Trailing suction hopper dredger yield end-to-end prediction system and method based on physical information neural network

Abstract

The invention provides an end-to-end prediction system and method for the output of a trailing suction dredger based on a physical information neural network, and the method achieves the precise prediction of the output of the trailing suction dredger through the fusion of a physical law and a neural network model. Data feature selection is optimized according to a construction stage database division method, and the generalization ability of the model is improved; the construction efficiency and the equipment utilization rate are improved through a method for optimizing construction parameters through real-time yield prediction; a visual decision support is provided for construction management through a method for displaying a yield prediction result in real time by a visual module; a set of working method is provided based on the drag suction hopper dredger yield prediction system, the end-to-end process of the system from data acquisition, preprocessing, model training to yield prediction can be realized, the operation process is simplified, and the prediction efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of physical information neural networks and production prediction, and in particular to a system and method for end-to-end prediction of the production of a trailing suction hopper dredger based on a physical information neural network. Background Art

[0002] Trailing suction hopper dredgers are core equipment in dredging projects, and their production forecasting is crucial for construction efficiency, cost control, and project management. However, due to the massive amount of construction data and the interconnected nature of the system, existing technologies suffer from the following issues:

[0003] 1. Data issues: There is a lack of a comprehensive dredging database, and the selection of data features under different construction scenarios is unclear.

[0004] 2. Model problem: The system models of the drag head, mud pump-pipeline, mud tank, etc. are complex, and lack a complete model mechanism analysis and establishment.

[0005] 3. Forecasting problem: There is a lack of accurate production forecasting methods, which makes it difficult to meet actual construction needs.

[0006] To address the above-mentioned problems, prior art patent CN111271300B discloses an online identification and control method for the mud pump speed of a dredger based on a weighted direct determination of a neural network, but does not predict the output of the neural network itself. Prior art CN115600746A discloses a method for predicting the energy efficiency of a trailing hopper dredger based on a convolutional neural network, but mainly focuses on the construction of the neural network and lacks physical model support, resulting in a need for improved prediction accuracy. Prior art CN111859777A discloses a method for calculating the process parameters of an intelligent excavation vessel for a trailing hopper dredger, but mainly relies on data feature extraction and regression prediction of a neural network and lacks physical model support, resulting in a need for improved prediction accuracy. Prior art CN117407799A discloses a method and system for construction process identification and process analysis using ship navigation tracks and GIS information data provided by a trailing hopper dredger automatic identification system, but the prediction and optimization methods are completely based on historical databases, require a large amount of data, and have low versatility.

[0007] Therefore, in order to solve the problem of difficult production prediction of trailing suction hopper dredgers and improve work efficiency, it is urgently necessary to build a comprehensive dredging database and select data features under different construction scenarios, build and analyze a complete mechanism model, and provide an accurate production prediction method to meet actual construction needs. Summary of the Invention

[0008] In view of the defects in the prior art, the purpose of the present invention is to provide an end-to-end prediction system and method for the output of a trailing suction hopper dredger based on a physical information neural network.

[0009] The end-to-end prediction system for trailing suction hopper dredger production based on physical information neural network provided by the present invention includes a hardware system and a software system;

[0010] The hardware system includes:

[0011] The sensor data acquisition terminal collects construction data in real time through mud pump speed sensors, pressure sensors, drag head depth sensors and ship motion sensors;

[0012] Construction data display, used to display construction data and production forecast results in real time;

[0013] Electronic chart display, used to overlay the electronic chart to display the vessel position, construction area and predicted production distribution heat map;

[0014] The software system includes:

[0015] The data preprocessing module is used to clean, denoise and standardize the collected data, build a training data set, and divide the database according to the construction stage;

[0016] Physical information neural network model module, used to integrate physical law constraints to train neural networks, predict mud flow rate and density, and calculate real-time production;

[0017] Yield prediction module, which implements end-to-end yield prediction based on the trained model;

[0018] Visualization module displays prediction results and construction parameter optimization suggestions in real time.

[0019] Preferably, the data preprocessing module specifically performs the following steps:

[0020] The sensors collect mud pump speed, mud pump power, mud pump pressure, drag head digging depth, digging width and leakage data;

[0021] Cluster analysis and principal component analysis are used to optimize data feature selection and divide the training set and test set into preset proportions;

[0022] The data are normalized, and sub-databases for full-speed loading, loading overflow, full-load sailing and light-load sailing stages are constructed according to the construction stage division rules.

[0023] Preferably, the physical information neural network model module adopts an LSTM network architecture, including:

[0024] In the input layer, the time step is set to 60 seconds of historical data, and the input dimension is (60, n), where n is the total number of ship speed, mud pump parameters, and soil characteristic parameters;

[0025] The first hidden layer contains 128 LSTM units and uses the tanh activation function;

[0026] The second hidden layer contains 64 LSTM units, followed by a Dropout layer with a dropout rate of 0.2;

[0027] The output layer predicts the mud flow rate and density and calculates the real-time production based on the following formula:

[0028]

[0029] Among them, W t is the instantaneous value of excavation production, D is the inner diameter of the mud pipe, V is the mud flow rate, γ m is the mud density.

[0030] Preferably, the loss function of the physical information neural network model is a weighted sum of the prediction error term and the physical constraint term, specifically:

[0031] Forecast error term: Among them, y i is the true value, is the predicted value, N is the number of samples;

[0032] Physical constraints: Among them, f(y i ) Calculated based on physical laws such as the mud pump characteristic curve, the mud pump speed-power-pressure relationship, and the drag head digging depth-width-leakage relationship;

[0033] Total loss function: L total =αL pred +βL phy , where α and β are weight coefficients used to balance the impact of prediction error and physical constraints.

[0034] Preferably, it also includes a sensitivity analysis and model validation module for:

[0035] Use historical data of different lengths to train the model and analyze the impact of data length on prediction accuracy;

[0036] The root mean square error (RMSE) is calculated on the test set to verify the model. The formula is:

[0037]

[0038] Optimize data selection strategies and hyperparameters based on validation results.

[0039] Preferably, the visualization module implements the following functions:

[0040] Dynamically display the changing curves of mud flow rate, density and real-time production;

[0041] Provide construction parameter optimization suggestions, including mud pump speed adjustment range and drag head depth threshold;

[0042] Based on the electronic chart overlay, the ship position, construction area and predicted production distribution heat map are displayed.

[0043] The end-to-end prediction method for trailing suction hopper dredger production based on physical information neural network provided by the present invention comprises the following steps:

[0044] Step 1: Collect construction data in real time through sensors, and perform cleaning and standardization processing;

[0045] Step 2: Divide the database into the full-speed loading, loading overflow, fully loaded sailing and light-loaded sailing stages to optimize feature selection;

[0046] Step 3: Construct a physical information neural network model and integrate the mud pump characteristic curve and drag head dynamic constraints for training;

[0047] Step 4: Optimize data selection strategy and verify model accuracy through sensitivity analysis;

[0048] Step 5: Predict production in real time based on the trained model and output the results through the visualization module.

[0049] Preferably, the neural network training in step 3 adopts an adaptive optimization algorithm, including Adam or RMSprop, and the optimization goal is to minimize the total loss function L total , the learning rate is determined by grid search.

[0050] Preferably, the stage division rule in step 2 is specifically:

[0051] Full-speed loading stage: The key characteristics are mud pump speed, power, and pressure, which are correlated with loading efficiency and soil properties;

[0052] Loading overflow stage: overflow flow, mud tank level, and mud pump pressure are used as key features to correlate overflow efficiency;

[0053] Fully loaded sailing stage: ship speed, mud tank level, and mud pump power are used as key characteristics to correlate with sailing resistance;

[0054] Light-load sailing phase: The ship speed and mud tank liquid level are used as key characteristics, and the fuel consumption rate is correlated.

[0055] Preferably, the real-time prediction in step 5 includes:

[0056] Based on the output of mud flow rate and density from the LSTM network, instantaneous production is calculated in combination with the inner diameter of the mud pipe;

[0057] Accumulate the instantaneous production of each time step and generate hourly, daily and task-level production reports;

[0058] Dynamically adjust the mud pump speed threshold and drag head excavation depth based on the prediction results to optimize construction efficiency.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] (1) By adopting a method that integrates physical laws and neural network models, compared with traditional neural network models, the output parameters are constrained by the physical model to reduce errors, and the production of the trailing suction hopper dredger is accurately predicted, which solves the problem of difficult production prediction of the trailing suction hopper dredger;

[0061] (2) By adopting the method of dividing the database according to the construction stage, the data feature selection is optimized, the neural network model is established with a small amount of data, the dependence on the historical database is reduced, and the generalization ability of the model is improved;

[0062] (3) By adopting the real-time production prediction method, it provides technical support for optimizing construction parameters and improving construction efficiency and equipment utilization;

[0063] (4) Provide intuitive decision support for construction management by using a visualization module to display production forecast results in real time;

[0064] (5) Based on the above-mentioned system for predicting the production of a trailing suction hopper dredger, a set of working methods is provided, which can realize the end-to-end process of the system from data collection, preprocessing, model training to production prediction, simplify the operation process and improve the prediction efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0066] Figure 1 To invent the system architecture diagram;

[0067] Figure 2 This is the data collection and preprocessing flow chart;

[0068] Figure 3 Construct a flow chart for a typical soil and vessel dredging database;

[0069] Figure 4 Construct a flow chart for the physical information neural network model;

[0070] Figure 5 Flowchart for sensitivity analysis and model validation;

[0071] Figure 6 Provide a flow chart for end-to-end forecasting of production and system functionality. DETAILED DESCRIPTION

[0072] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0073] Example

[0074] The present invention provides an end-to-end prediction system for trailing suction hopper dredger production based on physical information neural network, which includes two parts: hardware and software system. Figure 1 shown.

[0075] Display hardware system: includes sensor data acquisition terminals, construction data displays, and electronic chart displays. The sensor data acquisition terminals, including dredge pump speed sensors, pressure sensors, and drag head depth sensors, collect real-time data such as dredge pump speed, pressure, drag head depth, and vessel motion. The construction data display displays construction data and production forecasts in real time. The electronic chart display overlays the electronic chart to display vessel position, construction area, and a heat map of predicted production distribution.

[0076] Software system: This includes a data preprocessing module, a physical information neural network model module, a yield prediction module, and a visualization module, enabling full-process functionality from data acquisition to yield prediction. The data preprocessing module cleans and standardizes the collected data, constructs a training dataset, and divides the database according to construction phases, setting phase division rules and optimizing data feature selection. The physical information neural network model module uses historical data to train the physical information neural network, optimizes hyperparameters and weight parameters, analyzes the model's sensitivity to data duration, optimizes data selection strategies, verifies prediction results with actual construction data, and continuously optimizes the model. The yield prediction module and visualization module, based on the trained model, implement an end-to-end process from data acquisition, preprocessing, model training, to yield prediction, predicting the yield of the trailing suction hopper dredger in real time and displaying it in real time.

[0077] Based on the above hardware and software systems, a control system and control method are proposed. The technical solution of the present invention includes the following steps:

[0078] Step 1: Data collection and preprocessing. Figure 2 As shown in the figure, various sensors are installed on the trailing suction hopper dredger to collect construction data in real time. The collected data is cleaned and standardized to construct a training dataset.

[0079] Step 101, using a variety of sensors installed on the trailing suction hopper dredger (such as a mud pump speed sensor, a pressure sensor, a drag head depth sensor, etc.) to collect construction data in real time, including mud pump speed, mud pump power, mud pump pressure, drag head digging depth, digging width, leakage amount, etc.

[0080] Step 102: Clean, denoise, and standardize the collected data to construct a high-quality training data set.

[0081] Step 2: Construction of typical soil and ship dredging database. Figure 3 As shown, the database is divided according to the construction stage, the stage division rules are set, and the data feature selection is optimized.

[0082] Step 201, construction stage division: according to different construction stages of the trailing suction hopper vessel (such as full-speed loading, loading overflow, full-load sailing and light-load sailing stages), stage division rules are set to divide the database.

[0083] Full-speed loading stage: The main characteristics are mud pump speed, mud pump power and mud pump pressure, and the loading efficiency and soil characteristics are recorded.

[0084] Loading and overflow stage: The overflow flow, mud tank liquid level and mud pump pressure are the main characteristics, and the overflow efficiency and soil characteristics are recorded.

[0085] Fully loaded sailing stage: The main characteristics are ship speed, mud tank level and mud pump power, and the sailing efficiency and soil characteristics are recorded.

[0086] Light-load navigation stage: The main characteristics are ship speed, mud tank level, and mud pump power, and the navigation efficiency and soil characteristics are recorded.

[0087] Step 202, data feature selection: extracting inherent patterns from data at different stages through machine learning methods (such as cluster analysis and principal component analysis) to optimize data feature selection.

[0088] Step 203: Divide the data into a training set and a test set in a ratio of 80% and 20%, and perform normalization processing on them.

[0089] Step 3: Construct the physical information neural network model. Figure 4 As shown in Figure 1, physical laws are incorporated into the neural network model as constraints. Historical data is used to train the physical information neural network and optimize hyperparameters and weight parameters.

[0090] Step 301: neural network model construction.

[0091] LSTM (Long Short-Term Memory) is a recurrent neural network (RNN) variant suitable for time series forecasting. It can effectively capture long-term dependencies in time series data. In this paper, LSTM neural networks are used to predict flow rate and density to calculate real-time production. The following is the specific architecture design of the neural network:

[0092] The time steps of the input layer are set based on actual needs. Data from the past several seconds or minutes is typically used as input. For example, if data from the past 60 seconds is selected and collected once per second, the input dimension is (60, n), where n is the number of input parameters (such as ship speed and mud pump sensor parameters).

[0093] The LSTM network consists of multiple hidden layers, each containing several LSTM units. The number of hidden layers and units can be selected based on data complexity and computing resources. For example:

[0094] The first LSTM layer contains 128 LSTM units and is used to preliminarily extract time series features.

[0095] The second LSTM layer contains 64 LSTM units, which are used to further extract high-order features.

[0096] Dropout layer: Add a Dropout layer (such as dropout rate = 0.2) after each LSTM layer to prevent overfitting.

[0097] The tanh activation function is used inside the LSTM unit to control the flow and memory of information.

[0098] The output layer predicts the mud flow rate and density and calculates the real-time production using the following formula:

[0099]

[0100] Among them, W t is the instantaneous value of excavation production, D is the inner diameter of the mud pipe, V is the mud flow rate, γ m is the mud density.

[0101] Step 302: Fusion of physical laws.

[0102] The loss function measures the error between the model's predictions and the true value. In this paper, the loss function not only considers the prediction error but also incorporates physical laws such as the mud pump characteristic curve, the relationship between mud pump speed, power, and pressure, and the relationship between the drag head's digging depth, width, and leakage as constraints. These constraints are then incorporated into the neural network model to enhance the model's physical consistency.

[0103] Use the mean squared error (MSE) as the prediction error term to calculate the difference between the predicted value and the true value:

[0104]

[0105] Among them, y i is the true value, is the predicted value, and N is the number of samples.

[0106] The physical models of the drag head and mud pump are used as constraints and incorporated into the loss function:

[0107]

[0108] For example, based on the relationship between the digging depth, width and leakage of the rake head, the predicted results are compared with the actual results. i ) Based on the physical laws such as the mud pump characteristic curve, the mud pump speed-power-pressure relationship, and the drag head digging depth-width-leakage relationship, the excavation output Q is calculated as:

[0109] Q=H cutter ×B cutter ×v cutter ×(1-n overflow )

[0110] Among them, H cutter B is the effective cutting depth of the rake head, cutter is the cutting width, v cutter is the dredging speed, n overflow is the overflow rate.

[0111] The total loss function is the weighted sum of the prediction error term and the physical constraint term:

[0112] L total =αL pred +βL phy

[0113] Among them, α and β are weight coefficients used to balance the influence of prediction error and physical constraints.

[0114] Step 303, hyperparameter optimization: Optimize the hyperparameters of the neural network model (such as learning rate, number of hidden layers, number of neurons, etc.) through grid search, Bayesian optimization and other methods.

[0115] Step 304, model training: using the training set data in the historical data as input and output parameters, training a physical information neural network (PINN) to predict the output of the trailing suction hopper dredger.

[0116] Step 305, weight parameter optimization: Adopt an adaptive optimization algorithm (such as Adam, RMSprop) to optimize the weight parameters of the neural network to improve the convergence speed and prediction accuracy of the model.

[0117] Step 4: Sensitivity analysis and model validation. Figure 5 As shown, we analyze the model's sensitivity to data duration and optimize the data selection strategy. We validate the prediction results with actual construction data and continuously optimize the model.

[0118] Step 401, sensitivity analysis of data of different durations: Use historical data of different durations (such as short-term, medium-term, and long-term) to train a neural network model, analyze the model's sensitivity to the duration of the data, and optimize the data selection strategy.

[0119] Step 402, model validation: using the test set data from the actual construction data as input parameters, verifying the accuracy of the prediction results through evaluation parameters such as the root mean square error (RMSE), and continuously optimizing the model to reduce the error;

[0120]

[0121] Step 5: End-to-end production forecast and system function realization. Figure 6 As shown in the figure, based on the trained model, an end-to-end process from data collection, preprocessing, model training to yield prediction is implemented, and the yield of a trailing suction hopper dredger is predicted in real time. The hardware and software systems are designed to implement the full process from data collection to yield prediction.

[0122] Step 501, end-to-end production prediction: Based on the trained physical information neural network model, an end-to-end process from data collection, preprocessing, model training to production prediction is implemented to predict the production of the trailing suction hopper dredger in real time.

[0123] Those skilled in the art will appreciate that, in addition to implementing the system, device, and various modules provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, embedded microcontrollers, and the like by logically programming the method steps. Therefore, the system, device, and various modules provided by the present invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; the modules for implementing various functions can also be considered both software programs for implementing the method and structures within the hardware component.

[0124] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

Claims

1. An end-to-end prediction system for trailing suction hopper dredger production based on physical information neural network, characterized in that: Including hardware system and software system; The hardware system includes: The sensor data acquisition terminal collects construction data in real time through mud pump speed sensors, pressure sensors, drag head depth sensors and ship motion sensors; Construction data display, used to display construction data and production forecast results in real time; Electronic chart display, used to overlay the electronic chart to display the vessel position, construction area and predicted production distribution heat map; The software system includes: The data preprocessing module is used to clean, denoise and standardize the collected data, build a training data set, and divide the database according to the construction stage; Physical information neural network model module, used to integrate physical law constraints to train neural networks, predict mud flow rate and density, and calculate real-time production; Yield prediction module, which implements end-to-end yield prediction based on the trained model; Visualization module displays prediction results and construction parameter optimization suggestions in real time.

2. The end-to-end prediction system for trailing suction hopper dredger production based on physical information neural network according to claim 1 is characterized in that: The data preprocessing module specifically performs the following steps: The sensors collect mud pump speed, mud pump power, mud pump pressure, drag head digging depth, digging width and leakage data; Cluster analysis and principal component analysis are used to optimize data feature selection and divide the training set and test set into preset proportions; The data are normalized, and sub-databases for full-speed loading, loading overflow, full-load sailing and light-load sailing stages are constructed according to the construction stage division rules.

3. The end-to-end prediction system for trailing suction hopper dredger production based on physical information neural network according to claim 1 is characterized in that: The physical information neural network model module adopts the LSTM network architecture, including: In the input layer, the time step is set to 60 seconds of historical data, and the input dimension is (60, n), where n is the total number of ship speed, mud pump parameters, and soil characteristic parameters; The first hidden layer contains 128 LSTM units and uses the tanh activation function; The second hidden layer contains 64 LSTM units, followed by a Dropout layer with a dropout rate of 0.2; The output layer predicts the mud flow rate and density and calculates the real-time production based on the following formula: Among them, W t is the instantaneous value of excavation production, D is the inner diameter of the mud pipe, V is the mud flow rate, γ m is the mud density.

4. The end-to-end prediction system for trailing suction hopper dredger production based on physical information neural network according to claim 3 is characterized in that: The loss function of the physical information neural network model is the weighted sum of the prediction error term and the physical constraint term, specifically: Forecast error term: Among them, y i is the true value, is the predicted value, N is the number of samples; Physical constraints: Among them, f(y i ) Calculated based on the mud pump characteristic curve, mud pump speed-power-pressure relationship, and drag head digging depth-width-leakage relationship; Total loss function: L total =αL pred +βL phy , where α and β are weight coefficients used to balance the impact of prediction error and physical constraints.

5. The end-to-end prediction system for trailing suction hopper dredger production based on physical information neural network according to claim 4 is characterized in that: Also includes sensitivity analysis and model validation modules for: Use historical data of different lengths to train the model and analyze the impact of data length on prediction accuracy; The root mean square error (RMSE) is calculated on the test set to verify the model. The formula is: Optimize data selection strategies and hyperparameters based on validation results.

6. The end-to-end prediction system for trailing suction hopper dredger production based on physical information neural network according to claim 4 is characterized in that: The visualization module implements the following functions: Dynamically display the changing curves of mud flow rate, density and real-time production; Provide construction parameter optimization suggestions, including mud pump speed adjustment range and drag head depth threshold; Based on the electronic chart overlay, the ship position, construction area and predicted production distribution heat map are displayed.

7. A method for predicting the output of a trailing suction hopper dredger based on the system for predicting the output of a trailing suction hopper dredger based on a physical information neural network according to any one of claims 1 to 6, characterized in that: The following steps are involved: Step 1: Collect construction data in real time through sensors, and perform cleaning and standardization processing; Step 2: Divide the database into the full-speed loading, loading overflow, fully loaded sailing and light-loaded sailing stages to optimize feature selection; Step 3: Construct a physical information neural network model and integrate the mud pump characteristic curve and drag head dynamic constraints for training; Step 4: Optimize data selection strategy and verify model accuracy through sensitivity analysis; Step 5: Predict production in real time based on the trained model and output the results through the visualization module.

8. The end-to-end prediction method for trailing suction hopper dredger production based on physical information neural network according to claim 7 is characterized in that: In step 3, the neural network training adopts an adaptive optimization algorithm, including Adam or RMSprop, and the optimization goal is to minimize the total loss function L total , the learning rate is determined by grid search.

9. The end-to-end prediction method for trailing suction hopper dredger production based on physical information neural network according to claim 7, characterized in that: The specific stage division rules in step 2 are: Full-speed loading stage: The key characteristics are mud pump speed, power, and pressure, which are correlated with loading efficiency and soil properties; Loading overflow stage: overflow flow, mud tank level, and mud pump pressure are used as key features to correlate overflow efficiency; Fully loaded sailing stage: ship speed, mud tank level, and mud pump power are used as key characteristics to correlate with sailing resistance; Light-load sailing phase: The ship speed and mud tank liquid level are used as key characteristics, and the fuel consumption rate is correlated.

10. The end-to-end prediction method for trailing suction hopper dredger production based on physical information neural network according to claim 7, characterized in that: The real-time prediction in step 5 includes: Based on the output of mud flow rate and density from the LSTM network, instantaneous production is calculated in combination with the inner diameter of the mud pipe; Accumulate the instantaneous production of each time step and generate hourly, daily and task-level production reports; Dynamically adjust the mud pump speed threshold and drag head excavation depth based on the prediction results to optimize construction efficiency.

Images