A Smart Control Method for the Data Acquisition Head of a Subsea Mining Vehicle Considering Multi-Factor Coupling

Abstract

This invention discloses an intelligent control method for the seabed mining vehicle's acquisition head, considering multi-factor coupling. By establishing a multi-source parameter fusion-based intelligent hierarchical system for seabed mining, it provides precise working conditions and attitude data for acquisition head regulation, dynamically matching optimal acquisition head angle, height, and jet velocity parameters, thus achieving an intelligent transformation of acquisition regulation strategies. This intelligent control method includes a data perception layer, a model prediction layer, an execution control layer, and a feedback optimization layer. By constructing a three-dimensional standardized dataset containing basic, interactive, and statistical features, it fully leverages the synergistic prediction advantages of the XGBoost dual model and the XGBoost-LSSVR stacked model, significantly improving parameter prediction accuracy in multi-factor coupling scenarios. Combined with multi-dimensional real-time monitoring data and closed-loop feedback control, it constructs a "monitoring-prediction-regulation" intelligent control system, which can proactively adjust acquisition head parameters, effectively suppressing over-mining and under-mining, reducing acquisition head wear, and improving mineral acquisition efficiency and operational safety.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for deep-sea mining equipment, specifically to an intelligent control method for the seabed mining vehicle's data acquisition head that considers the coupling of multiple factors. Background Technology

[0002] Deep-sea mining, as a core means of developing marine mineral resources, directly determines the economic viability and feasibility of mineral development through its operational efficiency and equipment adaptability. Deep-sea mining vehicles, with their stable mobility, have become the mainstream equipment for harvesting polymetallic nodules, cobalt-rich crusts, and other mineral resources from the seabed. The harvesting head, as the core actuator that directly contacts the minerals, has its relative angle and height parameters with the seabed being key factors affecting harvesting efficiency. Optimal angle and height configurations maximize mineral recovery and reduce harvesting resistance, while inappropriate configurations can easily lead to problems such as missed or over-harvested samples and harvesting head wear.

[0003] However, the deep-sea seabed is rugged and varied, generally characterized by undulating slopes, uneven mineral distribution, and differences in sediment thickness. When mining vehicles travel on such terrain, the vehicle body is prone to pitching and rolling changes, causing the relative position and angle of the sampling head to deviate from the preset optimal value, directly affecting the sampling results. Currently, there is a clear technological gap in the control technology of deep-sea mining vehicle sampling heads: there is a lack of a technical solution that can intelligently optimize the angle and height of the sampling head based on dynamic changes in vehicle posture using artificial intelligence algorithms. Specifically, this deficiency manifests in the following three aspects: The control methods are passive and lack a dynamic linkage mechanism with the vehicle's attitude. Existing technologies mostly use fixed threshold triggering or manual remote intervention modes, without establishing a real-time linkage relationship between the acquisition head parameters and the vehicle's attitude. When the vehicle's attitude changes due to terrain undulations, the angle and height adjustment of the acquisition head lags significantly, failing to compensate for deviations in real time. This leads to near-distance collision wear or long-distance missed sampling, and improper angles increase sampling resistance and reduce mineral breakage rate.

[0004] Without the introduction of intelligent optimization algorithms, it is difficult to adapt to changing operating scenarios. Existing technologies rely on empirical parameters or simple proportional adjustment logic, without incorporating intelligent algorithms such as machine learning to achieve dynamic optimization of collected parameters. Due to significant differences in terrain features (slope, undulation frequency) and mineral distribution characteristics in different mining areas, fixed control logic cannot adapt to complex and changing operating environments, making it difficult to achieve optimal matching of "attitude-parameter-efficiency".

[0005] The control precision is insufficient under multi-factor coupling conditions, and a collaborative optimization model is lacking. The optimal angle and height of the acquisition head are not isolated parameters, but are strongly coupled with multiple factors such as vehicle attitude (pitch angle, roll angle), driving speed, mineral density, and sediment characteristics. Existing technologies have not established a multivariate collaborative optimization model, making it difficult to quantify the influence weight of each factor on acquisition efficiency. This leads to frequent parameter deviations during the control process, which not only reduces acquisition efficiency but also exacerbates the wear and tear on the acquisition head and transmission system due to long-term uneven stress, increasing maintenance costs and operational risks.

[0006] The aforementioned shortcomings prevent the deep-sea mining vehicle's sampling head from fully realizing its operational capabilities, hindering mineral recovery rates and increasing equipment energy consumption and the probability of malfunctions. Therefore, there is an urgent need to develop a technical solution that can intelligently optimize the sampling head's angle and height based on dynamic changes in the vehicle's posture using artificial intelligence algorithms, in order to overcome the existing bottlenecks. Summary of the Invention

[0007] In response to the aforementioned engineering challenges in deep-sea mining operations, such as complex and variable seabed topography, passive and rigid parameter control of the acquisition head, and insufficient control precision under the coupling of multiple factors, resulting in low mineral recovery rates, high acquisition head wear, and poor operational efficiency, this invention provides a full-process control method that combines seabed topography and mineral distribution characteristics for working condition identification, intelligent acquisition parameter selection based on machine learning algorithms, real-time monitoring of multi-source data, and dynamic closed-loop control. This method solves the problems of parameter reliance on experience, single-factor monitoring, and lack of multi-factor collaborative optimization in traditional mining vehicle acquisition head control. It can effectively predict and proactively adjust the optimal parameters of the acquisition head, thereby improving the safety, efficiency, and intelligence level of seabed mining vehicle acquisition operations.

[0008] To achieve the above objectives, this invention provides an intelligent control method for the seabed mining vehicle's data acquisition head that considers the coupling of multiple factors, characterized by comprising the following steps: S1. Employs multiple sensors such as lidar and pressure sensors (flow and attitude sensors can be added) to collect vehicle attitude, seabed environment, and acquisition head status parameters, which are then transmitted to the data processing module for aggregation. Key indicators are selected as input features, and acquisition head control parameters are selected as output targets. A standardized dataset is constructed after normalization. For extreme seabed environments, an adaptive filtering and feature enhancement module is added to improve the accuracy of attitude data detection.

[0009] S2. The constructed standardized dataset is cleaned, and the cleaned data is used for XGBoost model training. By setting a reasonable initial range of hyperparameters, the foundation for model training is laid. The XGBoost model adopts a dual-model framework, including an angle prediction model and a height prediction model.

[0010] S3. Optimize the XGBoost model hyperparameters using K-valued cross-validation (m). Divide the cleaned dataset into K subsets, alternating between them as training and validation sets for iterative training to minimize model error and avoid overfitting. After training, input test data to predict the optimal angle of the acquisition head. and optimal angle height During the training process, a multi-objective balance optimization mechanism is embedded, with collection efficiency as the core and mineral recovery rate and collection head loss rate as auxiliary objectives.

[0011] S4. Based on the optimal parameters of the acquisition head, commands are sent to the actuator to synchronously adjust the vehicle attitude, acquisition head parameters, and jet velocity; a redundant control command mechanism is designed to avoid control inaccuracies in extreme seabed environments; sensors collect and feedback data in real time, employing... The principle is to eliminate outliers and retain valid data; at the same time, the amount of minerals collected per unit time is collected using a flow sensor to calculate the actual collection efficiency. .

[0012] S5. Input the effective feedback data into the XGBoost-LSSVR stacked model. XGBoost extracts features and makes preliminary predictions. The results are then input into LSSVR for further fitting. Combined with the environment and attitude trends, the optimal parameters and attitude changes within a preset time period are predicted. The stacked model integrates the advantages of extreme gradient boosting algorithm and least squares support vector regression algorithm. The model has terrain feature recognition function and can infer the seabed terrain type through real-time attitude data.

[0013] S6. Based on the comparison between the prediction results and the test set, verify whether the accuracy of the XGBoost-LSSVR stacked machine learning model meets the engineering requirements. If the accuracy meets the standard, dynamically fine-tune the acquisition head parameters; when the acquisition efficiency is monitored... If the efficiency falls below 90% of the optimal acquisition threshold, the online learning mechanism is triggered; if the threshold is not met, the process returns to S3 to re-optimize the model hyperparameters until the requirements are met.

[0014] S7 combines the initial optimal parameters output from S3 with the fine-tuning results from S6. Real-time adjustment parameters are obtained by inputting real-time data collected by sensors into the stacked model. The relevant parameters are continuously adjusted through closed-loop feedback logic (response delay controlled in milliseconds) to achieve intelligent control of the acquisition head throughout the process. A feedback optimization layer is added to the closed-loop control, and the incremental XGBoost algorithm is used to periodically iterate and update the parameters of the two models to adapt to changes in seabed topography. A digital twin model of the seabed mining scenario is constructed, and extreme terrain control strategies are pre-trained to improve the system's adaptability.

[0015] Achieving dynamic linkage between the data acquisition head and the vehicle body's posture requires real-time adjustment of the acquisition head's angle and height by a physical actuator. Without a hydraulic drive or servo motor drive mechanism, the optimal parameters predicted by the model cannot be translated into the actual mechanical movements of the acquisition head. Therefore, in the technical solution of this invention, the acquisition head is connected to the vehicle body via a hydraulic drive mechanism or a servo motor drive mechanism. This drive mechanism receives commands from the controller and drives the acquisition head to perform angle and height changes, thus forming a complete "perception-decision-execution" closed loop. (Explanation of the necessity of adding necessary technical features).

[0016] Preferably, the method for constructing the standardized dataset includes: selecting basic features such as vehicle posture, seabed environment, and head state. Interactive features are obtained through parameter coupling calculation. Statistical characteristics (within the last 10 seconds) were calculated based on time-series data. The variance of the fluctuations is integrated with the three types of features and normalized to construct a standardized "input-output" dataset.

[0017] It should be noted that the mathematical expressions for the above interaction characteristics have clear physical meanings and have undergone dimensional normalization: The product of the vehicle's pitch angle and the initial angle of the data acquisition head represents the coupling effect between the vehicle's longitudinal tilt and the data acquisition head's own angle. When the vehicle's pitch angle increases, the data acquisition head needs to adjust its angle in the opposite direction to maintain parallelism with the seabed. This product term can quantify the cooperative relationship between the two and be used to predict the optimal angle compensation amount under sloping terrain.

[0018] The product of the vehicle roll angle and the initial height of the data acquisition head reflects the impact of the vehicle's lateral tilt on the height of the data acquisition head off the ground. The larger the roll angle, the more significant the difference in height between the two sides of the data acquisition head. This product term is used to correct the height control strategy to avoid over-sampling or under-sampling on one side.

[0019] The product of the seabed distance and the sum of the head's angle and height comprehensively characterizes the relative spatial pose of the head and the seabed. This represents the overall pose state of the acquisition head. Multiplying it by the distance H can amplify or reduce the weight of the overall pose state's influence on acquisition efficiency.

[0020] All of the above parameters have been approved before constructing the interaction features. Normalization method mapping to Intervals eliminate differences in units (e.g., angles are in degrees, heights are in centimeters, and distances are in meters). Therefore, addition in each product term... Multiplication All calculations are performed between dimensionless values, which is mathematically valid and physically interpretable, and there is no issue of inconsistent dimensions.

[0021] Preferably, step S3 includes: constructing input features comprising three types of features: basic features, interaction features, and statistical features; and optimizing the height of the acquisition head. Optimal control angle The output dataset is the target dataset; outliers in the dataset are identified based on... The preprocessed dataset is divided into training and validation sets. K-fold cross-validation is used to train and evaluate the XGBoost dual model, outputting the optimal model for predicting the acquisition head parameters. Preferably, in step S4, real-time monitoring of the acquisition operation status includes: collecting the mineral acquisition volume per unit time using a flow sensor and calculating the actual acquisition efficiency. Simultaneously, it monitors the wear status of the acquisition head and the stability of the jet velocity, forming a multi-dimensional monitoring data system to provide data support for feedback optimization.

[0022] The XGBoost dual model uses an additive strategy to update the predicted values, and its objective function is composed of both the prediction error and the model complexity.

[0023] In the formula, The loss function is (mean squared error is used in this embodiment). This is the complexity penalty term for the k-th tree. By performing a second-order Taylor expansion on the loss function, the objective function for the t-th iteration is transformed into:

[0024] Among them, the first-order gradient Second gradient Regularization term , The number of leaf nodes. The penalty coefficient for leaf nodes. The regularization coefficient is . The weights are those of the leaf nodes.

[0025] The training process calculates the split gain and selects the optimal split feature and threshold, including: (1) Calculate the first and second gradient statistics of the nodes. (2) Try each feature and its candidate split threshold, and sort the samples in the node according to the feature value from smallest to largest. Traverse the median value of two adjacent samples as candidate split points and calculate the statistics of the left and right child nodes.

[0026] (3) Calculate the split gain, i.e. the reduction in the objective function after splitting.

[0027] In the formula: The L2 regularization coefficient is set to 1.2. γ The penalty coefficient for the number of leaf nodes is set to 0.2. The objective function return when the left child node is used as a leaf node alone. For the right child node's profit, For the benefit of not splitting (4) Calculate the split gain for all features and all candidate thresholds, and select the feature and threshold with the largest split gain as the optimal split for the current node.

[0028] (5) When the gain of all candidate splits is ≤0, it means that splitting will not reduce the objective function or may even worsen it. Therefore, splitting is stopped and the current node is set as a leaf node. In addition, even if the split gain is >0, if the depth of the tree after splitting reaches the preset value of 6, or the number of samples in the leaf node is too small, it will also stop early.

[0029] Through the above mechanism, the XGBoost dual model balances prediction accuracy and generalization ability by introducing a regularization term, while accelerating convergence using second-order gradients. Simultaneously, a multi-objective balancing optimization mechanism is embedded during training, with acquisition efficiency as the core and mineral recovery rate and acquisition head loss rate as auxiliary objectives. The weighting coefficients are dynamically adjusted to adapt to the control requirements under different terrains. Finally, the optimal angle of the acquisition head is output. and optimal height To achieve high-precision prediction of acquisition head parameters in scenarios with multiple coupled factors. Preferably, the method for constructing the XGBoost-LSSVR stacked machine learning model includes: pre-training and predicting the XGBoost algorithm and the LSSVR algorithm respectively to obtain a first-layer prediction data matrix; using the first-layer data matrix as a new input feature for stacked regression training; adjusting the participation weights of the two algorithms and optimizing the hyperparameters to construct a fusion XGBoost-LSSVR stacked machine learning model, thereby achieving high-precision prediction of vehicle attitude and acquisition head parameters and adapting to dynamic changes in seabed topography.

[0030] A stacked model is used to combine the feature extraction capabilities of XGBoost and the nonlinear fitting advantages of LSSVR to predict vehicle attitude changes and optimal acquisition head parameters within the next 1-5 seconds. LSSVR is solved under constraints:

[0031] In the formula For the weight vector, For error variables, The regularization parameter is set to 200. This is a kernel function mapping.

[0032] Preferably, in step S6, the method for dynamically adjusting the parameters of the acquisition head includes: when the acquisition efficiency is lower than the optimal efficiency threshold of 90%, or when the acquisition head loss rate or mineral breakage rate exceeds the preset range, adjusting the acquisition head height, angle, and jet velocity step by step according to the model prediction results, dynamically adjusting the weighting coefficients in conjunction with a multi-objective balance optimization mechanism, increasing the equipment safety weight when the terrain is drastic, increasing the efficiency weight when the terrain is flat, and continuously monitoring until the operating parameters are restored to the optimal range.

[0033] Preferably, in S7, the intelligent closed-loop control also includes an online learning and model update mechanism: newly collected "feature-optimal parameter-operation efficiency" data are added to the training set, and the model parameters are periodically updated using the incremental XGBoost algorithm to achieve adaptive matching of the model to changes in seabed topography and operating environment, ensuring the adaptability of the control strategy.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a multi-dimensional feature fusion-based seabed operation condition identification system, providing precise environmental and attitude data for the control of acquisition head parameters. Based on this, it dynamically matches optimal core parameters such as acquisition head height, angle, and jet velocity, realizing a transformation of acquisition control strategies from empirical and fixed to intelligent and dynamic. By building a two-layer machine learning architecture of XGBoost dual models and XGBoost-LSSVR stacked models, and embedding a multi-objective balance optimization mechanism, it fully leverages the advantages of different algorithms, significantly improving the accuracy and stability of parameter prediction in multi-factor coupled scenarios. Combining real-time monitoring from multi-source sensors and closed-loop control logic, it constructs a full-process intelligent control system of "data acquisition - model prediction - parameter execution - feedback optimization". Combined with a digital twin model and online learning mechanism, it can proactively adjust acquisition head parameters, effectively improving mineral recovery rate and reducing acquisition head wear. This significantly improves the safety and equipment adaptability of deep-sea mining operations while reducing the need for manual intervention, increasing operational efficiency, and lowering equipment maintenance costs and operational energy consumption. Attached Figure Description

[0035] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart illustrating an embodiment of the present invention; Figure 2This is a flowchart of the XGBoost model construction process of this invention; Figure 3 This is a flowchart of the feedback optimization layer of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] Example like Figure 1 The diagram shown is a schematic representation of the method flow in this embodiment, and the steps include: S1. Multi-source sensor data acquisition and preprocessing: Utilizing a combination of multiple sensors including lidar, pressure sensors, and attitude sensors, this system achieves comprehensive data acquisition for seabed mining operations. Acquired parameters include vehicle attitude parameters, seabed environment parameters, and acquisition head status parameters. All acquired data is transmitted to the data processing module in real-time. The module first performs a data validity check; if the data does not conform to a normal distribution... If the data is deemed abnormal, it needs to be re-collected or manually reviewed; if the data is normal, normalization preprocessing is performed to map all parameters to... Intervals were defined to eliminate dimensional differences and complete the initial data aggregation.

[0040] S2. Construct a standardized feature dataset. Based on the preprocessed multi-source acquisition data, construct a three-dimensional feature system including basic features, interactive features, and statistical features. Use coupled features of multiple factors such as terrain slope, mineral density, vehicle pitch angle, and roll angle as input features (X), and use the optimal height (β) and optimal angle (α) of the historical acquisition head as output targets (Y) to construct an "input-output" standardized dataset. The specific feature design is as follows: (1) Basic characteristics: directly taken from the preprocessed original parameters, including ; (2) Interaction characteristics: obtained through parameter coupling calculation, reflecting the synergistic influence of multiple factors, including (Pitch angle coupled with acquisition head angle) (Roll angle and height coupling) (Distance and posture combination); (3) Statistical characteristics: Calculated based on time series data, reflecting the dynamic changes in terrain, including the vehicle pitch angle in the past 10 seconds. Roll angle The variance of the fluctuation.

[0041] After integrating the three types of features, the feature values ​​are standardized using the Min-Max normalization formula, ultimately forming a structured CSV format dataset, which provides a data foundation for subsequent model training.

[0042] S3. XGBoost Dual-Model Training and Optimal Parameter Prediction. The standardized dataset is cleaned, and the cleaned data is used to train the XGBoost machine learning algorithm and perform K-value cross-validation. This constructs an XGBoost dual-model framework for angle and height prediction, achieving accurate prediction of the optimal control parameters for the acquisition head. The XGBoost implementation steps in this embodiment are as follows: (1) Set up the XGBoost development environment in PyCharm, install and import the necessary Python libraries, including XGBoost, pandas and scikit-learn, and write the core code of the XGBoost dual model on this basis.

[0043] (2) The standardized dataset is divided into training and testing sets in a 7:3 ratio. Missing values ​​(NANs) in the dataset are automatically assigned a classification direction by the XGBoost algorithm without additional supplementation. For values ​​that do not conform to normality... Outliers are removed according to established principles. The input feature X is defined as 12 features of a 3D feature system, and the output target Y is the optimal angle of the acquisition head. Optimal height .

[0044] (3) K-value cross-validation is used. In this embodiment, k is 8. The training set is divided into 8 equal subsets. In each iteration, 1 subset is selected as the validation set and the remaining 7 subsets are used for model training. The iteration is repeated 8 times. The iteration results are combined to evaluate the model performance and avoid overfitting.

[0045] (4) XGBoost Model Core Settings: The model objective function consists of prediction error and model complexity. An additive strategy is used to update the predicted values. A second-order Taylor expansion is performed on the loss function to transform the objective function into a form that is only related to the first-order and second-order gradients. At the same time, a regularization term is introduced to penalize model complexity, balancing prediction accuracy and generalization ability. Single tree construction adopts a splitting strategy from the root node to the leaf node, calculates the splitting gain, and selects the feature with the largest gain and the threshold for splitting. Then the splitting will stop.

[0046] (5) Model hyperparameter settings. In this embodiment, the hyperparameters optimized by cross-validation are as follows: n_estimators=300 (A total of 300 trees will be built, with the optimal range being 100-500). max_depth=6 (each tree can be up to 6 layers deep, with an optimal range of 3-8). learning_rate=0.05 (the weight of each tree's influence on the final result, preferably in the range of 0.01-0.1) gamma = 0.2 (the threshold for the objective function to decrease, preferably in the range of 0.1-0.5) reg_lambda=1.2 (Leaf node weight penalty coefficient, preferably in the range of 0.5-2.0) subsample=0.85 (the proportion of samples randomly drawn from each tree, preferably in the range of 0.7-1.0) colsample_bytree=0.8 (Random feature extraction ratio, preferably within the range of 0.7-1.0) random_state=42 (fixed random number seed) n_jobs=-1 (Number of CPU cores used during model training).

[0047] (6) Overfitting suppression: adopt the early stopping method and stop training when the validation set error does not decrease for 20 consecutive rounds.

[0048] (7) Performance evaluation: R², MAE, and RMSE are calculated using the test set to evaluate the model accuracy. In this embodiment, the angle prediction model... High-precision prediction model It meets the precision requirements of deep-sea mining engineering.

[0049] (8) Optimal parameter output: Input the preprocessed data collected in real time into the trained XGBoost dual model, and output the optimal angle of the acquisition head under the current working scenario. and optimal height .

[0050] During the training process, a multi-objective balance optimization mechanism is embedded, with collection efficiency as the core and mineral recovery rate and collection head loss rate as auxiliary objectives. The weighting coefficients are dynamically adjusted to adapt to different terrain control needs.

[0051] S4. Acquisition Head Parameter Execution and Real-time Monitoring: Optimal Acquisition Head Control Parameters Based on XGBoost Dual Model Prediction It sends commands to the hydraulic drive / servo motor drive device in the execution control layer to simultaneously complete three core control operations: (1) Adjust the angle of the acquisition head to α and the height of the acquisition head to β; (2) Adjust the vehicle attitude according to the parameters of the acquisition head to compensate for the pitch and roll deviations of the vehicle body, so that the acquisition head maintains the optimal relative position with the seabed; (3) Matching the jet velocity (v), in this embodiment the jet velocity is positively correlated with the height of the acquisition head; the higher the height, the greater the jet velocity. Specifically: hour hour .

[0052] During the control process, multi-source sensors continuously monitor the vehicle's attitude, seabed environment, and acquisition operation status in real time. The actual acquisition efficiency η is calculated by collecting the amount of minerals collected per unit time through the flow sensor. At the same time, the wear status of the acquisition head and the stability of the jet flow velocity are monitored to form multi-dimensional real-time operating data, which is transmitted to the feedback optimization layer.

[0053] S5. XGBoost-LSSVR stacked model training and look-ahead prediction: The real-time running data acquired in S4 is input into the XGBoost-LSSVR stacked machine learning model to achieve look-ahead prediction of the vehicle attitude change and optimal parameters of the acquisition head in the next operation stage (the preset prediction time period in this embodiment is 3 seconds). The specific implementation steps are as follows: (1) Model development environment: The core code of XGBoost-LSSVR was programmed on the MATLAB software platform, and integrated training was carried out using the "Machine Learning and Visualization" extension package; (2) Data preprocessing: Outliers were removed from the real-time running data using the 3σ principle, and rows with missing values ​​were deleted row by row. The input variables were determined to be the coupling parameters of vehicle attitude and seabed environment (including...). (etc.), the output variable is the vehicle attitude parameters predicted after 3 seconds. and optimal parameters of the acquisition head ; (3) Stacked model construction: A two-layer stacked architecture is adopted, which fully integrates the feature mining capability of the XGBoost algorithm and the nonlinear fitting advantage of the LSSVR algorithm: First-layer prediction: The XGBoost algorithm and LSSVR algorithm are pre-trained and predicted respectively to obtain their respective prediction results, forming the first-layer prediction data matrix; Second-layer stacked regression: The first-layer predicted data matrix is ​​used as a new input feature, and a linear regression model is trained as a meta-model. The participation weights of XGBoost and LSSVR are adjusted. In this embodiment, the optimized weight ratio is 7:3. (4) LSSVR algorithm hyperparameter settings: kernel='rbf' (radial basis kernel function, adapted for nonlinear regression), C=200 (regularization parameter), gamma=0.05 (kernel coefficient), epsilon=0.1 (insensitive band width); (5) Model training and validation: The stacked model was trained using k=10 fold cross-validation. Overfitting was suppressed by early stopping. Training was stopped when the validation set error did not decrease for 15 consecutive rounds. (6) Outlook prediction output: Input the preprocessed real-time data into the trained XGBoost-LSSVR stacked model and output the vehicle attitude prediction value after 3 seconds. And the forward prediction values ​​(α, β) of the optimal parameters of the acquisition head.

[0054] S6. Model Accuracy Verification and Dynamic Parameter Fine-Tuning: Compare the prediction results from S5 with the actual test set data to verify whether the accuracy of the XGBoost-LSSVR stacked machine learning model meets the requirements of deep-sea mining engineering operations. In this embodiment, the model accuracy judgment criterion is the AUC area. Overall prediction error Perform different operations based on the verification results: (1) Accuracy achieved: The height and angle of the acquisition head are dynamically fine-tuned based on the forward prediction values ​​(α, β), with the fine-tuning step size being the angle. ,high This enables proactive adaptation of the acquisition head parameters; (2) Accuracy not met: Return to step S3, re-optimize the hyperparameters of the XGBoost dual model, add new real-time data to the training set, retrain the model and verify again until the accuracy meets the requirements.

[0055] When the acquisition efficiency η is lower than 90% of the optimal acquisition efficiency threshold, or the acquisition head loss rate... Mineral breakage rate When the optimal range is not restored within 10 seconds, the step-by-step parameter adjustment mechanism is triggered: first, the angle / height of the acquisition head is adjusted; if the optimal range is not restored within 10 seconds, the jet velocity is adjusted accordingly, and monitoring continues until the operating parameters are restored to the preset optimal range.

[0056] The determination of the 90% threshold for collection efficiency η is based on statistical analysis of historical operational data (cumulative collection time exceeding 500 hours) from a deep-sea polymetallic nodule mining test area. When the collection efficiency is lower than 90% of the optimal efficiency, the decrease in mineral recovery per unit time significantly exceeds the energy savings, resulting in a marked reduction in economic efficiency. Meanwhile, the collection head wear rate threshold of 0.02% / h and the mineral breakage rate threshold of 8% are determined comprehensively based on the equipment design life (cumulative operating life of the collection head not less than 2000 hours) and mineral quality requirements (a breakage rate higher than 8% will affect subsequent mineral processing).

[0057] To implement efficiency monitoring and model update functions in the feedback optimization layer, this embodiment configures the following hardware and software modules: Efficiency Monitoring Module: Composed of a flow sensor, a timer, and a controller (embedded industrial computer). The flow sensor is installed at the inlet of the mineral conveying pipeline to collect the amount of mineral collected per unit time (unit: kg / s) in real time; the timer records the collection duration; the controller calculates the actual collection efficiency η according to the formula η = (actual collection amount / theoretical maximum collection amount) × 100%, and compares it with the preset "optimal collection efficiency 90% threshold". When η is lower than this threshold for 10 consecutive seconds, the controller determines that the current control strategy has failed and triggers the model update module.

[0058] The model update module consists of non-volatile memory (storing historical training datasets and current XGBoost dual-model parameters) and an incremental learning unit (an online learning algorithm integrated into the controller). Upon receiving a trigger signal from the efficiency monitoring module, the controller stores the "feature-optimal parameter-actual efficiency" data (300 sampling points at a sampling frequency of 10Hz) from the last 30 seconds as incremental training samples in the memory. Then, it calls the incremental XGBoost algorithm, using these newly added samples as the primary data source and historical samples as secondary data sources (weighted by time decay), to iteratively update the parameters of the XGBoost dual-model. The number of iterations is set to 50 rounds, and the learning rate is maintained at 0.05. The updated model parameters overwrite the original model and are used for subsequent header parameter prediction. The entire update process is completed within 1 second and does not affect continuous operation.

[0059] The efficiency monitoring and model update described above together constitute the feedback optimization layer; its closed-loop logic is shown in the appendix. Figure 3 (Flowchart of feedback optimization layer). Through this mechanism, the system can adapt to the time-varying characteristics of seabed topography and maintain control accuracy over a long period of time.

[0060] S7. Intelligent closed-loop control and online model update: Based on the optimal parameters of the initial acquisition head set by the XGBoost algorithm, combined with the look-ahead prediction fine-tuning results of the XGBoost-LSSVR stacked model, it continuously adapts and adjusts the jet velocity and the overall attitude of the vehicle body, and simultaneously conducts real-time monitoring of the entire acquisition process, constructing an intelligent closed-loop control system of "data acquisition - model prediction - parameter execution - feedback optimization", specifically implemented as follows: (1) The sensor continuously inputs real-time monitoring data into the XGBoost-LSSVR stacked model, and the model dynamically outputs the adjusted acquisition head control parameters. The actuator responds and executes in real time, and the closed-loop control logic response delay is controlled in milliseconds. (2) The model prediction layer uses real-time attitude data to infer the type of seabed topographic relief and matches differentiated control strategies: gradual slope adopts smooth parameter transition (small adjustment step size and low frequency), and sudden bulge / depression adopts rapid response adjustment (large adjustment step size and high frequency). (3) Introduce a multi-objective balance optimization mechanism with collection efficiency as the core objective and mineral recovery rate and collection head loss rate as auxiliary objectives. Dynamically adjust the weighting coefficients: increase the equipment safety weight (weighting coefficient accounts for 50%) when the terrain is drastic and reduce the collection head loss rate; increase the efficiency weight (weighting coefficient accounts for 70%) when the terrain is flat and maximize collection efficiency and mineral recovery rate. (4) Online learning and updating of the model: An online learning mechanism is constructed to continuously add newly collected "feature-optimal parameter-operation efficiency" data to the training set. The incremental XGBoost algorithm is used to periodically update the XGBoost dual model parameters every 30 minutes to achieve adaptive matching of the model to changes in seabed topography and operating environment, and to ensure the accuracy of parameter prediction in the long term.

[0061] Meanwhile, in response to extreme seabed environments, this embodiment adds an adaptive filtering and feature enhancement module to the data perception layer to improve the accuracy of attitude data detection; designs a redundant control command mechanism in the execution control layer to avoid control inaccuracies caused by environmental interference; and constructs a digital twin model of seabed mining scenarios to pre-train parameter control strategies under extreme terrains such as steep slopes (slope > 15°) and gullies, further improving the system's adaptability to complex seabed terrain.

[0062] Through the above steps, this embodiment realizes intelligent closed-loop control of the seabed mining vehicle's collection head under the coupled influence of multiple factors. The application results in a deep-sea polymetallic nodule mining test area show that the mineral recovery rate is increased to over 92%, the collection head loss rate is reduced to below 0.01% / h, the collection efficiency is increased by 35%, and the need for manual intervention is reduced by 80%, which significantly improves the safety, efficiency and intelligence level of deep-sea mining operations.

[0063] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for intelligent control of a seabed mining vehicle's data acquisition head considering multi-factor coupling, characterized in that... include: S1. The data perception layer first collects multi-source sensor data through lidar, pressure sensor, acquisition sensor and attitude sensor. The collected parameters are transmitted to the data processing module for aggregation. The data processing module selects vehicle attitude, seabed environment and acquisition head status parameters as input features, and acquisition head height and angle adjustment as output targets. After normalization processing, a standardized dataset is constructed. S2. Clean the constructed standardized dataset and use the cleaned data for XGBoost model training. By setting a reasonable initial range of hyperparameters, the learning rate is set to 0.05, the tree depth is set to 6, and the minimum sample weight of the leaf node is set to 3, laying the foundation for model training. S3. K-value cross-validation is used to optimize the hyperparameters of the XGBoost dual model. The cleaned dataset is divided into K subsets, which are used alternately as training and validation sets for iterative training to minimize model error and avoid overfitting. After training, the test data is input to predict the optimal angle α* and optimal height β* of the acquisition head, which constitutes the model prediction layer of this intelligent control method. S4. The execution control layer sends instructions to the hydraulic drive mechanism or servo motor drive mechanism according to the optimal parameters of the acquisition head, and synchronously adjusts the vehicle body posture, acquisition head parameters and jet flow velocity. The sensor collects and feedback data in real time, and uses the 3σ principle to remove outliers and retain valid data. S5. The feedback optimization layer inputs effective feedback data into the XGBoost-LSSVR stacked model. XGBoost extracts features and makes preliminary predictions. The results are input into LSSVR for further fitting. Combining environmental and attitude trends, the optimal parameters and attitude changes are predicted within a preset time period. S6. Based on the comparison between the prediction results and the test set, verify whether the accuracy of the XGBoost-LSSVR stacked machine learning model meets the engineering requirements. If the accuracy meets the requirements, dynamically fine-tune the acquisition head parameters. If the accuracy does not meet the requirements, return to S3 to re-optimize the model hyperparameters until the requirements are met. S7. Combining the initial optimal parameters output from S3 with the adjustment results from S6, the real-time collected data is input into the XGBoost-LSSVR stacked model to obtain real-time adjustment parameters; the feedback optimization layer embeds an incremental XGBoost dual model, using historical and real-time data as samples to periodically iterate and update the angle and height prediction models; the closed-loop logic outputs compensation commands for angle, height, and jet velocity in real time based on the updated dual models, driving the execution layer to form a closed-loop intelligent control.

2. The intelligent control method for the seabed mining vehicle's data acquisition head considering multi-factor coupling as described in claim 1, characterized in that, The intelligent control system that implements this method includes a data perception layer, a model prediction layer, an execution control layer, and a feedback optimization layer; Data perception layer: includes lidar, pressure sensor, acquisition sensor and attitude sensor, used to collect vehicle attitude parameters, seabed environment parameters and acquisition head status parameters, and transmit them to the data processing module for normalization processing; Model prediction layer: includes XGBoost dual model and XGBoost-LSSVR stacked model. The XGBoost dual model is used to predict the initial optimal angle α* and optimal height β* of the acquisition head. The XGBoost-LSSVR stacked model is used to predict the optimal parameters and attitude changes within a preset time period. Execution control layer: includes hydraulic drive mechanism or servo motor drive mechanism, used to receive control commands and drive the acquisition head to adjust the angle, height and jet flow rate; Feedback optimization layer: includes an efficiency monitoring module and a model update module. The efficiency monitoring module is used to calculate the actual acquisition efficiency η and compare it with a threshold. The model update module is used to trigger online learning when the efficiency is lower than the threshold and iteratively update the XGBoost dual model parameters.

3. The intelligent control method for the seabed mining vehicle's data acquisition head considering multi-factor coupling as described in claim 1, characterized in that, When constructing the standardized dataset in S1, features are selected and output targets are determined through the feature engineering module. Three types of features are constructed to adapt to the subsequent XGBoost model training: the basic features are taken from various parameters after preprocessing. Interactive features are calculated through parameter coupling; Statistical characteristics are calculated based on time-series data, reflecting terrain stability and providing a basis for model optimization and trend prediction. The basic features include θ (pitch angle), φ (roll angle), H (seafloor distance), S (mineral concentration), R (water flow velocity), D (scan head spacing), α0 (initial angle of the scan head), and β0 (initial height of the scan head); Interactive features include θ×α0, φ×β0, and H×(α0+β0), which are used to improve the model's accuracy in recognizing parameter relationships. Statistical features include the variance of θ / φ over the past 10 seconds, which reflects terrain stability and provides terrain reference for S3 model optimization and S5 trend prediction.

4. The intelligent control method for the seabed mining vehicle's data acquisition head considering multi-factor coupling as described in claim 1, characterized in that, The specific method for predicting the optimal angle α* and optimal height β* using the XGBoost dual model in S2-S3 includes: using vehicle attitude, seabed environment, and acquisition head state parameters as input features, wherein the vehicle attitude, seabed environment, and acquisition head state parameters respectively include: (1) Vehicle attitude parameters: vehicle pitch angle (θ), vehicle roll angle (φ); (2) Seafloor environmental parameters: topographic slope (S), mineral density (D), sediment thickness (R); (3) Acquisition head status parameters: acquisition head angle (α0), acquisition head height (β0), jet velocity (v); The optimal angle α* and optimal height β* of the acquisition head are used as the output targets.

5. The intelligent control method for the seabed mining vehicle's data acquisition head considering multi-factor coupling as described in claim 1, characterized in that, In step S3, K-value cross-validation is used to optimize hyperparameters, finding a set of optimal hyperparameters that maximize the model's generalization ability on unseen data, thus avoiding overfitting. The steps include: (1) The cleaned and standardized dataset D is divided into a training set and a test set in a 7:3 ratio. The test set is only used in the final evaluation and does not participate in cross-validation. (2) Determine that K is 8, and randomly divide the training set into 8 subsets; (3) Perform an outer loop to perform a grid search for candidate combinations for each set of candidate hyperparameters; (4) Perform the inner loop and calculate the average score of K validations for each group of hyperparameters. The calculation formula is as follows: ； (5) Select the optimal hyperparameter and compare the average score of all candidate hyperparameter combinations. avg The optimal combination of evaluation metrics is selected as the final hyperparameter. (6) The model retrained on the full training set using the optimal hyperparameters is then tested on the test set to verify whether the accuracy meets R. 2 Requirements: ≥0.92, angle ≤0.5 degrees, height ≤2 cm.

6. The intelligent control method for the seabed mining vehicle's data acquisition head considering multi-factor coupling as described in claim 1, characterized in that, The objective function of the XGBoost dual model in S3 consists of prediction error and model complexity, and its expression is as follows: ； In the formula, Obj represents the overall objective function of the Kth iteration, and L(y) i ,y' i Ω(f) is the loss function used to measure the error between the predicted and actual values. k ) represents the complexity penalty for the k-th tree, used to control model overfitting; The XGBoost model uses an additive strategy to update predicted values. It performs a second-order Taylor expansion of the loss function, transforming the objective function into a form that depends only on the first and second gradients. Simultaneously, a regularization term is introduced to penalize model complexity, balancing prediction accuracy and generalization ability. The second-order Taylor expansion of the loss function is as follows: ； After removing the constant term, it simplifies to: ； ， ； In the formula g i f is the partial derivative of the loss function with respect to the previous round of predictions, reflecting the direction and magnitude of the error between the current prediction and the true value. t (X i ) is the output value of the t-th lesson tree for sample Xi, h i The second-order partial derivative of the loss function with respect to the previous round of predictions reflects the curvature of the loss function. The regularization term for the t-th tree is used to prevent overfitting, and its expression is: ； In the formula: T is the number of leaf nodes in the tree, γ is the penalty coefficient for the number of leaf nodes used to control the structural complexity of the tree, λ is the regularization coefficient used to control the smoothness of the leaf weights, and w j This is the output value of the j-th leaf node.

7. The intelligent control method for the seabed mining vehicle's data acquisition head considering multi-factor coupling as described in claim 1, characterized in that, The hydraulic drive mechanism or servo motor drive mechanism in S4 is a hydraulic drive or servo motor drive device that receives α from the controller. 、β Commands such as jet velocity are used to simultaneously adjust the vehicle's attitude, acquisition head parameters, and jet velocity, ensuring precise control actions and adapting to S4 commands and S7 closed-loop control logic.

8. The intelligent control method for the seabed mining vehicle's data acquisition head considering multi-factor coupling as described in claim 1, characterized in that, In step S5, the method for constructing the XGBoost-LSSVR stacked model includes: pre-training and predicting XGBoost and LSSVR respectively to obtain a first-layer prediction data matrix; using the first-layer data matrix as new input features for stacked regression training; adjusting the weights of the two algorithms and optimizing hyperparameters to construct a fusion stacked model, wherein the preset prediction time period is 1-5 seconds. The specific prediction process is as follows: (1) The first layer of prediction in stacked regression, with N samples, is trained and predicted using XGBoost and LSSVR respectively: ， ； In the formula, X is the input feature matrix, which includes features such as vehicle attitude, seabed environment, and head state. XGB y' LSSVR The prediction results are shown for the two models, respectively. (2) Use the prediction results of the first layer as new features ； ； The model is trained using linear regression to obtain the final prediction: w XGB w LSSVR w is the weighting coefficient. XGB =0.7、w LSSVR =0.3, b is the bias term.

9. The intelligent control method for the seabed mining vehicle's data acquisition head considering multi-factor coupling as described in claim 1, characterized in that, In S6, the method for dynamically adjusting the parameters of the acquisition head includes: when the acquisition efficiency is lower than the optimal efficiency threshold of 90%, or when the acquisition head loss rate or mineral breakage rate exceeds the preset range, adjusting the acquisition head height, angle and jet velocity step by step according to the model prediction results, dynamically adjusting the weighting coefficients in combination with the multi-objective balance optimization mechanism, and continuously monitoring until the operating parameters are restored to the optimal range.

10. The intelligent control method for the seabed mining vehicle's data acquisition head considering multi-factor coupling as described in claim 1, characterized in that, The incremental XGBoost dual model in S7 uses historical and real-time data as samples to periodically iterate and update the angle and height prediction models. The closed-loop logic outputs compensation commands for angle, height, and jet velocity in real time based on the updated dual model, driving the execution layer to form a closed-loop intelligent control. At the same time, newly collected feature-optimal parameter-operation efficiency data are added to the training set to achieve adaptive matching of the model to changes in the operating environment.

Images