consistency between the simulated scenarios and actual search and rescue operations. All experiments were conducted on personal computers equipped with Intel® Core™ Ultra 7 155H processors and 32GB of memory.

[0118] Model training and evaluation were performed using the scikit-learn library within the Visual Studio Code environment. Within each discrete time step, the predicted position and trajectory of the target were updated using the drift prediction model. Candidate search nodes were generated using particle-based simulation and probabilistic clustering methods, and search paths were constructed based on node positions and their respective occurrence probabilities. When the drifting target entered the effective sensing radius of the search platform, the system determined successful detection and triggered the trajectory correction algorithm.

[0119] 2. Comparison of Predictive Performance

[0120] This section presents an experimental analysis of the trajectory prediction performance. First, multiple machine learning models were used for prediction, and the model with the best prediction performance was determined by comparing key indicators. Then, the performance of the model in different sea areas and months was evaluated through a generalization study. The interpretability of the machine learning method was explored using Shap analysis. Finally, the impact of wind, waves, and ocean currents on drift trajectory prediction in a specific time period was analyzed through a case study.

[0121] The machine learning models used in this study include LGBM, RF, XGBoost, and Support Vector Machine (SVM). The blue trajectory represents the actual drift path of a man who fell into the water. Among them, the LGBM model consistently achieved the highest spatial consistency with the actual drift path. Compared with other models, LGBM can accurately capture the observed overall drift shape and directional trend; its deviations in latitude and longitude are consistently smaller than other methods. Although models such as RF can sometimes approximate the overall drift trend, they often have significant positional errors or fail to capture subtle changes in the trajectory.

[0122] Table 1 presents the comparative analysis results across multiple sea areas. The data shows that LGBM exhibits stable and excellent performance in all regions, with consistently low MAE, RMSE, and NCSD values. This fully demonstrates the method's strong generalization ability and robustness in diverse marine environments.

[0123] Although the SVM has a low NCSD in the Yellow Sea, the LGBM maintains a balanced and reliable accuracy overall, making it the most suitable core model for cross-regional sea-level rise prediction.

[0124] Table 1 Performance of various machine learning models

[0125]

[0126] This embodiment designed a comparative experiment to compare the Leeway model, the LGBM model, and the physically enhanced hybrid LGBM model. The parameters of the Leeway model are derived from physical drift experiments conducted in different sea areas. Coefficients derived from these studies, which included experimental data, were applied to the corresponding sea areas and trajectories. As shown in Table 2, the prediction results of both the independent LGBM model and the physically enhanced hybrid LGBM model are more stable, with the latter exhibiting a smaller prediction error. Specifically, within the four prediction time windows of 12 to 48 hours, the physically enhanced hybrid LGBM model not only had the lowest prediction error but also demonstrated the best overall performance. Figure 4 ).

[0127] Table 2 Performance of various machine learning models

[0128]

[0129] To analyze the response mechanisms of various features in the LGBM drift prediction model, this embodiment conducted a Shap analysis. The results show that wind direction and speed are the most influential features in longitude and latitude predictions, confirming that wind field is the main factor affecting ocean drift, consistent with Leeway's theory. Initial location is also significantly important, indicating spatial variability in the flow field and the interaction between wind field and ocean current conditions in different sea areas.

[0130] Furthermore, the current velocity and direction have a significant impact on the prediction results, particularly affecting latitudinal changes by primarily controlling drift velocity and direction adjustments. Time-related features provide auxiliary corrections, capturing the influence of tides and diurnal variations on the drift process. These findings are consistent with the conclusions of several other studies.

[0131] Between 12 and 18 hours, the ocean current direction gradually shifted from south-north to east-west, and the current velocity in this region continued to decrease. All three methods captured this trend; however, the Leeway physics model exhibited larger prediction errors and poorer performance. In contrast, the hybrid model corrected for some of the inherent errors of the purely data-driven LGBM model, thus more accurately reflecting the actual impact of ocean currents on the drift position.

[0132] Trajectory Prediction: Over 12 to 24 hours, the direction angle of the northeasterly wind gradually decreases, while the wind speed remains nearly constant. During this phase, significant differences begin to emerge in the prediction results of different models, especially in the longitudinal dimension. This phenomenon illustrates how changes in environmental factors affect the predictive performance of different models and further validates the conclusions drawn from the Shap analysis: longitudinal changes are more sensitive to wind direction.

[0133] 3. Trajectory Correction Comparison

[0134] To evaluate the effectiveness of the dynamic search correction mechanism proposed in this embodiment, a series of controlled experiments were conducted. These experiments aimed to test the performance of the mechanism under different correction parameters, different personnel sizes in search and rescue scenarios, and various correction strategies. The search and rescue facility used in this experiment was the Donghai Rescue Bureau's 311 small maritime rescue vessel. This vessel has a maximum speed of 35 knots, a propulsion system consisting of two MAN-D2862 engines (each with a power output of 1400 horsepower), and a maximum rescue capacity of 40 people. The vessel's coverage area was set at 1000 meters. Personnel in the water were classified as non-motorized, and their drift was only driven by wind and ocean currents.

[0135] In actual SAR (Search and Rescue) operations, once the search equipment locates a moving object (MOB), it adjusts its subsequent trajectory based on the search location, thereby enabling faster location of other nearby events. The trajectory correction method proposed in this embodiment demonstrates higher effectiveness as the number of MOB events increases. Furthermore, due to the large number of events successfully detected in the middle of the search process, the correction error of this method is minimized during this stage. Figure 5 ).

[0136] In large-scale drowning incidents, employing a multi-detection correction strategy and implementing actual SAR operations along the corrected trajectory can extend the entire time window required for successful rescue. This advantage is particularly significant in the mid-to-late stages of search and rescue operations. These results demonstrate that dynamic correction effectively improves trajectory accuracy, while search simulation can pinpoint target locations earlier and more accurately within the same timeframe, thereby enhancing the overall efficiency of search and rescue operations.

[0137] In summary, this model combines physical principles with machine learning techniques. It utilizes high-resolution, multi-source heterogeneous data and is optimized using the EWMA algorithm. Predictive trajectory analysis demonstrates, as confirmed by the case study, that this model significantly improves the predictive accuracy of maritime search and rescue (SAR) operations. Among the machine learning algorithms employed, the LGBM model performs best—achieving optimal levels in all evaluation metrics (including mean absolute error (MAE), mean absolute error (RMSE), and unconditional standard deviation (NCSD) both before and after the correction phase. The dynamic search feedback-driven correction mechanism significantly enhances the model's ability to reduce prediction errors, thereby optimizing the handling of maritime oil spills and improving the overall efficiency of search and rescue operations. Notably, the model exhibits good generalization ability under different sea areas and seasonal conditions. The proposed trajectory correction strategy not only improves prediction accuracy but also seamlessly integrates with search and rescue procedures.

[0138] The above-described embodiments are merely preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A machine learning-based method for predicting drift in water accidents, characterized in that, The prediction method specifically includes the following steps: S1. Model Building: Build and train a physically constrained drift prediction model for a fall-in-water accident. This model is used to predict the future drift trajectory of the target based on its initial position and environmental data. S2. Search path generation: Using the prediction information output by the drift prediction model, a search path is generated for the target at sea; S3. Feedback Correction: During the search process, discrete events detected in the search are used as feedback to dynamically correct the subsequent predicted trajectory of the drift prediction model; S4. Replanning: Based on the corrected predicted trajectory, replan and execute the search path to form a closed loop of "prediction-search-correction-replanning".

2. The prediction method according to claim 1, characterized in that, The model construction specifically includes: S1.

1. Dataset Construction and Preprocessing: Integrate multi-source heterogeneous data, including actual or simulated drift trajectory data of targets falling into the water, as well as corresponding wind and ocean current environmental field data, extract spatiotemporal features, and construct a feature matrix containing environmental variables, spatial location, regional type and seasonal information; S1.

2. Based on Leeway drift theory, a physical baseline model is established, and the observed displacement of the target drift is decomposed into physically interpretable displacement components based on wind and current and a residual term. S1.

3. A machine learning algorithm is used to train the model, with the aforementioned feature matrix as input and the aforementioned residual term as the learning target, in order to predict the drift bias that the physical baseline model fails to capture. S1.

4. Combine the output of the physical baseline model with the residual terms predicted by machine learning to obtain the final drift displacement prediction.

3. The prediction method according to claim 2, characterized in that, The machine learning algorithm is a lightweight gradient booster.

4. The prediction method according to claim 1, characterized in that, The search path generation includes: S2.

1. Based on the output of the above drift prediction model, the Monte Carlo method is used to sample the predicted position to generate a set of particles representing the uncertainty of the target position; S2.

2. Perform kernel density estimation on the particle set to generate a probability density field for the existence of the target; S2.

3. Perform spatial clustering on the particle set, identify high-probability regions, and use the cluster centers as candidate search nodes; S2.

4. Based on the probability of the search node and its distance from the current search position, a series of nodes are selected using a greedy algorithm to form an executable search path.

5. The prediction method according to claim 1, characterized in that, The correction method is a dynamic correction, specifically a search event-triggered exponentially weighted moving average trajectory correction method, including: S3.

1. During the search time step When a target is detected, record the observation position. With model predicted location Calculate the detection residual ; S3.

2. Using the exponentially weighted moving average formula Update system bias estimate ,in This is an estimate of the prior bias before the update. For smoothing coefficients; S3.

3. In subsequent time steps In this case, the estimated deviation remains unchanged, i.e. ; S3.

4. Perform real-time correction on subsequent model-predicted locations: ,in This is for correcting the gain coefficient.

6. The prediction method according to claim 1, characterized in that, The replanning specifically includes: based on the corrected drift trajectory, re-execute step S2, update the probability density field and candidate nodes of the search path, and drive the search platform to search along the new path. If the target is found again, the dynamic correction process of step S3 is triggered again, and so on until the task ends.

7. The application of the prediction method according to any one of claims 1-6 in predicting and searching for people who have fallen into the water or other passively drifting targets.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the prediction method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the prediction method as described in any one of claims 1-6.

Images