Instantaneous Production Prediction and Control-Parameter Optimization in Trailing Suction Hopper Dredgers Based on Multi-Model Feature Fusion.
1. IntroductionThe continuous growth of global maritime traffic has made port modernization and expansion inevitable, with dredging as a key enabler [1]. As a core component of marine infrastructure, port expansion, channel maintenance, and land reclamation, dredging enhances port throughput and navigational safety, while supporting regional economic development and national maritime strategies [2]. Under the "Belt and Road" initiative and increasing interconnection of global ports, its role in port clustering and deep-water development has become more prominent [3]. TSHDs extract sediment via dragheads, transport it in onboard hoppers, and dispose of it with a high degree of automation and efficiency. However, increasingly complex construction conditions--such as hard bedrock, high-silt flows, and variable hydrodynamics--challenge traditional control methods, which struggle to meet demands for high productivity and low energy consumption [4]. Meanwhile, the industry is aligning with the United Nations Sustainable Development Goals by integrating green dredging, intelligent technologies, and social benefits [5].
Efficient and stable dredging operation reduces equipment wear, operating time, and energy consumption, thereby accelerating construction and lowering costs [6][7]. TSHD production capacity, typically defined as effective dredged volume per unit time, is affected by coupled environmental conditions (wave height, current velocity, water depth, sediment distribution), equipment parameters (engine power, suction pipe size, hopper structure), and operational strategies (draghead depth, sailing speed, pump speed) [8]. Conventional prediction methods based on empirical formulas or physical models rely on simplifying assumptions and struggle with nonlinearities, leading to limited applicability and notable prediction errors [9]. Moreover, empirical models often ignore the intricate coupling between soil conditions and operational strategies, constraining their ability to support dynamic optimization [10].
The deployment of intelligent sensors and advances in industrial data acquisition have produced large volumes of heterogeneous TSHD operation data, including engine parameters (speed, fuel consumption, load), process variables (suction pipe pressure, slurry density, pump speed), and environmental factors (wave height, water depth, soil density) [11][12]. These multi-dimensional datasets capture complex interactions among environment, equipment load, and sediment properties, providing a solid basis for accurate capacity modeling and optimization. With the widespread adoption of sensors in construction, data-driven approaches are gradually replacing purely experience- or physics-based models, enabling intelligent and automated operations [13]. This trend is also evident in dredging, where effective data fusion and analysis significantly improve TSHD efficiency and safety under complex conditions. Artificial intelligence--particularly machine learning and deep learning--has shown promise in productivity prediction, condition recognition, and energy optimization: Li et al. [6] employed a Deep Belief Network for high-dimensional parameter prediction and feature extraction; Bai et al. [8] proposed a hybrid XGBoost-SVR model to identify and model productivity-related factors; Chen et al. [14] developed a multi-objective framework integrating productivity and energy efficiency, while Huang et al. [15] used a Barabasi-Albert-based and GA-optimized LightGBM model for ship power system fault diagnosis, offering insights for TSHD intelligent maintenance. Zhang et al. [16] combined digital twins with deep learning via multi-similarity and gradient fusion for bearing fault detection, demonstrating the potential of digital-twin-based predictive maintenance. For Cutter Suction Dredgers, Liu et al. [17] showed that machine-learning and global search-based parameter optimization can markedly improve productivity and energy efficiency, underscoring the broader potential of Al-driven dredging control. More recently, Chen et al. [18] proposed a cluster-wise deep learning framework for scenario-adaptive energy consumption and productivity prediction in cutter suction dredgers. By integrating unsupervised clustering with multivariate time series forecasting, their approach demonstrates the importance of adapting predictive models to varying operational scenarios--a concept that aligns with the adaptive optimization strategy pursued in the present study. In parallel with these advances, researchers have also explored the integration of physical constraints into data-driven models to improve prediction robustness. Liu et al. [19] proposed a physics-assisted deep learning (PADL) model for cutter suction dredger productivity prediction, incorporating a physics-assisted loss function (PALF) that combines data-driven and physics-driven objectives. Their results demonstrate that embedding domain knowledge into the loss function significantly improves prediction accuracy, particularly in small-data scenarios. This "data + knowledge" fusion philosophy aligns closely with the multi-model feature fusion and engineering-guided optimization strategies adopted in the present study.
Related work also addresses draghead trajectory optimization, cutter state recognition, and system-level coordination. Li et al. [20] proposed a real-time monitoring-based efficiency evaluation method for CSDs, enabling quantitative assessment and optimization of operational efficiency to guide trajectory adjustment and scheduling. Chen et al. [21] further applied clustering-based deep learning to predict soil dredging difficulty, improving the representation of cutter-soil interactions and supporting operational strategy optimization. These studies suggest that TSHD productivity should be viewed as a coupled outcome of equipment, environmental, and operational factors.
To tackle the limitations of traditional dredging--unmodeled nonlinearities, limited prediction accuracy, and suboptimal control under complex conditions--recent research has increasingly combined data-driven modeling with intelligent optimization. Zhou et al. [22] used real-time monitoring to estimate mean particle size, sand bed height, and quality in TSHD operations, enabling online environmental sensing and providing reliable inputs for capacity modeling and efficiency evaluation. Su et al. [23] employed mathematical models and CFD simulation to analyze hopper sedimentation, revealing sediment-flow coupling and improving predictive accuracy beyond empirical approaches. Qi et al. [24] investigated wave compensator and hull failure modes under mid- and long-period waves, highlighting the importance of system-level risk assessment and structural optimization for TSHD safety and stability.
Despite these advances, several challenges remain. First, feature extraction lacks a unified framework: most studies rely on single-source data, with limited multimodal fusion of sensor, log, and environmental information, resulting in underutilization of available signals in complex scenarios. Second, optimization is often single-stage and heuristic, improving local performance but lacking fine-grained sensitivity adjustment and comprehensive global search, which restricts adaptability to highly dynamic operations. Third, issues of generalization and computational efficiency persist: models may perform well on training datasets but degrade under different soils and operating regimes, while complex algorithms impose high computational costs in large-scale and real-time applications, limiting engineering deployment. These gaps indicate that breakthroughs in cross-modal feature fusion, layered optimization, and efficient computational frameworks are needed for large-scale intelligent dredging.
In this context, this paper proposes an intelligent TSHD productivity regulation method based on multi-model feature extraction and co-evolutionary optimization. The main contributions of this study are threefold:
(1) A dynamically weighted multi-model feature selection framework that integrates Lasso regression, Random Forest (RF), and LightGBM to identify key operational parameters influencing TSHD productivity. Unlike conventional single-source feature extraction methods, this approach leverages the complementary strengths of linear regularization and ensemble learning, effectively eliminating redundant variables while preserving both linear constraints and nonlinear interactions.
(2) A high-precision production prediction model based on Bayesian Optimization-optimized XGBoost (BO-XGBoost). By employing Bayesian Optimization for hyperparameter tuning, the model achieves superior predictive stability and generalization capability compared to traditional XGBoost and other machine learning baselines, addressing the limitations of manual parameter tuning in complex dredging environments.
(3) A two-stage feature-guided co-evolutionary optimization strategy that couples Feature-guided Sequential Univariate Optimization (FSUO) with a Constrained Grey Wolf Optimizer (CGWO). This "coarse-to-fine" framework first performs sensitivity-based coarse screening and then refines the solution space under engineering constraints, significantly improving search efficiency and solution feasibility. To the best of our knowledge, this is the first study to integrate such a hierarchical optimization approach with feature-guided mechanisms for TSHD operational parameter tuning.
This framework enhances adaptive configuration of TSHD operational parameters and offers a generalizable solution for intelligent modeling and optimization of complex engineering systems.
2. TSHD system and basic construction process
A TSHD is a large self-propelled dredger equipped with a draghead excavation system and hydraulic suction. Compared with other dredgers, TSHDs combine self-propulsion, self-dredging, and self-unloading, offering superior seakeeping and suitability for open ports with long approach channels. Key subsystems include the draghead, dredge pump, hopper, and propulsion [25][23]. The draghead directly interfaces with the seabed, so its design critically affects excavation efficiency and soil-breaking capacity; variants such as adaptive or twin dragheads can be configured to match soil conditions and construction goals. The dredge pump generates the negative pressure needed to convey slurry through the suction pipe to the hopper; power, head, and delivery efficiency are core performance metrics. The hopper stores the slurry; its capacity sets per-cycle yield and thus impacts overall efficiency and cycle time. The propulsion system provides power for both transit and dredging, ensuring stable operation and precise maneuvering in complex marine environments.
Operating principle and cycle stages:
Stage I: Loading. Dredging begins and the hopper fills with the water-sediment mixture. Stage II: Overflow loading. Excess water is discharged via the overflow once the hopper is full.
Stage III: Deck clearing. Residual deck water is removed to prepare for sailing.
Stage IV: Transport. The vessel transits from the dredging site to the disposal area.
Stage V: Discharge. Dredged material is released via bottom doors or high-pressure jetting.
Stage VI: Return voyage. The empty vessel returns to the dredging site to start a new cycle.
3. Monitoring features and data collection
Data acquisition and preprocessing form the foundation of this study. This section first outlines the data collection process and then describes in detail the preprocessing steps applied to the raw data. The Xin Hailong TSHD is selected as the case study. The vessel is equipped with an advanced dredging system and a highly integrated automatic control platform. Its twindraghead suction system allows a maximum dredging depth of 42 meters, with a hopper capacity of 12,888 [m.sup.3], providing efficient dredging, reclamation, and self-propulsion capabilities. To ensure efficient and stable operation under complex underwater topography and variable sediment conditions, multiple sensors are installed on key operational units such as the draghead, suction pipe, dredge pump, and bow discharge section. These sensors are tightly integrated with the onboard dredging automation and control system, enabling realtime monitoring and intelligent regulation of the dredging process. In total, the monitoring system collects 170 categories of operational data, which serve as the data foundation for this study. Several sensors and instruments used for data acquisition are shown in Fig. 1. According to operability, these parameters are classified into three categories: uncontrollable parameters, direct control parameters, and feedback control parameters.
4. Method
4.1 Data preprocessing
4.1.1 Outlier Detection
This study adopts the boxplot method for outlier detection. A boxplot is a graphical method that defines upper and lower limits based on quartiles and the interquartile range (IQR). Specifically, IQR = Q3 - Q1, the lower bound is Q1 - 1.5 * IQR, and the upper bound is Q3 + 1.5 * IQR. Data points beyond this range are regarded as outliers and removed. This method effectively eliminates abnormal fluctuations or invalid data, laying the foundation for subsequent analysis.
Meanwhile, for parameters that approximately follow a normal distribution in physical measurements, the 3[sigma] rule is also applied for auxiliary detection. Assuming a variable follows a normal distribution with mean [micro] and standard deviation [sigma], the acceptable range is defined as [micro] [+ or -] 3[sigma]. Data points outside this interval are considered to exhibit abnormal behavior.
4.1.2 Noise Reduction
After removing outliers and missing values, noise reduction is equally important. Since raw data often contain random noise, this study employs the Savitzky-Golay (S-G) filter to smooth the data, thereby reducing noise interference and enhancing smoothness. The S-G filter is based on local polynomial least-squares fitting, which can effectively remove noise while preserving the shape and width of the data sequence. Assume that the selected time-domain window length is 2n+1, and the data set is represented as {I[i]}={Ii[DELTA]z\-n[less than or equal to]i<n}. The polynomial fitting form derived from this process is expressed as in Equation (1).
[Please download the PDF to view the formula] (1)
Where, m denotes the highest order of the polynomial fitting, and bk represents the fitting coefficient corresponding to the k-th order term. The coefficients of this polynomial can be solved using the least-squares method, and the sum of squared residuals can be expressed as shown in Equation (2).
[Please download the PDF to view the formula] (2)
Where, n denotes the number of samples, Pm(i)represents the fitted value, and T(i) is the actual value.
4.2 Dynamically weighted multi-hybrid feature selection model
The core objective of feature selection is to identify and extract key control and state parameters from the raw monitoring data of TSHD operations that are significantly associated with critical performance indicators, such as instantaneous output, thereby providing a reliable foundation for constructing high-accuracy predictive models. During TSHD operations, the high-dimensional feature space generated by multi-sensor monitoring contains a large number of redundant features (e.g., low-correlation operating parameters or highly coupled process variables). Such irrelevant or redundant features not only weaken the predictive accuracy and generalization capability of the model by introducing noise but also increase model complexity and reduce interpretability. Machine learning has demonstrated great potential in handling high-dimensional data and analyzing complex problems [26]. Therefore, this study integrates three complementary algorithms--L1-regularized linear model (Lasso regression), Random Forest (RF), and Light Gradient Boosting Machine (LightGBM)--to construct a hybrid feature selection mechanism based on multiple strategies. This mechanism comprehensively identifies and extracts the core parameters that decisively affect dredging efficiency. In addition, machine learning models effectively overcome the limitations of manual expertise, making them more adaptable and generalizable [27]. The overall architecture of the TSHD feature selection model is shown in Fig. 2.
Lasso regression, as a widely used method for variable selection and classification in high-dimensional data, achieves sparse modeling by introducing L1 regularization. This mechanism enables the model to automatically shrink the regression coefficients of redundant or weakly correlated features to zero during training, thereby addressing the common problem of multicollinearity in high-dimensional data. As a result, it not only performs feature selection but also enhances both the predictive accuracy and interpretability of the model [28]. Random Forest (RF) demonstrates outstanding performance in tackling complex engineering problems. It can capture nonlinear interactions among features and, owing to its ensemble learning mechanism, exhibits strong robustness against noise and overfitting. Related studies have also confirmed the effectiveness of RF in addressing complex nonlinear problems such as ground settlement prediction [29]. LightGBM, built on the gradient boosting framework, adopts strategies such as leaf-wise growth and histogram-based optimization, which allow it to efficiently identify key features with high information gain when handling imbalanced data. This makes LightGBM highly effective in complex modeling tasks [30]. In the feature selection framework constructed in this study, the importance scores of features produced by the three models are first standardized. Then, weights are assigned to each model according to its performance, and a weighted integrated feature importance score is established, as expressed in Equation (3).
[Please download the PDF to view the formula] (3)
here, Sj(*)denotes the normalized score of the feature indexed by j under the corresponding model, where [[omega].sub.1], [[omega].sub.2], [[omega].sub.3][member of][0,1] and [[omega].sub.1] + [[omega].sub.2] + [[omega].sub.3] = 1.To ensure objectivity and reproducibility, the weights were determined via a grid search strategy that minimizes the prediction error of the downstream XGBoost model. Specifically, weight combinations were traversed with a step of 0.1 under the unit-sum constraint, and for each combination, a fused feature importance score was computed. The corresponding feature subset was then used to train an XGBoost model, evaluated using 5-fold cross-validation with root mean square error (RMSE) as the criterion. The optimal weights were found to be [[omega].sub.1] = 0.2 (Lasso), [[omega].sub.2]=0.4(Random Forest), and [[omega].sub.3] =0.4(LightGBM). This allocation reflects the superior ability of ensemble-based methods (RF and LightGBM) to capture nonlinear interactions in the TSHD operational data compared to the linear Lasso model. Finally, features are ranked in descending order according to the aggregated importance scores, and the top kmost important features are selected for subsequent prediction and optimization.
4.3 Adaptive prediction model based on Bayesian-Optimized XGBoost
4.3.1 XGBoost model
XGBoost demonstrates notable advantages in terms of computational efficiency, regularization control, and robustness. Compared with traditional GBDT, XGBoost introduces a regularization term into the loss function, structurally penalizing overly complex models and thereby enhancing generalization while maintaining predictive accuracy [31]. In this study, XGBoost is employed as the primary modeling tool to fit and predict the nonlinear mapping relationship between the operational parameters of the TSHD and unit production. Its objective function can be expressed as follows:
[Please download the PDF to view the formula] (4)
In this formulation, [OMEGA]([f.sub.k]) represents the regularization term, defined as [ohm]([f.sub.k]) = [gamma]T + [1/2][lambda][parallel]w[[parallel].sup.2]. Where, l([[^.y].sub.[iota]], [y.sub.i]) denotes the loss function for the i-th sample, with [[^.y].sub.[iota]]and [y.sub.i] representing the predicted and actual values of the i-th sample, respectively. [f.sub.k] indicates the k-th regression tree, T denotes the number of leaf nodes in the tree, and w refers to the weight vector of the leaf nodes. The parameters [gamma] and [lambda] are complexity factors in the regularization term [ohm], which serve to constrain excessive tree growth and prevent overfitting of the model.
Although the traditional XGBoost model is well known for its high training efficiency and strong nonlinear modeling capacity, its limitations in dredging productivity prediction tasks cannot be overlooked. On the one hand, XGBoost assumes equal importance across all input features [32], whereas in dredging operations, the impact of different features on productivity varies substantially. For example, suction flow rate and draghead depth often exert decisive influence on instantaneous output, while secondary factors such as wave height play only a minor role. Ignoring such feature-weight imbalance may lead to poor interpretability and reduced predictive accuracy. On the other hand, XGBoost is highly sensitive to hyperparameter configurations, such as learning rate, subsampling ratio, and tree depth, which require careful tuning. Traditional manual tuning based on empirical trial-and-error is time-consuming, labor-intensive, and often fails to guarantee stable results under complex operational conditions [33]. Therefore, to enhance the adaptability and predictive accuracy of the model under dredging conditions, it is necessary to introduce advanced parameter optimization mechanisms and tailor improvements to the XGBoost framework.
4.3.2 BO-XGBoost model
To enhance the modeling capability and generalization performance of XGBoost in dredging productivity prediction tasks, this study introduces Bayesian Optimization for adaptive hyperparameter tuning. Bayesian Optimization guides the hyperparameter search process by constructing a posterior probabilistic model combined with an acquisition function (AC), thereby reducing search cost while improving model performance [34]. The approach primarily consists of two steps: Gaussian Process (GP) regression modeling and acquisition function maximization. The optimization workflow is illustrated in Fig. 3.
Specifically, the target function is modeled through a Gaussian Process (GP), under the assumption that it follows the prior distribution:
f(x) ~ GP([mu](x), k(x, x')) (5)
Among them, x denotes the hyperparameter combination to be optimized, [mu](x) represents the mean function, and k(x, x') is the covariance kernel function. In this study, the Radial Basis Function (RBF) is adopted as the kernel function. Under the condition of known sample points, the posterior distribution at a new point [x.sub.*] can be derived through the Gaussian process as follows:
P([y.sub.*] |X, y, [x.sub.*]) ~ N([mu]([x.sub.*]), [[sigma].sup.2]([x.sub.*])) (6)
To effectively guide the next round of sampling based on the surrogate model's predictions, Bayesian optimization introduces an acquisition function (AC) to balance exploration and exploitation. In this study, the Expected Improvement (EI) function is adopted as the acquisition strategy, which is formally defined as follows:
[Please download the PDF to view the formula] (7)
In this context, f([x.sup.+])denotes the best-known objective value so far. The terms [mu](x) and [sigma](x) represent the predicted mean and standard deviation at the current point, respectively. [PHI] (Z) and [??] (Z) denote the cumulative distribution function (CDF) and the probability density function (PDF) of the standard normal distribution. The parameter [xi] is a constant used to adjust the exploration strength (set to 0.01 in this study), serving to balance exploration of new regions with exploitation of the current best solution.
4.4 Feature-Guided coevolutionary optimization method
To achieve practical automatic optimization of dredging operation parameters, this study proposes a feature-guided coevolutionary optimization method that balances efficient search with engineering feasibility. This approach integrates the strengths of Feature-Oriented Sequential Univariate Optimization (FSUO) and the Constrained Grey Wolf Optimizer (CGWO), constructing a two-stage "coarse screening-fine tuning" coevolutionary process.
4.4.1 Feature-Oriented Sequential Univariate Optimization (FSUO)
The Feature-Oriented Sequential Univariate Optimization (FSUO) method evolves from the classical idea of coordinate descent [35] and incorporates insights from the SPO/SPOT framework, which is commonly used in algorithm parameter tuning and sensitivity analysis. In addition, approaches such as batch sequential adaptive design [36] provide experimental design concepts that support its stepwise optimization procedure. In this method, the most influential dredging operation control variables, identified through feature selection, are designated as the optimization targets. By applying sequential local searches to these variables within the XGBoost-based production prediction model, the method aims to maximize the model's predictive output. Let the trained production prediction model be defined as follows:
[Please download the PDF to view the formula] (8)
Where, X denotes the feature vector containing d tunable parameters. Given the initial reference point X (0), the predicted output is y(0)= f(Z(0)). The objective of FSOU is to maximize production output while maintaining the required engineering and operational constraints. This is achieved by iteratively adjusting key parameters [x.sub.j] [member of] S, where S represents the subset of critical parameters identified through feature selection. During each iteration, only a single parameter xj is optimized, and the following expression is solved:
[Please download the PDF to view the formula] (9)
Subsequently, all parameters are updated to,
[Please download the PDF to view the formula] (10)
To prevent parameters from deviating excessively from the feasible solution space due to overly large step sizes in a single update, a gain constraint threshold [[delta].sub.max] is introduced at each iteration. Accordingly, each univariate optimization must satisfy the constraint improvement condition, as defined in Equation (11).
[Please download the PDF to view the formula] (11)
Where, the variable f(*) denotes the prediction model, S represents the subset of key tunable features, and [[L.sub.j], [U.sub.j]] defines the adjustment range of feature [x.sub.j]. The parameter [[delta].sub.max] denotes the maximum allowable improvement per single optimization step (set to 20%), which restricts the intergenerational adjustment magnitude. This constraint prevents the optimization process from being misled by local abnormal perturbations, thereby avoiding erroneous search trajectories.
4.4.2 Constrained Grey Wolf Optimizer (CGWO)
To further enhance the efficiency of the search process, this study introduces the Constrained Grey Wolf Optimizer (CGWO) to refine the parameter solution space based on the "coarse screening" results obtained from FSUO. CGWO extends the standard Grey Wolf Optimizer (GWO) by incorporating an adaptive leadership mechanism and constraint-handling strategies, thereby strengthening its ability to explore and exploit feasible regions within engineering constraints. Similar to the standard GWO, CGWO simulates the social hierarchy and cooperative hunting behavior of grey wolves. The position of each wolf in the population is updated according to the following formulation (see Fig. 4).
The overall search process of CGWO comprises three major steps: encircling the prey, attacking the prey, and updating positions, which are mathematically expressed as follows:
D = |C*[X.sub.p](t) -X(t)| (12)
X(t + 1) = [X.sub.p] (t) - A*D (13)
Where, [X.sub.p](t) denotes the current optimal solution (the prey position), X(t) represents the position of the current grey wolf individual, A and D are coefficient vectors that control the search scope and step size, defined as follows:
[Please download the PDF to view the formula] (14)
The parameter a decreases linearly from 2 to 0, while [r.sub.1], [r.sub.2]E [0,1]. In each iteration, the three best-performing grey wolves ([alpha], [beta], and [delta]) act as leaders to guide the remaining individuals in conducting global exploration and local exploitation. The mathematical formulation can be expressed as follows:
[Please download the PDF to view the formula] (15)
To meet the stringent requirements of parameter feasibility in practical engineering, this study incorporates a feasibility-preserving mechanism into the CGWO algorithm by combining projection repair and penalty functions. When a candidate solution exceeds the variable bounds, it is first projected back into the feasible domain. If projection repair alone cannot fully ensure feasibility, an additional fitness penalty term (Equation (16)) is introduced to flexibly reduce the relative importance of such solutions in the evolutionary process. This approach aligns with the "repair by projection" strategy adopted by Spettel and Beyer [37] in evolutionary strategies, demonstrating the effectiveness of ensuring solution feasibility through a combination of projection and penalty mechanisms in optimization algorithms. By integrating this constraint-handling mechanism with the prior "coarse screening" of FSUO, the method effectively enhances the global optimality and stability of dredging operation parameter optimization while maintaining computational efficiency and feasibility.
X(t + 1) = [[X.sub.[alpha]] + [X.sub.[beta]] + [X.sub.[delta]]/3] (16)
5. Case Study and Discussion
5.1 Data Preprocessing
Upon inspection, it was found that out of the 170 monitored variables, 120 were valid. To ensure the accuracy and reliability of the monitoring data, the boxplot method described in Section 4 was applied for anomaly detection.
Fig. 5 illustrates seven representative features selected from the 120-dimensional dataset after boxplot inspection. Features such as GPS 1 ground speed (F55), right density (F128), and right suction vacuum (F12) showed no outliers. In contrast, certain features (F121, F50, F71) contained multiple anomalous data points. A total of 2,055 abnormal records were identified, labeled, and removed, as indicated by the orange diamonds in the figure. Additionally, some features (F139, F94) exhibited constant values in the non-outlier range and were therefore excluded from further analysis. Finally, all retained features were denoised using the Savitzky-Golay (S-G) filter.
5.2 Key Feature Selection
SHAP values are essentially derived from the concept of Shapley values, which fairly distribute the contribution of each feature to the model's prediction outcome. They highlight the micro-level relationship between feature parameters and the target variable. Each feature has a corresponding SHAP value, representing its individual impact on the prediction (positive, negative, or the magnitude of influence). By analyzing the distribution patterns of SHAP values (i.e., their range and density), one can assess both the strength of a feature's influence on the model and the critical value intervals. Blue points correspond to negative SHAP values, indicating a decrease in the predicted instantaneous production rate, while red points correspond to positive SHAP values, suggesting an increase. As illustrated in Fig. 6, the SHAP value distribution for the right suction vacuum exhibits a wide range, implying that this feature exerts a bidirectional effect on the model's output. When the GPS ground speed is relatively low (blue region), SHAP values are predominantly negative, suggesting that lower vessel speed may adversely affect output. Conversely, the right high-pressure flushing pressure shows significant overlap with positive SHAP values, indicating that higher flushing pressure contributes positively to the predicted production rate. Additionally, the broad distribution of SHAP values for the right upper vertical angle demonstrates its strong influence on model output.
In the comparative feature importance analysis, significant differences emerged between LightGBM and Random Forest due to their distinct algorithmic principles: LightGBM places greater emphasis on dynamic variations, whereas Random Forest focuses more on static indicators. After incorporating SHAP-based importance (marginal contribution) and Lasso (regularization-based selection), six features consistently emerged as stable core features across models: right high-pressure flushing pressure, right active draghead angle, right upper vertical angle, right draghead depth, GPS ground speed, and right active draghead pressure. Following the dynamic weighted multiple-feature selection framework, this study ultimately retained 13 critical feature parameters, including right draghead depth, right active draghead angle, right upper vertical angle, GPS ground speed, right active draghead pressure, and right high-pressure flushing pressure, among others (see Fig. 7).
5.3 Analysis of predictive model performance
To improve the prediction accuracy of the target variable (instantaneous production), a Bayesian Optimization-enhanced XGBoost model (BO-XGBoost) was constructed based on the 13 selected input variables. This model was benchmarked against Random Forest (RF), Probabilistic Neural Network (PNN), and Backpropagation Neural Network (BPNN). The XGBoost framework was first initialized, and six core hyperparameters were defined for optimization: number of weak learners, learning rate, maximum tree depth, L2 regularization coefficient, subsample ratio, and feature column sampling ratio. The optimization objective function was set to the mean squared error (MSE) under 5-fold cross-validation. The dataset was randomly split into training (472,968 samples, 80%) and testing (118,243 samples, 20%) sets. Hyperparameter tuning was conducted using Bayesian Optimization over 100 iterations, resulting in an optimal configuration (Table 1), which was subsequently used to establish the mapping between key control parameters and instantaneous production. Model performance was quantitatively evaluated using four metrics: root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and the coefficient of determination ([R.sup.2]) (Table 2). To evaluate whether the performance improvement of BO-XGBoost over the baseline models is statistically significant, the Wilcoxon signed-rank test was conducted on the absolute prediction errors of each model against BO-XGBoost across the test set. As reported in Table 2, the p-values for all comparisons are below 0.01, indicating that the differences are statistically significant at the 95% confidence level. This confirms that the observed improvements are not due to random chance but reflect the genuine superiority of the proposed model.
The comparative results show that the BO-XGBoost model significantly outperformed all benchmarks, achieving the lowest RMSE, MAE, MAPE, and total computational time, along with the highest [R.sup.2] (0.938), indicating superior explanatory power. The baseline XGBoost model ranked second (RMSE = 124.6, [R.sup.2] = 0.902), confirming the effectiveness of the gradient boosting framework. Among neural network models, PNN (RMSE = 149.3, [R.sup.2] = 0.861) performed slightly better than BPNN (RMSE = 153.7, [R.sup.2] = 0.854), though both were markedly inferior to XGBoost variants. Random Forest (RF) exhibited the weakest performance (RMSE = 162.4, [R.sup.2] = 0.828), reflecting its modeling limitations in this context. As illustrated in Fig. 8, Bayesian Optimization significantly enhanced the predictive accuracy of XGBoost, demonstrating its superiority as the optimal solution for capacity prediction in TSHD operations.
5.4 Optimization of Controllable Operational Parameters
In the coarse screening stage, four key adjustable parameters were selected based on engineering expertise and parameter sensitivity analysis: the right active draghead angle, the right high-pressure flushing pressure, GPS1 ground speed, and the right upper vertical angle. Local perturbation experiments were conducted using Feature-Oriented Sequential Univariate Optimization (FSUO) to identify the optimal adjustment intervals (Fig. 9). The final optimal control ranges were comprehensively determined, as summarized in Table 3.
In the fine-tuning stage, the first 80% of historical data (472,968 samples) was used for model training, and the remaining 20% (118,243 samples) was employed for performance validation. Using the Taguchi orthogonal experimental method, an orthogonal table was designed to optimize five hyperparameters: wolf population size (num_wolves), maximum number of iterations (max_iter), initial convergence factor (a_init), position update weight (subsample), and boundary penalty coefficient (gamma) (Table 4). Each experiment was repeated ten times, and the average value was taken as the response variable, with mean production rate as the optimization objective (Table 5). Range analysis (Fig. 10) indicated the sensitivity order of the hyperparameters as: num_wolves > max_iter > a_init > subsample > gamma. The optimal combination was determined as num_wolves = 100, max_iter = 120, a_init = 2.5, subsample = 0.8, and gamma = 1.0. The orthogonal experiment required only 25 trials to obtain the optimal combination, whereas a grid search would require [5.sup.5] = 3125 trials under the same conditions. Thus, the orthogonal method reduced the number of tests by 99.20%, substantially lowering computational costs.
Subsequently, the Constrained Grey Wolf Optimizer (CGWO) was applied for parameter optimization. The dredging production rate increased from 4,670.37 [m.sup.3]/h to 4,871.93 [m.sup.3]/h, yielding an absolute gain of 201.56 [m.sup.3]/h, equivalent to an improvement of 4.32% (Table 6). Among the adjusted parameters, the right active draghead angle showed the largest variation (-1.4309[degrees]); a moderate reduction optimized the cutting attack angle, thereby improving soil penetration efficiency. Conversely, the GPS ground speed exhibited the smallest change (-0.1307 m/s), confirming that maintaining stable speed plays a critical role in ensuring consistent production performance.
To validate the effectiveness of the proposed optimization framework, a comparative study was conducted against three benchmark algorithms: standard Grey Wolf Optimizer (GWO), Particle Swarm Optimization (PSO), and Genetic Algorithm (GA). In addition, an ablation study was performed by removing the FSUO component to evaluate its contribution, denoted as "CGWO (w/o FSUO)". All algorithms were configured with identical constraints, the same XGBoost prediction model as the objective function, and the same initial baseline point. Each algorithm was independently run 10 times to account for stochastic variability.
As summarized in Table 7, the proposed FSUO+CGWO framework achieves the highest average production improvement (4.32%) and the fastest convergence (85 iterations), along with the lowest standard deviation (0.08). Compared to the standalone CGWO, the full framework improves production by an additional 0.45% and reduces convergence iterations by 27, confirming the necessity of the FSUO coarse-screening stage. Moreover, the proposed method outperforms standard GWO, PSO, and GA by a notable margin, demonstrating its superiority over conventional single-stage heuristic optimizers.
6. Conclusions
This study focuses on the problem of production capacity regulation during the loading operations of self-propelled Trailing Suction Hopper Dredgers (TSHD) and proposes an intelligent decision-making framework that integrates multi-model learning and cooperative optimization. The method fully accounts for the high dimensionality, nonlinearity, and engineering constraints of dredging data, with in-depth investigations conducted at three levels: feature selection, model construction, and parameter optimization. The main contributions and findings can be summarized as follows:
(1) Feature Selection: To address the challenge of redundant and overlapping information in dredging operations, a dynamically weighted hybrid feature selection model was developed. This model integrates Lasso regression, Random Forest (RF), and LightGBM to generate feature importance scores. By applying standardized fusion and weight allocation, the method effectively extracts variables that have critical impacts on productivity, thereby establishing a robust foundation for subsequent modeling and optimization. It balances interpretability with the ability to capture both linear constraints and nonlinear relationships, ensuring strong generalizability and robustness.
(2) Predictive Modeling: A high-precision predictive model combining Bayesian Optimization (BO) and XGBoost was constructed. By employing BO to efficiently search the hyperparameter space, the model achieves greater stability and generalization capacity in dredging productivity prediction tasks. The results confirm its superior performance under different sampling ratios and feature dimensions.
(3) Parameter Optimization: To further enhance the effectiveness of controllable parameter optimization, a two-stage cooperative optimization strategy was designed, coupling Feature-Oriented Sequential Univariate Optimization (FSUO) with a Constrained Grey Wolf Optimizer (CGWO). This framework ensures both engineering feasibility and operational safety while improving productivity. Experimental results show that the optimized strategy increases instantaneous output by 4.32%, with all parameter variations remaining within reasonable limits. The total optimization time was 3.6226 minutes, satisfying real-time operational requirements.
(4) Orthogonal Experiment Design: To determine the optimal configuration of CGWO hyperparameters, orthogonal experimental design was implemented. Five factors--number of wolves (num_wolves), maximum iterations (max_iter), initial convergence factor (a_init), subsample ratio, and local search coefficient (gamma)--were systematically evaluated. The results confirm that appropriately controlling the search granularity improves both optimization efficiency and performance, highlighting a practical strategy for model enhancement.
Despite the promising results, this study has several limitations that should be acknowledged. First, the proposed framework relies heavily on the quality and completeness of sensor data. Any sensor malfunction, calibration drift, or data loss could compromise the accuracy of both feature selection and prediction. Second, the model was trained and validated using data collected from a single TSHD operating in specific geological and environmental conditions. Its generalization capability to other vessels or significantly different sediment types (e.g., cohesive clay versus granular sand) has not been fully explored and may require domain adaptation or retraining. Third, the current optimization focuses solely on instantaneous production as the objective, without explicitly considering trade-offs with energy consumption, equipment wear, or environmental impact--factors that are increasingly important in sustainable dredging operations.
Future research will focus on extending the framework in three directions: (1) integrating multi-objective optimization to balance productivity with energy efficiency and environmental constraints; (2) developing transfer learning or meta-learning strategies to enhance model adaptability across different vessels and working conditions; and (3) embedding the proposed method into a real-time digital twin system to enable closed-loop adaptive control for TSHD operations.
Acknowledgments
This study is supported by Leading Industry Project of Shanghai Municipal Commission of Economy and Information Technology (No.2024-GZL-RGZN-01025).
References
[1] Cortez-Huerta, M., Echeverria, R. S., Garcia, G. F., Duran, R. E. A., Ramirez-Macias, J. I., & Kahl, J. D., "High-resolution atmospheric emissions estimate from dredging activities during port expansion in Veracruz, Mexico," Ocean Engineering, vol.310, part.1, Oct. 2024. Article(CrossRefLink)
[2] Dredging Market, Dredging Market Study by Application in Trade Activity, Trade Maintenance, Energy Infrastructure, Urban Development, Coastal Protection, and Leisure from 2024 to 2034. Fact.MR, 2024. [Online]. Available: https://www.factmr.com/report/922/dredging-market
[3] PIANC, Best Practices for Sustainable Dredging and Infrastructure Development, Permanent International Association of Navigation Congresses, 2016.
[4] Zhu, L., & Wu, W., Advances in dredging engineering, Springer Nature, 2022.
[5] IADC (International Association of Dredging Companies), Dredging for sustainable infrastructure, CEDA/IADC, 2020.
[6] Li, M., Lu, Q., Bai, S., Zhang, M., Tian, H., & Qin, L., "Digital twin-driven virtual sensor approach for safe construction operations of trailing suction hopper dredger," Automation in Construction, vol.132, Dec. 2021. Article(CrossRefLink)
[7] Wei, C., Wei, Y., & Ji, Z., "Model predictive control for slurry pipeline transportation of a cutter suction dredger," Ocean Engineering, vol.227, May 2021. Article(CrossRefLink)
[8] Bai, S., Li, M. C., Song, L. G., Ren, Q. B., Qin, L., & Fu, J. K., "Productivity analysis of trailing suction hopper dredgers using stacking strategy," Automation in Construction, vol. 122, Feb. 2021. Article(CrossRefLink)
[9] Wang, B., Fan, S., Jiang, P., Xing, T., Fang, Z., & Wen, Q., "Research on predicting the productivity of cutter suction dredgers based on data mining with model stacked generalization," Ocean Engineering, vol.217, Dec. 2020. Article(CrossRefLink)
[10] Tang, H., Chai, L., Shi, Q., & Zhang, J., "Modeling of Trailing Suction Hopper Dredger: A Two-Step Method Integrating Deep Neural Network and Human Knowledge," IEEE Sensors Journal, vol.22, no.2, pp.1552-1559, Jan. 2022. Article(CrossRefLink)
[11] Moshood, T. D., Rotimi, J. O. B., Shahzad, W., & Bamgbade, J. A., "Infrastructure digital twin technology: A new paradigm for future construction industry," Technology in Society, vol.77, Jun. 2024. Article(CrossRefLink)
[12] Wang, B., Fan, S., Chen, Y., Zheng, L., Zhu, H., Fang, Z., & Zhang, M., "The replacement of dysfunctional sensors based on the digital twin method during the cutter suction dredger construction process," Measurement, vol.189, Feb. 2022. Article(CrossRefLink)
[13] Wang, Z., Gonzalez, V A., Mei, Q., & Lee, G., "Sensor adoption in the construction industry: Barriers, opportunities, and strategies," Automation in Construction, vol.170, Feb. 2025. Article(CrossRefLink)
[14] Chen, Y., Ren, Q. B., Li, M. C., Tian, H. J., Qin, L., & Wu, D. C., "Data-driven multi-objective prediction and optimization of construction productivity and energy consumption in cutter suction dredging," Automation in Construction, vol.173, May 2025. Article(CrossRefLink)
[15] Zhang, X., Wang, X., Yao, H., Luo, W., Qiang, Z., & Luo, D., "Multi-similarity and gradient fusion digital twins for fault detection and diagnosis of rolling bearings," Computers in Industry, vol. 168, Jun. 2025. Article(CrossRefLink)
[16] Huang, K., Li, W., & Gao, F., "Barabasi-Albert model-enhanced genetic algorithm for optimizing LGBM in ship power grid fault diagnosis," Measurement, vol.249, May 2025. Article(CrossRefLink)
[17] Liu, H., Liu, R., Bai, S., Chen, Y., & Liu, L., "Efficient Productivity-Aware Control Parameter Optimization in Cutter Suction Dredger Construction Using Machine Learning with Parallel Global Search," Water, vol.16, no.21, Nov. 2024. Article(CrossRefLink)
[18] Chen, Y., Ren, Q., Li, M., Tian, H., Qin, L., & Wu, D., "Cluster-wise deep learning framework for scenario-adaptive energy consumption and productivity prediction of cutter suction dredger," Energy, vol.340, Dec. 2025. Article(CrossRefLink)
[19] Liu, R., Yu, G., Shi, K., Chen, Y., & Ren, Q., "Physics-Assisted Deep Learning Model for Improved Construction Performance Monitoring of Cutter Suction Dredger," Water, vol. 17, no.24, Dec. 2025. Article(CrossRefLink)
[20] Li, M. C., Kong, R., Han, S., Tian, G. P., & Qin, L., "Novel Method of Construction-Efficiency Evaluation of Cutter Suction Dredger Based on Real-Time Monitoring Data," Journal of Waterway, Port, Coastal, and Ocean Engineering, vol.144, no.6, 2018. Article(CrossRefLink)
[21] Chen, Y., Ren, Q. B., Li, M. C., Tian, H. J., Qin, L., & Wu, D. C., "Prediction of submarine soil dredging difficulty scale in cutter suction dredger construction with clustering-based deep learning," Engineering Applications of Artificial Intelligence, vol.147, May 2025. Article(CrossRefLink)
[22] Zhou, B., Yu, M., Zhou, X., Zhou, Z., & Liu, N., "Realtime estimation of average sand grain diameter and sand bed height and mass for a trailing suction hopper dredger," Scientific Reports, vol.14, no.1, 2024. Article(CrossRefLink)
[23] Su, Z., Zhang, X., Yao, Y., Zhu, D., & Fu, J., "Mathematic modeling and CFD simulation of the sedimentation process in a TSHD," Marine Georesources & Geotechnology, vol.41, no.5, pp.461-475, 2023. Article(CrossRefLink)
[24] Qi, J., Hao, Y., & Chai, W., "Failure analysis on wave compensator and hull structure of trailing suction hopper dredger (TSHD) under Medium-Long period wave," Engineering Failure Analysis, vol.152, Oct. 2023. Article(CrossRefLink)
[25] Cheng, T., Lu, Q. R., Kang, H. R., Fan, Z. Y., & Bai, S., "Productivity Prediction and Analysis Method of Large Trailing Suction Hopper Dredger Based on Construction Big Data," Buildings, vol.12, no.10, Oct. 2022. Article(CrossRefLink)
[26] Zhao, C., Yue, W., Xia, Q., Yang, H., Chen, A., & Liu, X., "Prediction of CO2 adsorption performance of biochar based on machine learning," Biomass and Bioenergy, vol.201, Oct. 2025. Article(CrossRefLink)
[27] Shang, G., Xu, L. Y., Tian, J. Z., Cai, D. W., Xu, Z., & Zhou, Z., "A real-time green construction optimization strategy for engineering vessels considering fuel consumption and productivity: A case study on a cutter suction dredger," Energy, vol.274, Jul. 2023. Article(CrossRefLink)
[28] Wang, J. K., Fei, L. H., & Sun, L., "Multi-level network Lasso for multi-task personalized learning," Pattern Recognition, vol.161, May 2025. Article(CrossRefLink)
[29] Zhang, R. H., Li, Y. Q., Goh, A. T. C., Zhang, W. G., & Chen, Z. X., "Analysis of ground surface settlement in anisotropic clays using extreme gradient boosting and random forest regression models," Journal of Rock Mechanics and Geotechnical Engineering, vol.13, no.6, pp.1478-1484, Dec. 2021. Article(CrossRefLink)
[30] Zhou, C. Y., Wang, Z., Wang, X. L., Guo, R., Zhang, Z., Xiang, X. W., & Wu, Y. Q., "Deciphering the nonlinear and synergistic role of building energy variables in shaping carbon emissions: A LightGBM- SHAP framework in office buildings," Building and Environment, vol.266, Dec. 2024. Article(CrossRefLink)
[31] Mao, F., Chen, M., Zhong, K., Zeng, J., & Liang, Z., "An XGBoost-assisted evolutionary algorithm for expensive multiobjective optimization problems," Information Sciences, vol.666, May 2024. Article(CrossRefLink)
[32] Shao, Z. F., Ahmad, M. N., & Javed, A., "Comparison of Random Forest and XGBoost Classifiers Using Integrated Optical and SAR Features for Mapping Urban Impervious Surface," Remote Sensing, vol.16, no.4, Feb. 2024. Article(CrossRefLink)
[33] Bo, Y., Guo, X., Liu, Q., Pan, Y., Zhang, L., & Lu, Y., "Prediction of tunnel deformation using PSO variant integrated with XGBoost and its TBM jamming application," Tunnelling and Underground Space Technology, vol.150, Aug. 2024. Article(CrossRefLink)
[34] Li, Q., & Wang, X., "Bayesian optimization of stacking ensemble learning model for HPC compressive strength prediction," Expert Systems with Applications, vol.288, Sep. 2025. Article(CrossRefLink)
[35] Wright, S. J., "Coordinate descent algorithms," Mathematical Programming, vol.151, no.1, pp.3-34, Jun. 2015. Article(CrossRefLink)
[36] Xiao, Y., Ning, J., Xiong, Z., & Qin, H., "Batch sequential adaptive designs for global optimization," Journal of the Korean Statistical Society, vol.51, no.3, pp.780-802, Sep. 2022. Article(CrossRefLink)
[37] Spettel, P., & Beyer, H.-G., "Analysis of the ([mu]/[micro]I, [lambda])-[sigma]-Self-Adaptation Evolution Strategy With Repair by Projection Applied to a Conically Constrained Problem," IEEE Transactions on Evolutionary Computation, vol.24, no.3, pp.593-602, Jun. 2020. Article(CrossRefLink)
Liyun Xu was born in Zhejiang, China in 1973. He obtained his bachelor's and master's degrees from Harbin University of Science and Technology in 1996 and 1998, respectively. He obtained his PhD from Shanghai Jiao Tong University in 2001. He is now a professor at the School of Mechanical Engineering, Tongji University. His main research areas are intelligent manufacturing and automated equipment design. He has published over 100 academic papers focusing on these fields.
Zhixian Jin was born in Zhejiang, China in 1991. He graduated from Zhejiang University of Technology with a bachelor's degree in 2014 and obtained a master's degree in Industrial Engineering from the School of Mechanical Engineering, Zhejiang University in 2017. In September 2023, he enrolled in the School of Mechanical Engineering at Tongji University to pursue a doctoral degree. His research direction is engineering artificial intelligence.
Jing Liang obtained his bachelor's degree from Jiangnan University in 2007. He earned his master's degree from the University of Technology and Art Dortmund in 2011. He obtained his doctoral degree from Tongji University in 2025. He joined Tongji University in 2012 and is currently an associate professor. One of his research directions is artificial intelligence.
Fangrong Chen is a Master's student enrolled in 2024 at the College of Logistics Engineering, Shanghai Maritime University. She received her Bachelor of Management degree from Jiangxi University of Finance and Economics in 2021. Her research interests focus on intelligent decision-making in scheduling and planning, dynamic scheduling, deep reinforcement learning, agent-assisted evolutionary algorithms, and cloud manufacturing task scheduling.
Jianguo Duan was born in Shandong, China in 1980. He obtained his Doctor's degree in Mechanical Manufacturing and Automation from Tongji University, and currently serves as an Associate Professor and Master's Supervisor at Shanghai Maritime University.Since 2017, he has been an Associate Professor at the Pilot Free Trade Zone Supply Chain Research Institute, Shanghai Maritime University. His research direction is artificial intelligence and its applications.
Liyun Xu (1,*), Zhixian Jin (1), Jing Liang (2,*), Fangrong Chen (3) and Jianguo Duan (4)
(1) School of Mechanical Engineering, Tongji University, Shanghai, 201804 China [e-mail: Lyxu@tongji.edu.cn]
(2) School of Design and Innovation, Tongji University, Shanghai, 200092 China [e-mail: 12046@tongji.edu.cn]
(3) Logistics Engineering College, Shanghai Maritime University, Shanghai, 201306 China [e-mail: 18721435074@163.com]
(4) China Institute of FTZ Supply Chain, Shanghai Maritime University, Shanghai, 201306 China [e-mail: jgduan@shmtu.edu.cn]
(*) Corresponding author: Liyun Xu, Jing Liang
Received December 10, 2025; revised April 4, 2026; accepted April 27, 2026; published July 31, 2026
http://doi.org/10.3837/tiis.2026.07.006
Table 1. Optimal hyperparameters of Bayesian Optimization
Parameter Name Parameter Interpretation Value Range
learning_rate Learning rate 0.001~0.3
max_depth Maximum tree depth 3~12
n_estimators Number of trees 100~1500
gamma Minimum loss reduction for 0~5
splitting nodes
min_child_weight Minimum sum of instance weights 1~10
for leaf nodes
subsample Subsample ratio of training 0.5~1
instances
colsample_bytree Subsample ratio of features for tree 0.5~1
construction
Parameter Name Final Value
learning_rate 0.1089
max_depth 11
n_estimators 587
gamma 1.9859
min_child_weight 6
subsample 0.7856
colsample_bytree 0.9185
Table 2. Optimal hyperparameters of Bayesian Optimization
Model RMSE MAPE MAE [R.sup.2] p-value (vs.
BO-XGBoost)
XGBoost 738.9384 43.55% 501.3198 0.8784 < 0.01
RF 1061.9580 75.23% 728.6645 0.7489 < 0.01
PNN 862.3521 48.36% 583.8472 0.8348 < 0.01
BPNN 931.0777 55.74% 638.7504 0.8070 < 0.01
BO 590.5462 32.5772% 398.4488 0.9223 --
-XGBoost
Model Total Time
XGBoost 193.5827 s
RF 344.7823 s
PNN 15797.6293s
BPNN 7225.553 s
BO 187.3609 s
-XGBoost
Table 3. Parameter control range Parameter Range (min ~ max) Adjustment Rationale Right Active -2.5 ~ 5.0 Constrained to stable mid-range, Draghead Angle avoiding extreme angles GPS1 Speed 2.3 ~ 2.6 Reduced upper limit for moderate Over Ground forward movement Right High 4.8 ~ 5.3 Focused around operational plateau -Pressure Water to prevent under/overshoot Pressure Right Upper 17.5 ~ 23 Narrowed to optimal mud-entry Vertical Angle range
Table 4. Table of hyperparameter levels for CGWO orthogonal experiment Level Num-wolves Max-iter A-init Subsample Gamma 1 10 10 0.5 0.4 0.2 2 20 30 1 0.6 0.5 3 50 50 1.5 0.7 1 4 80 80 2 0.8 1.5 5 100 120 2.5 1 2.5
Table 5. Results of the CGWO model orthogonal experiment scheme Order Num-wolves Max-iter A-init Subsample Gamma Output 1 10 10 0.5 0.4 0.2 4520.1333 2 10 30 1 0.6 0.5 4642.192 3 10 50 1.5 0.7 1 4711.195 4 10 80 2 0.8 1.5 4655.409 5 10 120 2.5 1 2.5 4699.489 6 20 10 1 0.7 1.5 4549.69 7 20 30 1.5 0.8 2.5 4684.738 8 20 50 2 1 0.2 4576.791 9 20 80 2.5 0.4 0.5 4725.857 10 20 120 0.5 0.6 1 4715.9624 11 50 10 1.5 1 0.5 4627.3267 12 50 30 2 0.4 1 4707.8164 13 50 50 2.5 0.6 1.5 4729.3745 14 50 80 0.5 0.7 2.5 4673.4946 15 50 120 1 0.8 0.2 4815.4897 16 80 10 2 0.6 0.5 4636.6797 17 80 30 2.5 0.7 1 4789.9087 18 80 50 0.5 0.8 1.5 4784.9883 19 80 80 1 1 1 4867.877 20 80 120 1.5 0.4 1.5 4808.8804 21 100 10 2.5 0.8 1 4754.9043 22 100 30 0.5 1 1.5 4837.824 23 100 50 1 0.4 2.5 4796.0444 24 100 80 1.5 0.6 0.2 4863.5835 25 100 120 2 0.7 0.5 4872.6343
Table 6. Parameter optimization results
Parameter Baseline Optimized Value Change
Value
Right Active 0.9516 -0.4793 -1.4309
Draghead
Angle ([degrees])
GPS1 Speed 2.4700 2.3393 -0.1307
Over Ground
(m/s)
Right High 5.5549 5.0729 -0.4820
-Pressure Water
Pressure (MPa)
TRight Upper 21.6636 23.5810 1.9174
Vertical
Angle ([degrees])
Production 4670.3727 4871.9263 201.5534
([m.sup.3]/h)
Parameter Total Optimization Time (min)
Right Active
Draghead
Angle ([degrees])
GPS1 Speed
Over Ground
(m/s)
Right High 3.6226
-Pressure Water
Pressure (MPa)
TRight Upper
Vertical
Angle ([degrees])
Production
([m.sup.3]/h)
Table 7. Performance comparison of optimization algorithms
Algorithm Avg. Production Best Production Avg. Std. Dev.
Improvement (%) Improvement Convergence (10 runs)
(%) Iterations
FSUO + 4.32 4.35 85 0.08
CGWO
(Proposed)
CGWO 3.87 3.95 112 0.15
GWO 3.62 3.71 108 0.18
PSO 3.56 3.68 96 0.21
GA 3.41 3.52 134 0.24
 Printer friendly
 Cite/link
 Email
 Feedback
| |
| Author: | Xu, Liyun; Jin, Zhixian; Liang, Jing; Chen, Fangrong; Duan, Jianguo |
|---|---|
| Publication: | KSII Transactions on Internet and Information Systems |
| Article Type: | Report |
| Geographic Code: | 9CHIN |
| Date: | Jul 1, 2026 |
| Words: | 8281 |
| Previous Article: | Engagement Level Assessment via Computer Camera Using Machine Learning Based on Facial Movement Analysis. |
| Next Article: | CST-Net: A CNN-SE-Transformer Hybrid Model for Robust LoRa Indoor Positioning. |
| Topics: | |