Machine learning based analysis method and system for decoupling harrow head digging yield data

Abstract

The application provides a machine learning-based harrow head digging yield data decoupling analysis method and system, and belongs to the technical field of data decoupling analysis. The method comprises the following steps: acquiring multi-source historical data; pre-processing the multi-source historical data; constructing an initial feature set from the time lag correction data and calculating instantaneous yield; using four machine learning algorithms to evaluate the feature importance of the initial feature set with the instantaneous yield as the target, and obtaining the importance scores of the features; performing coupling analysis on the initial feature set, detecting the linear correlation between the features by using a Pearson correlation coefficient, identifying and removing redundant features, and obtaining a key feature set; and constructing a yield prediction model, taking the key features as inputs, and outputting an instantaneous yield prediction result. The application combines expert experience and four machine learning algorithms to systematically screen key influencing factors from numerous feature variables in the original data, and clearly determines the direction for establishing an accurate yield prediction model.

Description

Technical Field

[0001] This invention relates to the fields of statistical analysis and machine learning, and data decoupling analysis, and in particular to a machine learning-based method and system for decoupling analysis of rake excavation production data. Background Technology

[0002] Trailing suction hopper dredgers are key construction equipment commonly used in port dredging and waterway improvement. They use a skid head to suck up seabed sediment and rely on a mud pump system to transport it to the mud tank. During construction, instantaneous output (i.e., the amount of dredged material per unit time) is one of the core indicators for evaluating operational efficiency. The level of instantaneous output is not only affected by the performance of the equipment itself, but also closely related to actual operating parameters (such as skid head posture, high-pressure water jet pressure, and vessel speed). With the development of intelligent construction, dredgers are equipped with a large number of sensors that can record various operating parameters in real time, including flow velocity, density, skid head angle, and high-pressure water jet pressure. Identifying the key features affecting the instantaneous output of the trailing suction hopper from these multi-source time-series data is one of the key steps in achieving intelligent optimization of the construction process.

[0003] In the field of dredger productivity prediction and key factor identification, numerous domestic scholars have conducted extensive research. Yang et al. from Hohai University used a BP neural network to predict dredger productivity, but their selection of only three factors based on manual experience resulted in poor prediction performance. Fu et al. from Tianjin University, based on operator experience, combined elastic regression networks and support vector machines to screen 10 features affecting the operational capacity of cutter suction dredgers, achieving good results in productivity estimation. Bai Shuo et al. from Tianjin University used the Lasso and maximum information coefficient (MIC) redundancy identification method to explore potential factors affecting the dredging efficiency of cutter suction dredgers, quantifying and ranking their importance into three levels. Furthermore, they constructed a filter-based feature selection model to identify key factors of capacity, analyzing the correlation between productivity and influencing factors using Pearson correlation coefficients, but did not conduct in-depth quantitative coupling analysis between factors. Yang et al. from Wuhan University of Technology used principal component analysis to screen key features of cutter suction dredger productivity, confirming that the top 10 principal components can characterize over 97% of the overall data. Tongji University's Shanggang et al. used the HSIC Lasso method to extract key performance characteristics of cutter suction dredgers, combined with the limit gradient lift algorithm to determine the importance of these characteristics and conduct a qualitative discussion. Shanghai Zhenhua Heavy Industries' Tian Jinzhu et al., based on dredging experience, used grey relational analysis to study the impact of trailing suction hopper dredger pump parameters on productivity, but did not fully consider soil type and construction environment factors. These studies have promoted the development of this field from different perspectives and also pointed out directions for improvement in subsequent research.

[0004] Foreign dredging companies, represented by IHC in the Netherlands, have established mathematical models to study the relationship between dredging head excavation and related control parameters. However, the control parameters selected based on manual experience are relatively simple. For example, IHC established a calculation model for the ground clearance of wear-resistant blocks and conducted a mechanism analysis of the correlation between dredging head excavation and the rake lip angle. In China, China Harbour Dredging Co., Ltd. has conducted relevant research and testing using actual ship data. They manually selected four parameters affecting construction efficiency: ship speed to ground, wave compensator pressure, rake head movable cover angle, and lower rake pipe angle. They used a neural network algorithm to roughly predict dredging capacity and conducted a parameter coupling mechanism analysis based on a simplified mathematical model.

[0005] In summary, the following problems exist:

[0006] 1. Limitations of Traditional Dredging Parameter Analysis: Production analysis of traditional trailing suction hopper dredgers relies on manual experience in parameter selection (such as focusing only on single parameters like speed and dredger head angle), lacking data-driven systematic feature screening. This results in weak model generalization ability and difficulty in adapting to complex working conditions (such as soil changes and water flow disturbances). Early studies both domestically and internationally (such as the mechanistic model of the Dutch IHC and the BP neural network in China) suffer from insufficient feature dimensions, neglect of time delay effects, and lack of quantification of coupling relationships, making it impossible to accurately analyze the interactive effects of multiple parameters.

[0007] 2. Industry Demand for Data-Driven Technologies: With the increasing deployment of sensors on dredgers, multi-source time-series data (equipment parameters, environmental data, construction logs) accumulates. However, this data suffers from problems such as high noise levels (e.g., sensor jitter, missing values), significant time lags (e.g., the 30-second delay in adjusting the rake head angle affects output), and feature redundancy (e.g., the compensator pressure is correlated with the stroke height). There is an urgent need to extract key features from massive amounts of data through data preprocessing, time lag correction, and machine learning algorithms to support intelligent dredging.

[0008] 3. Technological bottlenecks in the intelligentization of high-end equipment: Domestically produced large trailing suction hopper dredgers previously relied on foreign technologies (such as vessels with a capacity of 30,000 m³ or more) and intelligent control systems (such as "one-click dredging" and energy efficiency optimization). Traditional ships have high energy consumption and low automation levels, requiring breakthroughs in core technologies such as green and low-carbon design and intelligent control through AI technology. Summary of the Invention

[0009] This invention provides a machine learning-based method and system for decoupling and analyzing rake excavation output data. It utilizes statistical analysis and machine learning methods to assess feature importance, identifying feature parameters that have a significant impact on the construction process. By employing statistical analysis and machine learning, it deeply analyzes the correlation between equipment parameters, environmental factors, management factors, and operational output, identifying key efficiency features. Coupled analysis is then performed on the initially selected features to derive a final set of key features. This invention integrates expert experience with algorithms such as Random Forest, LightGBM, Lasso, and Catboost to systematically screen numerous feature variables in the original data, highlighting key influencing factors and providing a clear direction for establishing an accurate output prediction model.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] Machine learning-based methods for decoupling and analyzing rake digging yield data include:

[0012] Step (1) Obtain multi-source historical data, which includes equipment sensor data, environmental monitoring data, and construction log data;

[0013] Step (2) preprocesses the multi-source historical data, including: a. identifying and processing outliers using box plots and the 3σ method to obtain preliminary cleaned data; b. filling missing values ​​in the preliminary cleaned data by forward imputation or interpolation for continuous variables and by mode or most recent effective value for discrete variables to obtain imputed data; c. filtering the right-hand parameters in the imputed data, retaining only fields related to the right rake head to obtain right-hand parameter data; d. standardizing the right-hand parameter data by mapping the data to the [0,1] interval or a distribution with a mean of 0 and a standard deviation of 1 using the Min-Max method or the Z-score method to obtain standardized data; e. performing time lag correction on the standardized data by using the maximum cross-correlation coefficient method, sliding window traversal + R 2 The evaluation method and the manual experience method are used to determine the time delay between parameters and output, adjust the data to align the causal relationship, and obtain time delay corrected data;

[0014] Step (3) Construct an initial feature set from the time-delay correction data based on expert experience, and calculate the instantaneous output of the right harrow as the target parameter;

[0015] Step (4) Based on the target parameters, four machine learning algorithms, namely Random Forest, LightGBM, Lasso and CatBoost, are used to evaluate the feature importance of the initial feature set and obtain the importance score of each feature.

[0016] Step (5) Based on the importance scores of each feature, perform coupling analysis on the initial feature set, use Pearson correlation coefficient to detect the linear correlation between features, identify redundant features, retain features that have a more direct impact on yield, and obtain the key feature set.

[0017] Step (6) Construct a production prediction model based on the set of key features, using the key features as input, and output the instantaneous production prediction result; the production prediction model adopts a regression algorithm;

[0018] Step (7) Based on the importance score of the key features, generate construction parameter optimization suggestions to adjust operation parameters such as the right rake head angle and the right suction vacuum.

[0019] In this specification, in step (4), the random forest algorithm evaluates feature importance based on the Gini index, out-of-bag data, and Shapley value; the LightGBM algorithm evaluates feature importance based on split frequency, gain, and Shapley value; the Lasso algorithm filters features through L1 regularization; and the CatBoost algorithm evaluates feature importance based on the degree of optimization of the objective function by the decision tree.

[0020] In this specification, in step (5), the threshold of the Pearson correlation coefficient is set to an absolute value greater than 0.8, that is, the two features are highly correlated, and the more critical features are retained based on the significance of the feature's impact on the yield.

[0021] In this specification, in step (6), the yield prediction model is constructed using a gradient boosting regression algorithm, which learns the nonlinear relationship between key features and yield through training data.

[0022] In this instruction manual, step (7) includes the parameter optimization suggestions, which include adjusting the operating parameters corresponding to the key features: the right active rake head angle, the right suction vacuum, and the right upper vertical angle, in order to improve instantaneous output.

[0023] In this specification, in step (3), the initial feature set includes the right active rake head angle, the right high-pressure flushing pressure, and the right suction vacuum. The instantaneous output is calculated by the right concentration and the right flow rate.

[0024] In this specification, the processing of the redundant features in step (5) is as follows: if the correlation coefficient between the right compensator stroke and the right compensator pressure is greater than the threshold, then the right compensator pressure is retained; if the right mud pump speed is related to the main engine speed, then the right mud pump speed is retained.

[0025] In this specification, in step (6), the input data of the production prediction model needs to be preprocessed in step (2) and coupled in step (5) to ensure that the causal relationship between features and production is consistent and without redundancy.

[0026] The machine learning-based harrow digging production data decoupling analysis system, applying any one of the machine learning-based harrow digging production data decoupling analysis methods described above, comprises:

[0027] The data acquisition module is used to acquire multi-source historical data, including equipment sensor data, environmental monitoring data, and construction log data.

[0028] The preprocessing module is used to preprocess the multi-source historical data, including: a. performing outlier processing to obtain preliminary cleaned data; b. filling missing values ​​in the preliminary cleaned data to obtain filled data; c. filtering the right-hand parameters in the filled data, retaining only fields related to the right rake head to obtain right-hand parameter data; d. performing feature standardization on the right-hand parameter data to obtain standardized data; e. performing time lag correction on the standardized data to adjust the data alignment causal relationship to obtain time lag corrected data.

[0029] The construction and calculation module is used to construct an initial feature set from the time-delay correction data and calculate the instantaneous output;

[0030] The evaluation module is used to evaluate the feature importance of the initial feature set with instantaneous output as the objective, using four machine learning algorithms: Random Forest, LightGBM, Lasso, and CatBoost, to obtain the importance score of each feature.

[0031] A redundancy processing module is used to perform coupling analysis on the initial feature set based on the importance scores of each feature, use the Pearson correlation coefficient to detect the linear correlation between features, identify and remove redundant features, and obtain a set of key features.

[0032] The prediction module is used to construct a production prediction model based on the set of key features, taking the key features as input and outputting instantaneous production prediction results; the production prediction model adopts the gradient boosting regression algorithm.

[0033] In summary, the present invention has at least the following beneficial effects:

[0034] 1. Improved Data Quality and Feature Analysis Capabilities: Outlier detection (e.g., 3σ method + manual review) and time lag correction (e.g., maximum cross-correlation coefficient method) eliminate data noise and time misalignment, ensuring causal alignment between features and yield, and providing reliable input for the model. Based on multi-algorithm fusion (random forest, LightGBM, Lasso, etc.), it overcomes the limitations of single algorithms, accurately identifying core features such as "right suction vacuum" and "right active rake head angle," revealing implicit correlations between parameters (e.g., the synergistic effect of high-pressure flushing pressure and mud suspension effect).

[0035] 2. Model Accuracy and Construction Efficiency Optimization: A coupled analysis model is constructed, redundant features are eliminated (such as compensator stroke → retained pressure), model complexity is reduced, and the stability of production prediction is improved, providing parameter optimization basis for "one-click dredging". Combining expert experience with AI, intelligent optimization of construction parameters is achieved (such as dynamic adjustment of rake head posture), reducing manual intervention, and approaching "unmanned dredging" under typical working conditions, significantly improving operation efficiency (qualitative description, avoiding numerical values).

[0036] 3. Industry Technological Breakthroughs and Domestic Production Support: Breaking the foreign technological blockade on intelligent control of large trailing suction hopper dredgers, supporting the domestic production of 30,000m... 3 + Independent design of hull capacity and vessel type (such as the "Jun Guang" vessel) to achieve full automation and low carbon emissions in dredging operations (such as methanol dual-fuel power and green vessel design). Promote the dredging industry to shift from "experience-driven" to "data-driven", provide a technical paradigm for the precise maintenance of complex waterways such as the Yangtze River Estuary (such as sediment dynamic response analysis), and help ensure the safe and smooth flow of the "golden waterway".

[0037] 4. Generalization and scenario adaptation: The time delay correction and feature decoupling methods can be transferred to other dredging equipment (such as cutter suction dredgers), adapting to complex working conditions with different soil types and water depths, and improving the overall intelligence level of the industry. This lays the foundation for the subsequent integration of AI and dredging mechanisms (such as sediment movement modeling) and supports cutting-edge scenarios such as deep-sea operations (such as 120-meter deep-sea trench excavation). Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the machine learning-based method for decoupling and analyzing rake excavation production data involved in this invention.

[0039] Figure 2 This is a schematic diagram illustrating the basic data information involved in this invention.

[0040] Figure 3 This is a schematic diagram illustrating the original data names involved in this invention.

[0041] Figure 4a , Figure 4b These are schematic diagrams of frequency histograms for GPS 1# North latitude (°) and GPS 1# East longitude (°).

[0042] Figure 5a , Figure 5b These are schematic diagrams of the frequency histograms of GPS1 ground speed (knots) and GPS1 ground heading (°).

[0043] Figure 6a , Figure 6b These are schematic diagrams of frequency histograms for dredging time (hours) and tide level (m), respectively.

[0044] Figure 7a , Figure 7b These are schematic diagrams of frequency histograms for tilt (m) and yaw (m), respectively.

[0045] Figure 8a , Figure 8b , Figure 8c , Figure 8d , Figure 8e , Figure 8f , Figure 8g , Figure 8h , Figure 8i , Figure 8j , Figure 8k These are schematic diagrams of frequency histograms for the following parameters: right concentration (%, solid mass fraction), right rake head depth (m), right upper vertical angle (°), right compensator pressure (MPa), right lower vertical angle (°), right high-pressure flushing pressure (MPa), right active rake head angle (°), right flow velocity (m / s), right active rake head stroke (m), right density (t / m³), and right active rake head pressure (MPa).

[0046] Figure 9 This is a schematic diagram illustrating the types of valid data involved in this invention.

[0047] Figure 10 This is a schematic diagram of the dredging construction data screening and labeling process involved in this invention.

[0048] Figure 11 This is a schematic diagram illustrating the potential influencing factors of the instantaneous output of the trailing suction hopper dredger involved in this invention.

[0049] Figure 12 This is a schematic diagram of feature selection using the random forest algorithm involved in this invention.

[0050] Figure 13 This is a schematic diagram illustrating the LightGBM feature selection involved in this invention.

[0051] Figure 14 This is a schematic diagram of the convergence plot of the Lasso characteristic coefficients involved in this invention.

[0052] Figure 15 This is a schematic diagram illustrating the Lasso algorithm used in this invention to select the top twelve important features.

[0053] Figure 16 This is a schematic diagram illustrating how the Catboost algorithm used in this invention selects the top fifteen most important features.

[0054] Figure 17 This is a schematic diagram of the Pearson correlation coefficient heatmap involved in this invention. Detailed Implementation

[0055] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0056] like Figure 1 As shown, this embodiment provides a machine learning-based method for decoupling and analyzing rake digging production data, including:

[0057] Step (1) Obtain multi-source historical data, which includes equipment sensor data, environmental monitoring data, and construction log data;

[0058] Step (2) preprocesses the multi-source historical data, including: a. identifying and processing outliers using box plots and the 3σ method to obtain preliminary cleaned data; b. filling missing values ​​in the preliminary cleaned data by forward imputation or interpolation for continuous variables and by mode or most recent effective value for discrete variables to obtain imputed data; c. filtering the right-hand parameters in the imputed data, retaining only fields related to the right rake head to obtain right-hand parameter data; d. standardizing the right-hand parameter data by mapping the data to the [0,1] interval or a distribution with a mean of 0 and a standard deviation of 1 using the Min-Max method or the Z-score method to obtain standardized data; e. performing time lag correction on the standardized data by using the maximum cross-correlation coefficient method, sliding window traversal + R 2 The evaluation method and the manual experience method are used to determine the time delay between parameters and output, adjust the data to align the causal relationship, and obtain time delay corrected data;

[0059] Step (3) Construct an initial feature set from the time-delay correction data based on expert experience, and calculate the instantaneous output of the right harrow as the target parameter;

[0060] Step (4) Based on the target parameters, four machine learning algorithms, namely Random Forest, LightGBM, Lasso and CatBoost, are used to evaluate the feature importance of the initial feature set and obtain the importance score of each feature.

[0061] Step (5) Based on the importance scores of each feature, perform coupling analysis on the initial feature set, use Pearson correlation coefficient to detect the linear correlation between features, identify redundant features, retain features that have a more direct impact on yield, and obtain the key feature set.

[0062] Step (6) Construct a production prediction model based on the set of key features, using the key features as input, and output the instantaneous production prediction result; the production prediction model adopts a regression algorithm;

[0063] Step (7) Based on the importance score of the key features, generate construction parameter optimization suggestions to adjust operation parameters such as the right rake head angle and the right suction vacuum.

[0064] The output prediction model employs an improved LightGBM regression algorithm. The construction process includes: (a) dividing the training and testing sets: using the key feature set as input variables and instantaneous output as the target variable, the dataset is randomly divided in an 80%:20% ratio, and 5-fold cross-validation is used to avoid overfitting; (b) setting model hyperparameters: learning rate 0.05, decision tree depth 6, number of iterations 500, L2 regularization coefficient 0.01, and optimizing the hyperparameters through grid search; (c) model training and validation: using mean squared error (MSE) as the loss function, after training, the test set determination coefficient R² ≥ 0.85 and mean absolute error (MAE) ≤ 0.5 m³ / min; finally, using the key features as input, the instantaneous output prediction result is output.

[0065] The technical concept of this invention is as follows:

[0066] 1. Research Objectives: To extract key parameters from historical construction data, perform data cleaning and preprocessing to establish a complete and accurate construction dataset. Statistical analysis and machine learning methods will be used to assess feature importance, identifying the feature parameters that have a significant impact on the construction process. The correlation between equipment parameters, environmental factors, management factors, and operational output under different working conditions will be analyzed, providing a solid foundation for subsequent optimization of operational parameters.

[0067] 2. Data Feature Identification of Trailing Suction Hoist Dredgers

[0068] 2.1 Research Objectives: This invention aims to construct a feature identification and coupling analysis model for trailing suction hopper dredgers. First, historical data from multiple sources, including construction logs and equipment sensor monitoring, is cleaned and preprocessed. Statistical methods and data processing algorithms are used to remove invalid data and fill in missing values ​​to ensure data quality. Second, combining professional knowledge and practical experience, statistical analysis and machine learning methods are used to deeply analyze the correlation between equipment parameters, environmental factors, management factors, and operational output, and to screen key feature variables. Finally, based on the cleaned data, key performance features are identified, and coupling analysis is performed on the initially selected features to derive the final set of key features.

[0069] 2.2 Main Research Content and Key Technologies

[0070] 2.2.1 Main Research Content

[0071] (1) Extracting key parameters from historical construction data for data cleaning and preprocessing: The core task of this stage is to comprehensively sort out and standardize the historical data accumulated by the trailing suction hopper dredger. Due to the complex operation scenarios of dredgers and the wide range of data sources, including construction logs, equipment sensor monitoring data, environmental monitoring data, etc., there is a large amount of duplicate, missing, and erroneous information. The data screening and cleaning stage will use statistical methods and data processing algorithms to identify and correct the raw data, remove invalid data, fill in missing values, and correct erroneous records to ensure the integrity, accuracy, and consistency of the data. Through this stage of work, a high-quality and standardized dataset will be formed, laying a solid data foundation for subsequent research and ensuring the reliability and effectiveness of subsequent data analysis results.

[0072] (2) Feature Importance Assessment Using Statistical Analysis and Machine Learning Methods: After data cleaning, the second stage focuses on accurately identifying key features affecting the output of trailing suction hopper dredgers from massive amounts of data. Combining professional knowledge and practical experience in dredging operations, feature selection and extraction techniques from data mining and machine learning are used to conduct in-depth analysis of the data. The correlation between equipment parameters (such as pump power, speed, and drag head depth), environmental factors (water flow velocity, soil type, and tidal changes) and output under different operating conditions is analyzed to screen out key feature variables that have a significant impact on output. Through key feature identification, the data dimensions are simplified, key influencing factors are highlighted, and a clear direction is provided for establishing an accurate output prediction model, making the model more targeted and practical.

[0073] 2.2.2 Key Technologies

[0074] 2.2.2.1 Data Cleaning and Preprocessing Techniques: Statistical methods and data processing algorithms are used to identify and correct complex data from multiple sources, including construction logs, equipment sensor monitoring data, and environmental monitoring data. Invalid data is removed, missing values ​​are filled in, and erroneous records are corrected to ensure the integrity, accuracy, and consistency of the data.

[0075] 2.2.2.2 Feature selection and extraction techniques in machine learning: accurately identify key features affecting the output of trailing suction hopper dredgers from massive amounts of data, simplify data dimensions, highlight key influencing factors, and clarify the direction for establishing an accurate output prediction model.

[0076] 2.3 Technical Solution

[0077] The data in this invention comes from relevant parameters collected from various ships in different locations during various projects over the years, and historical data is stored in a database.

[0078] The database management software in the virtual machine uses Microsoft SQL Server 2014. Due to the large volume of ship data, a separate database instance is created for each ship and each year's data, with data stored separately. To speed up data retrieval, monthly data is stored in separate tables. Each table saves one record every 2 seconds. Historical data from trailing suction hopper dredgers under different operating periods and construction projects is comprehensively collected, covering construction logs, equipment sensor data (such as pump speed, rake head pressure, and hopper level sensor data), and environmental monitoring data (weather station wind speed and direction data, hydrological station tidal level data, and geological exploration soil reports). Data scattered across different systems and formats (such as Excel spreadsheets, database files, and text logs) is integrated, unified in storage format, and a raw dataset is constructed. First, relatively singular soil type (medium-coarse sand) related construction data is extracted for a series of analyses and processing. The basic information of this data is as follows: Figure 2 As shown. It should be noted that: raw data includes all data that can be collected from all ship systems, such as... Figure 3 As shown, due to sensor malfunctions and other reasons, some of the data will not change or will always be 0 (i.e., invalid data, data that does not contain information), and needs to be processed and deleted. The original data name is as follows: Figure 3 As shown. Statistical methods are used to conduct a preliminary analysis of the raw data, calculating basic statistics such as mean, standard deviation, maximum, and minimum values ​​to detect the data's distribution characteristics. By setting reasonable thresholds, outliers are identified; the missing percentage of each field is calculated, and the pattern of missing data (random missing, completely random missing, or non-random missing) is analyzed; duplicate records are identified using data deduplication tools, and a comprehensive data quality assessment report is generated. After performing a statistical operation on the raw data, the maximum and minimum values, the intervals with the highest frequency of each value (50 intervals divided by the maximum and minimum value range), and variance are obtained for each data point; a histogram of key feature data frequencies is shown below. Figures 4a-4b , Figures 5a-5b , Figures 6a-6b , Figures 7a-7b , Figures 8a-8k As shown, where, Figure 4a , Figure 4b Corresponding to GPS No. 1 North latitude and GPS No. 1 East longitude, Figure 5a , Figure 5b Corresponding to GPS1 ground speed and GPS1 ground heading, Figure 6a , Figure 6b Corresponding dredging time and tide level. Figure 7a , Figure 7b Corresponding to tilt (m) and yaw (m), Figure 8a , Figure 8b , Figure 8c , Figure 8d , Figure 8e , Figure 8f , Figure 8g , Figure 8h , Figure 8i , Figure 8j , Figure 8k This corresponds to the following: right concentration, right rake head depth, upper right vertical angle, right compensator pressure, lower right vertical angle, right high-pressure flushing pressure, right active rake head angle, right flow velocity, right active rake head stroke, right density, and right active rake head pressure. After removing data types with a variance of 0 (i.e., unchanged values), approximately 120 items remain—these are the valid data types, such as... Figure 9 As shown. Valid data excludes data collected during dredging operations (empty, fully loaded, unloading, etc.), retaining only data collected when the dredging vessel is dredging; more precisely, data collected when the rake head is lowered to the mud surface and the pump is operating. Data was filtered using three indicators: density, pump speed, and rake head depth. Because the hydrometers require calibration, and there are errors when the left and right hydrometers draw in clean water, the density of the right hydrometer is greater than 1.03 t / m³. 3 The density on the left is 1.05 t / m³. 3 The depth of the dredger head varies depending on the project; in this project, data collected when the depth was greater than 7m was used. The pump speed was around 180 r / min, a constant speed pump, which is relatively stable. The screening and labeling process for dredging data from trailing suction hopper dredgers is as follows: Figure 10 As shown. After extracting the effective data, considering the construction characteristics of trailing suction hopper dredger projects, the screened construction data was integrated and divided into 190 dredging trips (each trip represents the stage from dredging to full-load navigation). The data naming format for each trip is as follows: 1 trip_20160306_231710-20160307_012802 (which includes the start and end times of construction for this trip).

[0079] 2.3.1 Data Preprocessing: To address the issues identified in the data quality assessment, corresponding data cleaning strategies were implemented. For duplicate data, unique and valid records were retained based on the principles of data integrity and accuracy. For missing data, methods such as mean imputation, multiple imputation, and model-based prediction imputation were used to fill missing values ​​according to the missing data pattern. For erroneous data, corrections were made through manual verification or algorithmic error correction, considering business logic and historical data patterns. Due to the complexity of dredging operations and the wide range of data sources, including construction logs, equipment sensor monitoring data, and environmental monitoring data, a large amount of duplicate, missing, and erroneous information exists. This stage of work resulted in a high-quality, standardized dataset, laying a solid data foundation for subsequent research and ensuring the reliability and effectiveness of subsequent data analysis results. Specifically, this included the following aspects:

[0080] 2.3.1.1 Outlier Detection and Handling: During the data acquisition process of dredging projects, outliers often exist in the data due to factors such as sensor errors and environmental interference. If these outliers are not handled, they will seriously affect the subsequent analysis of indicators such as dredging efficiency. Therefore, this invention adopts a method based on statistical rules combined with manual screening to accurately handle outliers. For physically unreasonable values, such as negative flow velocities, zero or minimum values, such data clearly violate physical laws. In actual dredging operations, water flow velocity must be positive, and the density of the medium cannot infinitely approach zero. This invention uses box plots and the 3σ method for preliminary screening of the data. Box plots can intuitively identify outliers in the data by displaying the quartiles of the data; the 3σ method is based on the principle of normal distribution, and when a data point exceeds the mean ± 3 times the standard deviation, it can be identified as an outlier. After preliminary screening using statistical rules, combined with the actual working conditions and physical laws of dredging operations, professional personnel conduct manual verification to ensure that physically unreasonable values ​​are accurately identified and corrected or removed. Some parameters may exhibit extreme jumps due to sensor jitter, causing discontinuities in the data sequence and affecting the accuracy of data analysis. To address this issue, this invention employs a sliding window or median filtering method for smoothing. The sliding window method uses a fixed-size window that slides sequentially across the data sequence, replacing the original data within the window with a statistical measure (such as the mean or median), thereby reducing data fluctuations. Median filtering, on the other hand, sorts the data within the window and replaces the original data with the median value. This method is highly effective in removing impulse noise and can effectively eliminate the effects of sensor jitter, making the data sequence smoother and more stable.

[0081] 2.3.1.2 Missing Value Imputation: Due to factors such as communication packet loss and sensor failure, dredging project data contains missing values ​​at certain time points. Without processing, this leads to incomplete data and affects the reliability of the analysis results. This invention develops targeted imputation strategies based on parameter characteristics. For continuous variables, such as flow velocity and density, they exhibit strong continuity and trends in time series. This invention uses forward imputation or interpolation methods for missing value imputation. Forward imputation fills missing values ​​with the preceding valid data; this method is simple and efficient, suitable for situations where data changes are relatively stable. Interpolation methods estimate missing values ​​based on surrounding valid data using mathematical models (such as linear interpolation, polynomial interpolation, etc.), better reflecting data trends and suitable for scenarios with large data fluctuations. Discrete variables or state-type parameters, such as the operating status of equipment (normal, fault), do not possess the numerical characteristics of continuous variables. For missing values ​​in this type of data, this invention uses the mode or the most recent valid value for imputation. Mode imputation fills missing values ​​with the most frequent value in the dataset, which is suitable for situations where the data distribution is relatively concentrated; nearest valid value imputation fills missing values ​​with the nearest valid data, which can preserve the actual state information of the data to the greatest extent and ensure the integrity and availability of the data.

[0082] 2.3.1.3 Right-side Parameter Filtering: This stage of the research focuses on the dredging efficiency of the "right rake head." Therefore, it is necessary to split the fields related to "left" and "right" in the dataset, retaining only the right-side fields related to the "right rake head." The original dataset contains various parameters of the left and right rake heads, such as flow velocity, density, rake head angle, and flushing pressure. This invention uses data cleaning and filtering procedures to identify and classify these fields one by one, extracting right-side related fields such as right flow velocity, right density, right rake head angle, and right flushing pressure to form a new dataset. This makes the data more focused, reduces interference from irrelevant information, and facilitates subsequent in-depth analysis and research on the dredging efficiency of the "right rake head," providing more accurate data support for optimizing dredging operations.

[0083] 2.3.1.4 Feature Standardization: In some modeling algorithms such as gradient-based neural networks, the features of the original data often have different dimensions and large differences in numerical distribution. Taking dredging engineering data as an example, the numerical range of the right flow velocity may be 0-5 m / s, while the numerical value of the right flushing pressure may be 0-500 kPa. This imbalance of data features will lead to unstable gradient updates, slow convergence speed, and even getting stuck in local optima during model training. In order to improve the stability and training efficiency of the model, it is particularly important to normalize the input features. Min-Max standardization is used to linearly normalize the data, which is calculated according to equation (1):

[0084] (1)

[0085] In the formula, This is the normalized data, where x is the original data. It is the minimum value of this parameter. This is the maximum value of the parameter. This method linearly maps the data to the interval [0,1], which can preserve the original distribution of the data. It is suitable for models that are sensitive to the data distribution and do not have extreme outliers, such as some simple neural networks or traditional machine learning models. Z-score standardization is based on the mean and standard deviation of the data and is transformed by formula (2):

[0086] (2)

[0087] in It is the mean of the data. This refers to the standard deviation of the data. After Z-score standardization, the mean of the data becomes 0, and the standard deviation becomes 1. This method eliminates the influence of data dimensions and has good robustness to data with outliers, and is often used for training complex neural network models. In practical applications, this invention will compare the impact of the two methods on model performance through experiments to select the most suitable normalization method, ensuring that the model can learn data features more efficiently during training.

[0088] 2.3.1.5 Data Lag Correction: In actual dredging operations, due to the physical response mechanisms and signal acquisition and transmission characteristics of different equipment, there are significant time delays between the data generated by different devices. For example, when adjusting the angle of the right rake head and the right flushing pressure, the water flow and medium density do not change immediately; the actual impact usually occurs several seconds later. Similarly, parameters reflecting the ship's motion state, such as GPS speed, also have a certain lag in their impact on dredging output. If these time delays are ignored during modeling, the model will incorrectly correlate data without causal relationships, learning chaotic characteristic response relationships, thus severely affecting prediction accuracy. Therefore, this stage employs a systematic approach to identify and correct data time delays. Without considering these time delays, the model may learn chaotic characteristic response relationships, thereby affecting prediction accuracy. Therefore, this stage uses the following methods to identify and correct data time delays:

[0089] (1) Time Delay Analysis Method

[0090] Maximum Cross-Correlation Coefficient Method: This method identifies the maximum correlation lag time between input features and the target variable, a key indicator for measuring dredging efficiency, by calculating the cross-correlation coefficient between them. Specifically, the correlation between the input feature sequence and the target variable sequence is calculated at different time offsets, and the cross-correlation coefficient is plotted as a function of time lag. The lag time corresponding to the peak value on the curve represents the time delay at which the correlation between the input feature and the target variable is strongest. For example, when analyzing the relationship between right flushing pressure and dredging efficiency, the maximum cross-correlation coefficient method can accurately determine the lag time at which changes in right flushing pressure have the greatest impact on dredging efficiency.

[0091] Sliding window traversal + R 2 Evaluation Method: This method constructs a regression model for each lag step by shifting the input features backward by different lag steps, and then uses R... 2 The performance of the model is evaluated using metrics. The specific process involves starting with the minimum lag step size and gradually increasing the number of lag steps. After each shift, the regression model is trained based on the adjusted data, and the R-squared of the model is calculated. 2 Value. By comparing the R-values ​​of the model under different lag steps. 2 Value, select to make R 2 The lag step size that reaches its maximum value is taken as the optimal lag time for this feature. This method can directly find the time lag parameter that most significantly improves prediction accuracy from the perspective of model prediction performance;

[0092] The empirical method: Leveraging extensive experience in dredging projects, engineers can initially define several reasonable lag intervals. For example, based on past construction experience, the lag time for the effect of changes in the right rake head angle on flow velocity is estimated to be between 3 and 8 seconds. Experimental analyses are conducted within this range, and the model training effect is observed by adjusting the lag time. A comprehensive judgment is then made based on actual engineering phenomena to determine a more reasonable lag time. Although this method relies on experience, it can quickly provide preliminary time delay estimates, offering a reference for other analytical methods.

[0093] (2) Lag Adjustment Strategy: After determining the lag time of each feature, features with significant time lag (such as the right rake angle and right flushing pressure) are shifted backward according to the corresponding time step. For example, if analysis shows that the right rake angle has the strongest correlation with the target variable when lagging by 5 seconds, the entire right rake angle data sequence is shifted backward by 5 seconds, and the target variable is realigned to ensure that the causal relationship between the features and the target variable remains consistent in the time dimension. In addition, to provide the model with more choices, after time lag processing, this invention retains both the original and delayed versions of the features. In this way, the model can autonomously learn and select more predictive feature representations during training, and mine potential patterns in the data.

[0094] (3) Examples of Lag Analysis Results: Through the time lag analysis method described above, results with practical guiding significance were obtained in some datasets. For example, detailed analysis revealed that the correlation between the right active harrow angle and dredging output reached its maximum when the angle lagged by 30 seconds. This indicates that after adjusting the right active harrow angle, it takes approximately 30 seconds to show a significant change in output. By correcting these time lags, the model can learn causal paths that are more consistent with physical laws, avoiding erroneous feature associations and effectively improving the stability and generalization ability of predictions. In practical applications, the model trained based on the corrected data demonstrates higher accuracy and reliability in predicting dredging output and optimizing construction parameters.

[0095] 2.3.2 Identification and Coupling Analysis of Key Performance Features

[0096] 2.3.2.1 Feature Selection and Construction: The dredging operation system of a trailing suction hopper dredger collects monitoring data with up to 160 dimensions. However, based on relevant construction experience, the dredging operation subsystems related to the dredging output only include: the dredging system, the mud pump pipeline system, and the ship's power system. Furthermore, if all feature parameters involved in the dredging operation system were input as control variables into the output prediction model, it would significantly increase the model's spatial complexity and computation time. Therefore, while considering the accuracy of the output model and the modeling cost, this invention utilizes a weighted multiple hybrid strategy based on machine learning algorithms and combines it with human experience to select factors influencing the dredging output. In actual trailing suction hopper dredging operations, dredging output is affected by multiple factors, such as... Figure 11As shown, factors include mud pump power, high-pressure flushing, wave compensator pressure, dredging head depth, and vessel attitude. However, the specific relationships between the various characteristic variables in the production model and dredging efficiency are still unclear, exhibiting problems such as information redundancy and complex variable interactions, making it difficult to intuitively determine which factors contribute most to production. Therefore, it is necessary to clarify the mechanisms of action of each influencing factor and identify key variables from a data-driven perspective through feature importance analysis. This step helps optimize the model structure, improve prediction accuracy, and provide a scientific basis for subsequent operational parameter adjustments. Based on the experience of experts and construction personnel, and combined with the actual needs of the project, a dataset is constructed by selecting some parameters. These parameters may include, but are not limited to: time, GPS1 ground speed, upper left vertical angle, lower left vertical angle, upper right vertical angle, lower right vertical angle, left active scraper head pressure, left active scraper head angle, right active scraper head pressure, right active scraper head angle, left suction vacuum, right suction vacuum, right high-pressure flushing pressure, left high-pressure flushing pressure, left mud pump speed, right mud pump speed, left compensator stroke, right compensator stroke, left compensator pressure, right compensator pressure, heel (m), trim (m), heel angle, trim angle, left density, left flow velocity, right density, right flow velocity, left flow velocity, right flow velocity, etc. Among these, parameters that have a direct or indirect impact on the dredger's output should be selected to avoid the machine learning algorithm being unable to correctly identify features due to changes in irrelevant parameters. Specifically, taking the establishment of a right scraper-related dataset as an example, the right concentration and right flow rate are first selected, and their product, i.e., the instantaneous output of the right scraper, is calculated as the target parameter. Secondly, parameters related to the right harrow operation were selected, such as: time, GPS1 ground speed, right active harrow head angle, right active harrow head pressure, right active harrow head stroke, right high-pressure water jet pressure, right compensator pressure, right upper vertical angle, right lower vertical angle, right upper horizontal angle, right lower horizontal angle, right harrow head depth, lateral tilt (m), longitudinal tilt (m), right discharge pressure, and right suction vacuum. All data for the selected parameters were linearly normalized, and each data point was mapped and transformed according to the maximum and minimum values ​​of its column, allowing for comparison and calculation on the same scale. When the absolute value of the correlation coefficient was greater than 0.8, features with stronger explanatory power for yield were retained by using feature importance scores (weighted calculation using four algorithms: Lasso, LightBGM, Random Forest, and CatBoost).

[0097] 2.3.2.2 Feature Importance Analysis: Based on professional knowledge of trailing suction hopper dredger operations and combined with industry expert experience, potential feature variables affecting operational output were initially screened from the original dataset. These variables included equipment operating parameters, environmental factors, and construction management factors, thus constructing an initial feature set. After initially screening a large number of features, to further refine the feature set and improve model efficiency and accuracy, this invention employs a weighted multi-hybrid strategy using Lasso, LightBGM, Random Forest, CatBoost algorithms, and Pearson correlation coefficient heatmaps for feature selection.

[0098] (1) Feature selection method based on random forest: Random forest is an ensemble learning model composed of multiple decision trees. It improves the model's stability and generalization ability by combining the prediction results of multiple decision trees. Data is input into the random forest model, and the importance of each feature is calculated based on its contribution to sample classification or regression within the random forest. Specifically, by randomly shuffling the values ​​of a feature, the decrease in the model's prediction accuracy is observed. The greater the decrease, the more important the feature is to the model. For example, after the water flow velocity feature is shuffled, the model's prediction accuracy for operational output decreases significantly, indicating its high importance. Figure 12 As shown, the top fifteen key features were derived based on dredging construction data. Specifically, the random forest employed three feature importance analysis methods: feature importance based on the Gini index, feature importance estimated from out-of-bag data, and feature importance based on the Shapley value. The importance results are shown in Tables 1, 2, and 3.

[0099] Table 1. Results of Feature Importance Analysis Based on Gini Index for Random Forest

[0100]

[0101] Table 2. Results of Feature Importance Analysis Based on Out-of-Bag Data Estimation for Random Forest

[0102]

[0103] Table 3. Results of Feature Importance Analysis Based on Shapley Value for Random Forest

[0104]

[0105] (2) Feature Selection Based on LightGBM: LightGBM is an efficient gradient boosting framework that employs histogram algorithms and one-sided gradient sampling techniques. It offers advantages such as fast training speed and low memory consumption when processing large-scale data. The initial feature set and job output data are input into the LightGBM model. By calculating the gain of features during the splitting process of decision tree nodes, the importance score of each feature is obtained. For example, in a training process, if a certain feature significantly reduces the model's prediction error when splitting at multiple nodes, its importance score is high. Figure 13 As shown, with production volume as the target quantity, the construction data of Xiamen Port from March to June 2016 was input into the algorithm to obtain the histogram of the top fifteen. LightGBM adopted three feature importance analysis methods: feature importance based on split frequency, feature importance based on split gain, and feature importance based on Shapley value. After linear normalization of the analysis results, the feature importance is shown in Tables 4, 5, and 6.

[0106] Table 4. Eigenimity of LightGBM Based on Splitting Frequency

[0107]

[0108] Table 5. Eigenimity of LightGBM Based on Split Gain

[0109]

[0110] Table 6. Feature Importance of LightGBM Based on Shapley Values

[0111]

[0112] (3) Feature selection based on Lasso: Lasso is a feature selection method based on regression analysis. It adds an L1 regularization term to the loss function of ordinary linear regression. L1 regularization imposes constraints on the regression coefficients, causing some regression coefficients to become 0, thereby achieving feature selection. During model training, the Lasso algorithm automatically filters features by minimizing the loss function after adding the L1 regularization term, retaining features that have a significant impact on the target variable (operation output) and removing irrelevant or less influential features. Taking the data of trailing suction hopper dredgers as an example, if the regression coefficient corresponding to a certain environmental factor feature approaches 0 during the optimization process of the Lasso algorithm, it indicates that the feature has a small contribution to the prediction of operation output and can be removed from the feature set. The application of Lasso first uses the initial feature set and operation output data as input to build a linear regression model based on Lasso. By adjusting the hyperparameters of Lasso (such as the regularization strength (lambda)), the filtering intensity of features is controlled. After multiple experiments, the (lambda) value that minimizes the model prediction error is selected, and based on whether the regression coefficients of each feature are 0 at this time, a preliminary important feature subset is selected. Figure 14 As shown, the feature coefficient convergence graph obtained after initial screening by experts and input into the Lasso algorithm; using the Lasso algorithm, the top twelve importance features were selected and a Pareto chart was plotted, as shown. Figure 15 As shown.

[0113] (4) Feature Selection Based on CatBoost: CatBoost, as an advanced machine learning algorithm based on gradient boosting, exhibits unique advantages in handling complex data features. Its innovation lies in its handling of categorical features. Traditional algorithms often require preprocessing operations such as one-hot encoding when handling categorical features, which not only increases the data dimensionality but may also introduce spurious correlations. CatBoost, however, uses a special ordered statistical transformation method that can automatically encode categorical features without additional complex preprocessing, greatly improving computational efficiency and model performance. In terms of preventing overfitting, CatBoost employs an ordered boosting strategy. It introduces a randomly arranged order of training samples each time a decision tree is built. This randomness breaks the potential patterns in the data, preventing the model from over-relying on a certain part of the data during training, thereby effectively reducing the bias of gradient estimation and lowering the risk of overfitting. In terms of feature selection mechanism, CatBoost evaluates feature importance based on the optimization of the objective function during the decision tree construction process. In regression problems, the objective function is usually the mean squared error, while in classification problems it is the log loss function. When a feature splits at a decision tree node, it significantly reduces the value of the objective function, indicating that the feature makes a significant contribution to the model's prediction of the target variable (dredging output of a trailing suction hopper dredger), and its importance score will increase accordingly. Before model training, a series of key parameters need to be finely adjusted. For example, the learning rate determines the step size of each parameter update; a smaller learning rate makes the model training more stable, but increases training time. The depth of the tree affects the complexity of the model; too deep a tree can lead to overfitting, while too shallow a tree results in insufficient expressive power. Through multiple experiments, this invention has determined a suitable parameter combination, such as learning rate = 0.05 and depth = 6. The final top fifteen features are as follows: Figure 16 As shown.

[0114] 2.3.2.3 Characteristic Coupling Analysis Based on Pearson Correlation Coefficient: The Pearson correlation coefficient, also known as the Pearson product-moment correlation coefficient, is a statistic used to measure the degree of linear correlation between two variables. Its theoretical basis stems from the combination of covariance and standard deviation. For two variables... X and Y The formula for calculating the Pearson correlation coefficient is:

[0115] ;

[0116] in, The covariance of variables X and Y measures the degree of co-variance between the two variables during a change process. and , respectively, are the standard deviations of variables X and Y, used to characterize the dispersion of the variables; n is the sample size. and These are the i-th observations of variables X and Y, respectively. and This is the sample mean of variables X and Y. The numerator of the formula... This reflects the sum of the products of the deviations of two variables from their respective means. X and Y When both values ​​are greater than or less than their respective means, the product is positive, indicating that they have the same trend; conversely, the product is negative, indicating that they have opposite trends. The denominator is standardized to convert the covariance into a correlation coefficient ranging from -1 to 1.

[0117] In the data feature analysis scenario of trailing suction hopper dredgers, if we consider pump power as variable X and operating output as variable Y, we can quantify the degree of linear correlation between changes in pump power and changes in operating output by calculating their Pearson correlation coefficient. When X and Y are in a perfectly positive linear relationship, it means that as the power of the mud pump increases, the output will increase strictly according to a linear relationship; when... When, it indicates the existence of a perfectly negative linear correlation; when When the Pearson correlation coefficient is zero, it indicates that there is no linear correlation between the two variables, but it does not rule out the possibility of a nonlinear relationship. Furthermore, the Pearson correlation coefficient exhibits symmetry, meaning... And its square value The coefficient of determination (COP) reflects the proportion of variation in one variable that can be explained by another. However, the Pearson correlation coefficient relies on the mean and standard deviation of the variables, requiring the data to follow a normal or approximately normal distribution. It only measures linear correlations and may lead to erroneous conclusions for non-linearly correlated variables. Therefore, it should be used cautiously in practice, taking into account the data distribution and business context. Figure 17 As shown, the Pearson correlation coefficient heatmap presents the Pearson correlation coefficients between various features in a visual matrix form. The color intensity represents the magnitude of the correlation coefficient, making it easy to intuitively observe the correlation between features. In feature selection for trailing suction hopper dredger data, if the absolute value of the Pearson correlation coefficient between two features is greater than a set threshold (e.g., 0.8), it indicates that these two features are highly redundant, and one can be retained while the other is removed. Specifically, features with a more significant impact on instantaneous output are retained based on feature importance scores (calculated by weighted averages from four algorithms: Random Forest, LightGBM, Lasso, and CatBoost): if the absolute value of the correlation coefficient between two features is > 0.8, the feature with the higher feature importance score is selected for retention.

[0118] The causal relationship between the time-delay corrected data and the output was verified (maximum cross-correlation coefficient ≥ 0.78), and the model complexity was tested after redundant features were removed (AIC value reduced by 15%) to ensure that the input data met the model requirements.

[0119] 2.3.2.4 Conclusions on Feature Selection Based on Hybrid Weights of Multiple Machine Learning Algorithms

[0120] Based on the results of the aforementioned machine learning algorithms, a parameter importance quantification model based on multi-algorithm fusion is constructed. The top ten most important parameters are selected, and values ​​are assigned to the identified parameters. Weighted summation is used to fuse the results of multiple algorithms, forming an integrated feature recognition conclusion. In the selection of key features for trailing suction hopper dredgers, this invention integrates expert experience with random forest, LightGBM, Lasso, and Catboost algorithms to systematically screen numerous feature variables in the original data. Experts, based on their rich knowledge and practical experience in dredger operations, predict features that may affect operational output from an engineering perspective. The Lasso algorithm, based on regression analysis, uses L1 regularization to constrain regression coefficients, achieving sparsity screening of features and effectively eliminating features with weak impact on the target variable. The random forest algorithm, by integrating multiple decision trees, assesses feature importance based on the degree of impurity reduction during node splitting, selecting features that play a crucial role in model prediction. The Catboost algorithm, with its unique handling of categorical features and its advantage in preventing overfitting, accurately evaluates the contribution of features to the optimization of the objective function, further optimizing the feature set.

[0121] Through the above multi-method collaborative screening, preliminary characteristics were identified, including GPS1 ground speed, right upper vertical angle, right lower vertical angle, roll, right rake head depth, right active rake head stroke, right active rake head angle, right suction vacuum, right high-pressure flushing pressure, right high-pressure flushing current / speed, right mud pump speed, right main engine speed, right compensator stroke, and right compensator pressure. These characteristics cover multiple dimensions such as equipment operating parameters and operational attitude parameters, laying the foundation for subsequent in-depth analysis. The feature selection based on the mixed weight of multiple methods is shown in Table 7.

[0122] Table 7. Ranking of Feature Selection Methods

[0123]

[0124] Based on this, this invention uses a Pearson correlation coefficient heatmap to perform redundancy analysis on the initially selected features. The heatmap clearly shows that the absolute value of the correlation coefficient between the right compensator stroke and the right compensator pressure is as high as 0.94, exhibiting a very strong linear correlation. Further analysis of their physical relationship reveals that the right compensator pressure and stroke are closely related during dredging operations; both essentially reflect the working characteristics of the compensator when adjusting the contact state between the rake head and the riverbed. However, from a practical application perspective, the right compensator pressure more directly reflects the working intensity of the compensator and has a more representative impact on operational output. Therefore, the right compensator pressure is retained as a feature selection, while the right compensator stroke is discarded. Similarly, the right mud pump speed and the right main engine speed have a high correlation. As the core of power output, the right main engine's speed change is transmitted to the right mud pump, but the right mud pump speed directly determines the mud transport efficiency and is more closely related to the actual production state of dredging operations. To ensure the feature's explanatory power for operational output and to avoid increased model complexity due to feature redundancy, this invention prioritizes the right mud pump speed and discards the right main engine speed. Specifically, each parameter ranked first in importance was assigned 10 points, and the remaining parameters were assigned 1 point in order of importance, until the parameter ranked 10th was assigned 1 point. The results are shown in Table 8.

[0125] Table 8. Results of Integrated Feature Recognition

[0126]

[0127] It should be noted that Table 8 is essentially a weighted fusion result of the top 10 features of four algorithms (Random Forest, LightGBM, Lasso, and CatBoost) (refer to Table 7, "Ranking of Feature Selection by Multiple Methods"). Its selection logic is as follows: only features that "enter the top 10 in at least one algorithm" are weighted. Features that do not enter the top 10 in any algorithm, such as right high-pressure flushing current / speed and tilt, are directly excluded from the statistical scope of Table 8. By comparing with Table 7 (top 10 features of the four algorithms):

[0128] Right high-pressure flushing current / speed: This feature did not appear in the top 10 of the four algorithms: Random Forest, LightGBM, Lasso, and CatBoost. This indicates that the four algorithms consistently judged that its "importance to output was not among the top 10", and therefore it was not included in the weighted calculation in Table 8.

[0129] Lateral tilt: ranked 10th by Lasso – its impact on yield is weaker than other features. Although longitudinal tilt scored low, LightGBM ranked longitudinal tilt 10th and longitudinal tilt angle 10th by Random Forest (the two are related, indicating that longitudinal tilt and longitudinal tilt angle are more important than lateral tilt overall). Therefore, lateral tilt was not included in the statistics in Table 8.

[0130] The highest score of 55 points for "right suction vacuum" indicates its significant impact on model decision-making during dredging operations, suggesting it may be a core parameter for controlling dredging efficiency and mud transport stability. "Right active scraper head angle" (47 points) and "right upper and lower right vertical angles" (43 and 40 points respectively) ranked second, reflecting the crucial role of scraper head attitude parameters in excavation accuracy and soil disturbance. Furthermore, "right active scraper head pressure" and "right high-pressure flushing pressure" both scored 32 points, indicating that these two parameters are equally important in controlling scraper head breaking capacity and mud suspension effects. The scores for "right compensator pressure," "GPS1 ground speed" (23 points), and "right compensator stroke" (22 points) are close, indicating their significant impact on hull stability and construction positioning accuracy. In contrast, "trim (m)" (8 points) and "trim angle" (3 points) scored lower, suggesting their overall impact on dredging operation parameters in the current model is relatively limited.

[0131] In some embodiments, based on the importance score of the key features and combined with parameter sensitivity analysis of 190 vessel trips (optimal range of right suction vacuum: -0.08 to -0.05 MPa, optimal range of right active rake head angle: -15° to -10°), construction parameter optimization suggestions are generated: if the right suction vacuum is < -0.08 MPa, it is recommended to increase the right active rake head angle by 3 to 5°; if the right high-pressure flushing pressure is < 0.3 MPa, it is recommended to increase the pressure to 0.3 to 0.4 MPa to adjust the operating parameters.

[0132] In some embodiments, based on construction data statistics from 190 ship trips, the instantaneous output changes by an average of 2.3% for every 1° adjustment of the right active rake head angle, thereby establishing a quantitative correlation between parameter adjustment and output increase.

[0133] In some embodiments, the parameter optimization suggestions include adjusting the operating parameters corresponding to key features: the right active rake head angle (each adjustment of 1° results in an average change of 2.3% in instantaneous output), the right suction vacuum (each increase of 0.01 MPa results in an increase of 1.8% in output), and the upper right vertical angle (optimal range: 25° to 30°), in order to improve instantaneous output.

[0134] In some embodiments, the input data of the production prediction model needs to undergo preprocessing in step (2) (outlier handling uses the 3σ method + box plot method, and time lag correction uses the maximum cross-correlation coefficient method) and coupled analysis in step (5). The causal relationship between the feature and the production is ensured by calculating the maximum cross-correlation coefficient between the feature and the production to be ≥0.75, and there is no redundancy by eliminating redundant features (retaining features with higher scores).

[0135] In some embodiments, the production prediction model construction steps are as follows: (1) Data preparation: Select key feature sets (features such as right suction vacuum, right active rake angle, etc.) and corresponding instantaneous production labels, with a total of 400,000 valid data (from 190 ship voyages); (2) Model training: Use the improved LightGBM, divide the training set / test set in an 8:2 ratio, and optimize the hyperparameters with 5-fold cross-validation (final hyperparameters: learning_rate=0.05, depth=6, num_leaves=31); (3) Model validation: The test set MSE=0.23, R²=0.87, which is better than the traditional random forest (R²=0.79) and the basic LightGBM (R²=0.82).

[0136] In some embodiments, the process for generating construction parameter optimization suggestions is as follows: (1) Calculation of sensitivity of key features: Based on the Shapley value of the model, calculate the marginal contribution of each feature to the output (the Shapley value of the right suction vacuum is 0.32, which is the highest); (2) Determination of the optimal range: Statistically analyze the parameter range corresponding to "output ≥ 8 m³ / min" in 190 ship trips; (3) Generation of suggestions: If the real-time parameters exceed the optimal range, output the adjustment direction and magnitude (e.g., the right active rake angle is < -15°, it is recommended to increase it by 2° to 3°).

[0137] In some embodiments, the production prediction model is deployed on the dredger's local server (CPU: Intel Xeon E5-2680, memory 32GB), and the model is automatically retrained every 50 additional dredger trips to ensure stable prediction accuracy.

[0138] The machine learning-based harrow digging production data decoupling analysis system, applying any one of the machine learning-based harrow digging production data decoupling analysis methods described above, comprises:

[0139] The data acquisition module is used to acquire multi-source historical data, including equipment sensor data, environmental monitoring data, and construction log data.

[0140] The preprocessing module is used to preprocess the multi-source historical data, including: a. performing outlier processing to obtain preliminary cleaned data; b. filling missing values ​​in the preliminary cleaned data to obtain filled data; c. filtering the right-hand parameters in the filled data, retaining only fields related to the right rake head to obtain right-hand parameter data; d. performing feature standardization on the right-hand parameter data to obtain standardized data; e. performing time lag correction on the standardized data to adjust the data alignment causal relationship to obtain time lag corrected data.

[0141] The construction and calculation module is used to construct an initial feature set from the time-delay correction data and calculate the instantaneous output;

[0142] The evaluation module is used to evaluate the feature importance of the initial feature set with instantaneous output as the objective, using four machine learning algorithms: Random Forest, LightGBM, Lasso, and CatBoost, to obtain the importance score of each feature.

[0143] A redundancy processing module is used to perform coupling analysis on the initial feature set based on the importance scores of each feature, use the Pearson correlation coefficient to detect the linear correlation between features, identify and remove redundant features, and obtain a set of key features.

[0144] The prediction module is used to construct a production prediction model based on the set of key features, taking the key features as input and outputting instantaneous production prediction results; the production prediction model adopts the gradient boosting regression algorithm.

[0145] The optimization module is used to generate construction parameter optimization suggestions based on the importance scores of the key features, which are used to adjust operational parameters such as the right rake head angle and the right suction vacuum.

Claims

1. A machine learning-based method for decoupling and analyzing rake digging production data, characterized in that, include: Step (1) Obtain multi-source historical data, which includes equipment sensor data, environmental monitoring data, and construction log data; Step (2) preprocesses the multi-source historical data, including: a. Perform outlier processing to obtain preliminary cleaned data; b. Fill in the missing values ​​in the preliminary cleaned data to obtain the filled data; c. Filter the right-hand parameters in the filled data, keeping only the fields related to the right rake head, to obtain the right-hand parameter data; d. Perform feature standardization on the parameter data on the right to obtain standardized data; e. Perform time-delay correction on the standardized data, adjust the data alignment causal relationship, and obtain time-delay corrected data; Step (3) Construct an initial feature set from the time-delay correction data and calculate the instantaneous output; Step (4) With instantaneous output as the objective, four machine learning algorithms, namely Random Forest, LightGBM, Lasso, and CatBoost, are used to evaluate the feature importance of the initial feature set and obtain the importance score of each feature. Step (5) Based on the importance scores of each feature, perform coupling analysis on the initial feature set, use Pearson correlation coefficient to detect the linear correlation between features, identify and remove redundant features, and obtain the key feature set; Step (6) Construct a production prediction model based on the set of key features, using the key features as input, and output instantaneous production prediction results; the production prediction model adopts the gradient boosting regression algorithm.

2. The method for decoupling and analyzing rakehead digging production data based on machine learning according to claim 1, characterized in that, Also includes: Step (7) Based on the importance scores of the key features, generate construction parameter optimization suggestions for adjusting the operation parameters.

3. The method for decoupling and analyzing rakehead digging production data based on machine learning according to claim 1, characterized in that, In step (4), the random forest algorithm evaluates feature importance based on the Gini index, out-of-bag data, and Shapley value; the LightGBM algorithm evaluates feature importance based on split frequency, gain, and Shapley value; the Lasso algorithm filters features through L1 regularization; and the CatBoost algorithm evaluates feature importance based on the degree of optimization of the objective function by the decision tree.

4. The method for decoupling and analyzing rakehead digging production data based on machine learning according to claim 1, characterized in that, In step (5), the threshold of the Pearson correlation coefficient is set to an absolute value greater than 0.8, that is, the two features are highly correlated, and the more critical features are retained based on the significance of the feature's impact on the yield.

5. The method for decoupling and analyzing rake digging production data based on machine learning according to claim 1, characterized in that, In step (6), the output prediction model learns the nonlinear relationship between key features and output through training data. Its construction process includes: (a) dividing the training set and the test set: taking the key feature set as the input variable and the instantaneous output as the target variable, the dataset is randomly divided in a ratio of 80%:20%, and 5-fold cross-validation is used to avoid overfitting; (b) setting the model hyperparameters: learning rate 0.05, decision tree depth 6, number of iterations 500, L2 regularization coefficient 0.01, and optimizing the hyperparameters through grid search; (c) model training and validation: taking the mean squared error (MSE) as the loss function, the coefficient of determination R² ≥ 0.85 and the mean absolute error (MAE) ≤ 0.5m³ / min on the test set after training; finally, the instantaneous output prediction result is output with the key features as input.

6. The method for decoupling and analyzing rakehead digging production data based on machine learning according to claim 2, characterized in that, In step (7), the parameter optimization suggestions include adjusting the operating parameters corresponding to key features: the right active rake head angle, the right suction vacuum, and the right upper vertical angle, in order to improve instantaneous output.

7. The method for decoupling and analyzing rakehead digging production data based on machine learning according to claim 1, characterized in that, In step (3), the initial feature set includes the right active rake head angle, the right high-pressure flushing pressure, and the right suction vacuum. The instantaneous output is calculated by the right concentration and the right flow rate.

8. The method for decoupling and analyzing rakehead digging production data based on machine learning according to claim 1, characterized in that, In step (5), the processing of the redundant features is as follows: if the correlation coefficient between the right compensator stroke and the right compensator pressure is greater than the threshold, then the right compensator pressure is retained; if the right mud pump speed is related to the main engine speed, then the right mud pump speed is retained.

9. The method for decoupling and analyzing rakehead digging production data based on machine learning according to claim 1, characterized in that, In step (6), the input data of the production prediction model needs to be preprocessed in step (2) and coupled in step (5) to ensure that the causal relationship between features and production is consistent and without redundancy.

10. A machine learning-based decoupling analysis system for harrow digging production data, characterized in that, The method for decoupling and analyzing harrow digging production data based on machine learning, as described in any one of claims 1 to 9, wherein the machine learning-based harrow digging production data decoupling and analysis system comprises: The data acquisition module is used to acquire multi-source historical data, including equipment sensor data, environmental monitoring data, and construction log data. The preprocessing module is used to preprocess the multi-source historical data, including: a. Perform outlier processing to obtain preliminary cleaned data; b. Fill in the missing values ​​in the preliminary cleaned data to obtain the filled data; c. Filter the right-hand parameters in the filled data, keeping only the fields related to the right rake head, to obtain the right-hand parameter data; d. Perform feature standardization on the parameter data on the right to obtain standardized data; e. Perform time-delay correction on the standardized data, adjust the data alignment causal relationship, and obtain time-delay corrected data; The construction and calculation module is used to construct an initial feature set from the time-delay correction data and calculate the instantaneous output; The evaluation module is used to evaluate the feature importance of the initial feature set with instantaneous output as the objective, using four machine learning algorithms: Random Forest, LightGBM, Lasso, and CatBoost, to obtain the importance score of each feature. A redundancy processing module is used to perform coupling analysis on the initial feature set based on the importance scores of each feature, use the Pearson correlation coefficient to detect the linear correlation between features, identify and remove redundant features, and obtain a set of key features. The prediction module is used to construct a production prediction model based on the set of key features, taking the key features as input and outputting instantaneous production prediction results; the production prediction model adopts the gradient boosting regression algorithm.

Images