A Method for Calculating the Safety Angle of the Trailer Arm of a Trailer Suction Hopper Based on Neural Networks and Geometric Analysis

The method for calculating the safe angle of the drag arm using neural networks and geometric analysis solves the problem of the lack of quantitative safety judgment for drag suction dredgers under crossflow conditions. It realizes the prediction of dynamic safe distance and the solution of static limit angle of the drag arm system, thereby improving operational safety and construction efficiency.

CN123115659APending Publication Date: 2026-11-17CCCC SHANGHAI DREDGING CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611277370.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, trailing suction hopper dredgers lack quantitative assessment of the safety status of the drag arm system under crossflow conditions, leading to frequent drag collision accidents. Traditional methods cannot adapt to changes in dredging depth and environmental parameters, and lack quantitative decision-making basis.

Method used

A method based on neural networks and geometric analysis is adopted. The mapping relationship between environmental parameters and the safe distance of the rake head-rake arm system is established through physical information neural network. Combined with the inherent geometric relationship of the rake arm system, a safety criterion for the horizontal angle of the rake arm is constructed to realize the prediction of dynamic safe distance and the solution of static limit horizontal angle, providing quantitative safety operation guidance.

Benefits of technology

It significantly improves the operational safety and construction efficiency of trailing suction hopper dredgers in complex sea conditions, reduces the probability of drag collisions, and realizes the transformation from qualitative judgment to quantitative and differentiated safety decision-making, adapting to adaptive safety assessments for different working conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115659A_ABST
    Figure CN123115659A_ABST
Patent Text Reader

Abstract

This invention provides a method for calculating the safe angle of the drag boom of a trailing suction hopper dredger based on neural networks and geometric analysis, comprising the following steps: S1: Physical perception D-optimal sequential sampling and simulation data acquisition; S2: Construction of a Bayesian regularized physical information neural network; S3: Hyperparameter optimization and two-stage cascaded network training and verification; S4: Solving the vertical angle of the ship-drag boom geometric relationship; S5: Constructing constraint equations and solving the limiting horizontal angle using the homotopy extension iterative method; S6: Constructing a safety criterion that integrates dynamic prediction and static limits, and closed-loop feedback. This invention transforms abstract spatial spacing constraints into angle indicators that can be directly monitored from the bridge, solving the technical challenge of rapid quantitative assessment of the drag boom's safe angle under crossflow conditions, and improving the safety and decision-making efficiency of crossflow dredging operations.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of safety monitoring technology for trailing suction hopper dredging vessels, and in particular to a method for calculating the safety angle of the trailing suction hopper boom based on neural networks and geometric analysis. Background Technology

[0002] Trailing suction hopper dredgers, with their ability to dredge while simultaneously navigating, have become the most widely used vessel type in waterway maintenance and port construction. The trailing arm system is the core dredging mechanism of a trailing suction hopper dredger, using a skid head to excavate sediment from the riverbed and then pumping it to the silt tank via a suction pipe. The trailing arm typically employs a multi-section articulated structure, connecting the upper and lower arms via universal joints, allowing it to swing horizontally and vertically to adapt to different dredging depths and water flow conditions.

[0003] During the operation of a trailing suction hopper dredging vessel, the operators must constantly monitor the status of the dredging arm during operations such as dredging, raising the dredging arm, and lowering the dredging arm. Under crosscurrent conditions, the dredging head-arm system experiences lateral oscillation in the horizontal plane under the combined excitation of waves and ocean currents. When the lateral oscillation amplitude exceeds a certain critical value, the spatial distance between the dredging arm system (including the universal joint rubber pad, dredging head rubber pad, dredging center sling, or dredging head sling) and the side of the hull shrinks to zero, causing a dredging collision accident.

[0004] In existing technologies, construction workers lack quantitative judgment on whether the current working conditions are suitable for operation, and decisions rely entirely on personal experience. Even when the rake arm is lowered during operation, the operator cannot know the precise safety margin corresponding to the current rake arm posture. Some existing technologies use fixed angle thresholds as safety criteria, but such methods cannot adapt to changes in digging depth, environmental parameters, and the physical reality of risk-critical components shifting with working conditions, leading to a significant risk of misjudgment in safety assessments. Although some studies involve rake arm angle control or rake arm posture measurement, none of them address methods for transforming the complex environmental-structural coupled dynamic response into quantitative operational indicators that can be directly monitored from the operator's platform. Summary of the Invention

[0005] This invention provides a method for calculating the safe angle of the drag arm of a drag suction hopper dredger based on neural networks and geometric analysis. By establishing a mapping relationship between environmental parameters (sense wave height, spectral peak period, and crossflow velocity) and the safe distance between the drag head and the drag arm system through a physical information neural network, a rapid prediction of the dynamic safe distance is achieved. At the same time, based on the inherent geometric relationship of the ship-drag arm system, an iterative solution algorithm is established to transform the abstract safe distance constraint into the horizontal angle of the upper and lower drag arm, which can be directly monitored by the bridge, thereby providing clear and quantitative safe operation guidance for the drag operator.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for calculating the safe angle of the trailing suction hopper dredger boom based on neural networks and geometric analysis includes the following steps: S1: Obtain environmental parameters of the sea area where the trailing suction hopper dredging vessel is operating, construct a candidate sample pool based on the range of variation of environmental parameters and select the optimal training sample set, then input it into the time-domain coupled dynamic simulation model for simulation, extract the safety distance statistics from the simulation results as output parameters, and form a training dataset. S2: Construct a physical information neural network, perform standardization processing on the training dataset, and train the physical information neural network using the standardized training dataset and preset physical constraints to obtain the trained physical information neural network. S3: Perform cross-validation on the trained physical information neural network to solidify the physical information neural network that meets the accuracy requirements into a prediction model; S4: Obtain the inherent geometric parameters and adjustable parameters of the trailing suction hopper dredging vessel, establish the geometric constraint relationship between the drag arm system and the hull based on the inherent geometric parameters and adjustable parameters, solve the vertical angle of the drag arm system according to the geometric constraint relationship, and determine the water depth range to which the current working condition belongs based on the vertical angle to identify the critical risk components. S5: Input environmental parameters into the prediction model and output the predicted dynamic safety distance between the rake arm system and the hull. Based on the predicted dynamic safety distance, critical risk components, vertical angle and preset minimum safety distance in the adjustable operation parameters, construct the rake arm horizontal angle constraint equation and solve it to obtain the static limit horizontal angle. S6: The predicted dynamic safety distance is fused with the static limit horizontal angle to generate a fused safety criterion. The safety angle of the rake arm is output based on the fused safety criterion, and the weight coefficients of the physical constraints are adjusted according to the fused safety criterion. The results are then fed back to S2 for subsequent incremental training.

[0007] In summary, this invention has at least the following beneficial effects: By using the physical perception D-optimal criterion to select the training samples with the most information from candidate working conditions, and combining them with a Bayesian regularized physical information neural network, it achieves a leap from small-sample simulation data to a high-precision prediction model, solving the bottleneck problem of traditional neural networks relying on large-scale training data. By simultaneously introducing first-order and second-order physical monotonicity constraint loss terms, it ensures both the basic physical law that the safety distance decreases monotonically with environmental parameters and constrains the downward convexity of the curve shape, enabling the model to maintain a reasonable shape consistent with physical laws even in extrapolation regions not covered by training data. This significantly improves the model's extrapolation ability and prediction reliability, overcoming the defect of pure data-driven models in predicting distortion in sparse data regions. Through an iterative solution algorithm for the limit horizontal angle based on the inherent geometric relationship between the ship and the boom, the abstract safety distance constraint is transformed into a boom horizontal angle index that can be directly monitored from the bridge. This allows boom operators to make safety decisions based on the angle instrument panel without understanding complex dynamic mechanisms, solving the technical problem of lacking quantitative decision-making basis for on-site operators. By constructing a fusion safety margin through a segmented approach combining dynamic prediction and static limits, a multi-level early warning system is introduced. Each early warning level has a clearly defined operational time window requirement. Combined with health threshold triggering graded decision-making instructions, this achieves a shift from qualitative judgment to quantitative and differentiated safety decision output, avoiding the limitations of traditional binary safety / hazard determination methods. Through automatic identification of critical components under different dredging depths and adaptive switching of constraint equations, adaptive safety assessment is achieved for three typical working conditions: shallow water, medium water depth, and deep water. This overcomes the fundamental defects of traditional fixed threshold methods, which cannot adapt to changes in working conditions and cannot identify the transfer of critical components. This effectively reduces the probability of drag-suction collisions and improves the operational safety and construction efficiency of trailing suction hopper dredgers in complex sea conditions. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of a method for calculating the safe angle of the drag boom of a trailing suction hopper dredger based on neural networks and geometric analysis.

[0009] Figure 2 This is a schematic diagram illustrating the process of training and dynamic prediction of a physical information neural network.

[0010] Figure 3 This is a flowchart illustrating the geometric analysis and the solution of the limit horizontal angle.

[0011] Figure 4 A flowchart illustrating the integration of safety criteria and closed-loop feedback. Detailed Implementation

[0012] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0013] like Figure 1As shown, this embodiment provides a method for calculating the safety angle of the trailing suction hopper dredger's boom based on neural networks and geometric analysis, including the following steps: S1: Physical sensing D-optimal sequential sampling and simulation data acquisition.

[0014] A three-dimensional environmental parameter space is constructed, covering the range of environmental parameter variations in the sea area where the target trailing suction hopper dredger is operating. The three dimensions of the three-dimensional environmental parameter space are: significant wave height, etc. Spectral peak period and crossflow velocity Based on historical environmental statistics of the target trailing suction hopper dredger in the Yangtze River estuary, the wave height is [missing information]. The range of values ​​is Spectral peak period The range of values ​​is Crossflow velocity The range of values ​​is The wave spectrum was simulated using the JONSWAP spectrum with a peak factor of 2.4, and the energy distribution of the wave spectrum was simulated according to the standard JONSWAP format.

[0015] The three-dimensional environmental parameter space is discretized along the three dimensions as follows: Each node. For the spectral peak period... A non-uniform discretization method is employed, with increased sampling density near the ship's natural roll period. This is based on the ship's natural roll frequency. Corresponding inherent period ,in The known design parameters of the target vessel can be obtained from the vessel loading manual or roll attenuation tests; for the spectral peak period... Nodes are generated using the following nonlinear mapping: , , ; in For the first Each spectral peak period node value For node sequence number, For parameters with uniform distribution, an exponential mapping is used to make the nodes more densely distributed at the end of the interval (near the natural period). Significant wave height. and crossflow velocity The discretization method is still used with equal spacing. A total of three dimensions are generated. The candidate sampling points constitute the candidate sample pool. , of which The input vector of candidate sampling points , This is the index number of the candidate sampling point in the candidate sample pool. .

[0016] To implement the D-optimal criterion, a basis function vector is defined to construct the design matrix. This basis function vector is used only to characterize the coordinate distribution of candidate sample points in the input space and is used for the subsequent calculation of the Fisher information matrix; it is not involved in fitting the safety margin output parameters. (Definition of basis function vector) : ; in For any environment, the input vector is .

[0017] For any candidate sampling point in the candidate sample pool Its corresponding basis function vector is For any selected subset of samples Define the design matrix The design matrix of the first Line by number The transpose of the basis function vectors corresponding to each selected sample point constitutes: ; in For sample subset The number of samples in the middle For sample subset The first in The input vector of each sample, .

[0018] Define the Fisher information matrix : ; A sequential forward selection algorithm is used to filter sample points from the candidate sample pool. Initially, the selected sample set is... Set it to an empty set. In the... In the next iteration Traverse the candidate sample pool All candidate sampling points that were not selected , calculation will Temporary addition The newly formed design matrix The increment of the determinant of the corresponding Fisher information matrix relative to the value before addition. : ; Select to increment the determinant value Largest candidate sampling point Add to selected sample set: ; ; Repeat the above process until the stopping condition is met. When a sample set is selected... After ensuring that the number of samples is no less than the dimension of the basis function vector is 10, the currently selected sample set is calculated after each iteration. The corresponding determinant value of the Fisher information matrix And calculate this value relative to the candidate sample pool. The determinant value of the Fisher information matrix of the design matrix consisting of all 1000 samples. Information coverage : ; when Stop iterating when the optimal training sample set is obtained. If during the iteration process Less than If it is determined that the currently selected sample set has approximately multicollinearity, then the current sample set will be... Clear the pool of candidate samples and restart the sequential selection process.

[0019] In the sequential forward selection process, the candidate sample pool The design matrix consists of all 1000 candidate points. Fisher information matrix determinant value As the baseline denominator, calculate the currently selected sample set. Information coverage It should be noted that the above Fisher information matrix only utilizes the input parameters. The calculation is performed using the basis function values, without involving simulation output. The specific values ​​are not specified, therefore this sampling criterion aims to maximize the information coverage of the input space and ensure the representativeness of the selected sample points in the three-dimensional environmental parameter space. Based on this calculation method, The evolution data is as follows: after the first 10 iterations After the 15th iteration After the 18th iteration After the 20th iteration After the 21st iteration After the 20th iteration The 95% threshold was not reached after the 21st iteration. Since the 95% threshold has been exceeded, 21 is indeed the minimum required number of samples. The optimal training sample set is then determined. .

[0020] The optimal training sample set The 21 sets of input parameters are input into a pre-established time-domain coupled dynamics simulation model of the trailing suction hopper dredger. This model includes a six-degree-of-freedom motion model of the hull, a multibody dynamics model of the trailing arm, a dynamics model of the slings, and a wave compensator model. The execution time for each set of input parameters is [duration missing]. A time-domain simulation was performed, extracting the time history data of the safe clearance between the rake head-rake arm system and the hull from the 600th second to the 11400th second within the simulation interval. Statistical analysis was performed on the time history data for each operating condition to obtain six safe clearance output parameters: the maximum value of the distance between the rake arm and the hull. Minimum distance between the rake arm and the hull Average distance between the rake arm and the hull The maximum distance between the rake head and the hull Minimum distance between the rake head and the hull Average distance between the rake head and the hull .

[0021] This constitutes the complete training dataset. , of which The output vector of each sample Defined as: ; S2: Construction of a Bayesian regularized physical information neural network.

[0022] The training dataset obtained from S1 Perform Z-score normalization. Input vector The standardized formula is: ; in Let the mean vector of all training sample input vectors be denoted as . ; Let covariance be the input vector. ; Covariance matrix The inverse matrix of the square root of , i.e. .

[0023] Output vector The standardized formula is: ; in The mean vector of the output vectors of all training samples. ; Let covariance be the output vector. ; Covariance matrix The inverse matrix of the square root of , i.e. The standardized dataset is denoted as . .

[0024] Construct a three-layer feedforward fully connected neural network, which consists of one input layer, two hidden layers, and one output layer. The input layer contains... One neuron, with three neurons corresponding to the standardized meaningful wave height. Standardized spectral peak period and standardized crossflow velocity Each of the two hidden layers contains The output layer contains 10 neurons. There are six neurons, each corresponding to a standardized safety distance output parameter.

[0025] Output of the first hidden layer The calculation expression is: ; in Standardize the input vector; This is the weight matrix of the first hidden layer; This is the bias vector of the first hidden layer; The hyperbolic tangent activation function is used. , Let be any real number as the independent variable.

[0026] Output of the second hidden layer The calculation expression is: ; in This is the weight matrix of the second hidden layer; This is the bias vector for the second hidden layer.

[0027] output layer prediction The calculation expression is: The output layer uses a linear activation function, not... Apply a nonlinear transformation.

[0028] in This is the weight matrix of the output layer; This is the bias vector for the output layer.

[0029] The total number of trainable parameters of a neural network is denoted as . The total number of trainable parameters .

[0030] Define the composite loss function The composite loss function consists of the data fitting loss term. Bayesian regularization loss term First-order physical monotonicity constraint loss term and second-order physical monotonicity constraint loss term Four weighted components: ; in The loss weight coefficients are used to fit the data. These are the weighting coefficients for the Bayesian regularization loss. These are the first-order physical constraint weighting coefficients. The first three are the weighting coefficients of the second-order physical constraints, and their optimal values ​​will be determined through a five-fold cross-validation process.

[0031] Data fitting loss term Using mean square error form: ; in For the neural network to the first Predicted output for each standardized training sample; This represents the Euclidean norm of a vector.

[0032] Bayesian regularization loss term Based on the principle of weight decay regularization: ; in For all trainable parameters of the neural network The first in The specific values ​​of each trainable parameter. .

[0033] First-order physical monotonicity constraint loss term Constructed using a perturbation testing method, the safety distance is required to decrease monotonically with any increase in environmental parameter. For the training dataset... Each standardized input sample Positive relative perturbations are applied to the three input parameters respectively. Construct three positive perturbation input vectors ,in , , These are unit vectors in the three coordinate directions of three-dimensional space. This is the index for the direction of the disturbance. The first-order constraint loss term is defined as the sum of the squares of the prediction behaviors for all violations of the monotonic relationship that the predicted safety distance after a positive disturbance is no greater than the predicted safety distance under the original input: ; in Represents the output vector of the neural network The One portion, ; Indicates taking and The larger value in Let be any real number as the independent variable.

[0034] Second-order physical monotonicity constraint loss term The second-order difference sign of the safety clearance as it varies with environmental parameters must conform to physical laws. For crossflow velocities... Physically, the safe clearance decreases monotonically with increasing flow velocity, and the rate of decrease typically increases with increasing flow velocity (the safe clearance-velocity curve is convex downwards, and the second-order difference is greater than zero). To reinforce this prior knowledge, a second-order constraint is applied to the flow velocity direction for each training sample, and similar constraints are extended to other parameters to guide the model in learning a reasonable curvature. For each training sample, a positive perturbation is also constructed. and negative disturbance The second-order constraint loss term is defined as: ; in Indicates taking and The smaller value in Let be any real number, where , and For the same neural network, the input is respectively , and The first time Each output component. When the second-order difference is less than zero (concave function shape), a positive loss value is generated, forcing the network to learn a convex, monotonically decreasing shape, so that the model still maintains a curve shape consistent with physical laws in the extrapolation region not covered by the training data.

[0035] Perform Xavier normal distribution initialization on the network weights. For connections... Each input neuron and The weight matrix of each output neuron, where each element... From normal distribution Independent sampling in, among which This refers to the row index of the weight matrix. These are the column indices of the weight matrix. All bias vectors are initialized to zero.

[0036] S3: Hyperparameter optimization and training and validation of two-stage cascaded networks.

[0037] The training and dynamic prediction process of physical information neural networks is as follows: Figure 2As shown. First, the hyperparameters defined in S2 are determined through five-fold cross-validation. , and The optimal value is found. To simplify the search space, a sequential selection strategy is adopted: first fix... , (Experience value), confirm ; then fix and ,Sure Finally fixed , ,Sure In a small-scale preliminary experiment, the order selection strategy was compared with... , The mean absolute error of the validation set obtained by the 4×4 joint grid search with the sequential selection strategy is 0.008 meters, and the optimal value of the joint grid search is 0.0075 meters. The difference between the two is less than 0.001 meters. Therefore, the sequential selection strategy and the joint grid search have comparable performance.

[0038] The 21 samples obtained in S1 are divided into 5 disjoint subsets according to the D-optimal sampling order. The first 4 subsets each contain 4 samples, and the last subset contains 5 samples. Four subsets are used as the training set, and the remaining subset as the validation set, for a total of 5 training-validation cycles. For each candidate hyperparameter set, the following two-stage training is performed on the training set, and the mean absolute error is evaluated on the validation set. The hyperparameter combination that minimizes the mean validation error is selected.

[0039] Bayesian regularization loss weight coefficient The candidate values ​​are in the interval The values ​​were taken at logarithmic intervals: 0.001, 0.003, 0.01, 0.03, 0.1, 0.3, 1.0. The mean absolute errors of the corresponding five-fold cross-validation were 0.023 meters, 0.019 meters, 0.015 meters, 0.012 meters, 0.008 meters, 0.009 meters, and 0.014 meters, respectively. This was verified. The mean absolute error of the five-fold cross-validation is the smallest, at 0.008 meters, therefore it is selected. .

[0040] First-order physical monotonicity constraint loss weight coefficient The candidate values ​​are 0.1, 0.2, 0.5, and 1.0. (Fixed) , Cross-validation was performed. The mean absolute errors for the corresponding five-fold cross-validation were 0.014 meters, 0.011 meters, 0.009 meters, and 0.015 meters, respectively. The validation set has the smallest mean absolute error, which is 0.009 meters, therefore it is selected. .

[0041] Second-order physical monotonicity constraint loss weight coefficient The candidate values ​​are 0.01, 0.05, 0.1, 0.2, 0.5, and 1.0. (Fixed) , Cross-validation was performed. The mean absolute errors of the corresponding validation sets were 0.015 m, 0.012 m, 0.009 m, 0.008 m, 0.011 m, and 0.018 m, respectively. The time error is minimized, therefore, it is selected. .

[0042] After determining the hyperparameters, a two-stage training strategy of cascading optimization of the Levenberg-Marquardt algorithm and the Adam algorithm was adopted to train the neural network.

[0043] The first stage is the Levenberg-Marquardt supervised learning stage, which aims to minimize the sum of the data fitting loss and the Bayesian regularization loss. To optimize the objective function, an augmented residual vector is constructed to accurately represent the objective function as a sum of squared residuals.

[0044] Define the data residual vector , its first The components are: ; in The global index for the sample-output pair satisfies ; For the neural network in the first Predicted values ​​under a global index; This corresponds to the standardized true value. Simultaneously, a parameter residual vector is defined. , its first The components are: ; in For the first One trainable parameter, The two are concatenated into an augmented residual vector. : ; The sum of squares of the augmented residual vector Strictly equal to .

[0045] Define the augmented Jacobian matrix The first 126 rows correspond to the partial derivatives of the data residuals with respect to each parameter, and the last 216 rows form a block diagonal matrix, with its diagonal blocks being... The LM parameter update formula is: ; Where the damping factor The initial value is set to 0.01; It is a 216th order identity matrix. If the loss function value decreases after iteration, then... Multiply by a decay factor of 0.1 and accept parameter updates; if the loss function value increases, Multiply by a growth factor of 10 and reject parameter updates. The convergence criterion is... Alternatively, by reaching the maximum number of iterations (1000), the optimal parameters for the first stage can be obtained. .

[0046] The second phase is the Adam physics constraint fine-tuning phase, with Using the initial values, the Adam algorithm is applied to the complete composite loss function. Fine-tuning is performed. The Adam algorithm in the... The update process of step is as follows ( ): ; ; ; ; ; in This is the gradient vector of the composite loss function with respect to the network parameters; For the first The gradient vector of the step; This is for the estimation of the first moment of the gradient; This is for estimating the second moment of the gradient; This represents element-wise multiplication; This is the first-order moment estimate after bias correction; This is the second-order moment estimate after bias correction; For the first The network parameter vector is calculated step by step. The maximum number of iterations in the fine-tuning phase is 500 to obtain the final optimal parameters. .

[0047] Leave-one-out cross-validation was used to evaluate the model's generalization performance. For the 21 training samples, the first... One sample is used as the validation set, and the remaining 20 are used as the training set. Repeat this complete training process (with the hyperparameters fixed at 0). , , (only two-stage training is performed) to obtain the target for the first Optimal parameters for the second verification .

[0048] Standardized predictions are obtained by inputting validation samples. Performing inverse standardization yields predicted values ​​for physical quantities: ; Calculate the root mean square error and the mean absolute error to obtain the global average. and .like Miqie The optimal parameters will be obtained based on all 21 samples trained. Solidification, together with standardized parameters, forms the prediction model. Predictive models Once solidified, it will be able to accept input parameters from any given environment. The ability to output six dynamic safety clearance predictions will be invoked in S5 and S6.

[0049] S4: Solving for the vertical angle of the ship-hook arm geometry.

[0050] The geometric analysis and limit horizontal angle solution process is as follows: Figure 3 As shown. The inherent geometric parameters of the trailing suction hopper dredger and its boom include: beam. Rice, drinking water Meters, length of bend meters, length of the upper rake arm Meters, distance from the lifting point of the upper rake arm to the horizontal hinge Meters, length of the harrow arm Meters, distance from the lower hoist arm lifting point to the universal joint Meters, distance from the mud compartment inlet to the bottom of the ship Meters, horizontal distance from the rake arm bend shaft to the side of the hull Meters, horizontal distance from the center of the universal joint to the rubber pad Meters, vertical distance from the center of the universal joint to the rubber pad Meters, horizontal distance from the center of the rake head to the rubber pad Meters, height from the center of the rake head to the rubber pad Meters, vertical distance from the center of the rake head to the water nozzle Meters, the horizontal distance of the A-frame extending from the middle tube. Meters, the horizontal distance of the A-frame of the rake head extending out. Meters. Additionally, define the vertical distance of the suspension point eccentricity. Meters, this distance is the offset of the upper and lower drag arm lifting points from the drag arm centerline along a direction perpendicular to the drag arm axis. The above values ​​are specific to the Xin Haihun vessel. For other drag suction dredging vessel types, the following parameters need to be obtained from the actual hull and drag arm design drawings: hull width drinking water Length of bend Length of upper rake arm length of the lower rake arm Distance from mud compartment inlet to bottom of ship Horizontal distance from the rake arm bend shaft to the side of the ship Horizontal distance from the center of the universal joint to the rubber pad and vertical distance Horizontal distance from the center of the rake head to the rubber pad and height Vertical distance from the center of the rake head to the water nozzle The horizontal distance of the A-frame extending from the middle tube The horizontal distance of the A-frame of the rake head extending outwards. And adjust the parameter values ​​in the above formula accordingly. Mechanical limit angle of the rake arm system. For the new sea sturgeon vessel to be taken as This angle is determined by the physical interference position between the rake arm and the hull structure.

[0051] Intermediate pipe A-frame suspension point height The calculation formula is: ; in, m is the height of the A-frame base of the intermediate tube. m is the length of the A-frame strut in the middle tube. m is the horizontal distance from the midship longitudinal section to the base of the intermediate tube A-frame; A-frame hoisting point height of the rake head The calculation formula is: ; in, m is the height of the A-frame base of the rake head. m is the length of the A-frame support rod of the rake head. m is the horizontal distance from the midship longitudinal section to the base of the A-frame of the rake head.

[0052] Special attention should be paid to the above: , , , , , These are the specific design values ​​for the Xin Haihun vessel. For other trailing suction hopper dredging vessel types, the above parameters should be replaced with the actual hull arrangement drawings and A-frame structure drawings, while the rest of the calculation logic remains unchanged.

[0053] Obtain adjustable parameters for the operation, including: excavation depth. Preset minimum safety distance Safety angle of rubber hose Optimal vertical angle of the lower rake arm .

[0054] Determine the vertical angle of the upper harrow arm based on the target digging depth and the geometric conditions of the harrow head touching the bottom. perpendicular angle to the lower rake arm First, let's assume the vertical angle of the lower rake arm is at its optimal value. Vertical angle of the upper rake arm Determined by the geometric equations governing the contact between the rake head nozzle and the seabed: ; The vertical angle of the upper rake arm can be directly solved from the above formula: ; like The vertical angle is then determined as and .like The vertical angles of the upper and lower rake arms are limited by the maximum safe angle of the rubber hose, requiring the simultaneous solution of the following system of two nonlinear equations: ; The above system of two nonlinear equations is solved using the Newton-Raphson iterative method. Define the function vector. ,in , .Will Substitute into the bottoming geometry equation To provide a reasonable initial value for the Newton-Raphson iteration, the above-mentioned bivariate nonlinear equation system is dimensionally reduced. Substituting into the bottoming geometry equation, we obtain about The single-variable transcendental equation. To quickly obtain an initial guess for numerical solution, the following simplified formula is used for coarse estimation (this coarse estimate is only used to start the iteration; subsequent Newton-Raphson iterations will converge precisely to the true solution, and the coarse estimation error does not affect the final solution accuracy): ; get Afterwards, order This constitutes the initial vector for Newton's iteration.

[0055] Starting from this initial conjecture, iterative solutions are obtained. The Jacobian matrix The elements are , , , Iterate to Alternatively, if the maximum number of iterations (100) is reached, the result of the last iteration is taken as an approximate solution. If convergence is not achieved even after reaching the maximum number of iterations... If the error occurs, a warning message will be output and a binary search will be performed again within the feasible region.

[0056] Calculate the height of the midpoint of the universal joint Height of the rake head rubber pad : ; ; According to the excavation depth The operating conditions are divided into three intervals. The first interval is the deep-water operating condition, and the judgment condition is... and At this point, both the universal joint rubber pad and the shovel head rubber pad are below the hull bottom, and the critical components are the shovel center sling and the shovel head sling. The second interval is a medium water depth operating condition, and the judgment condition is... and At this point, the universal joint rubber pad is above or flush with the hull bottom, but the rake head rubber pad is below the hull bottom. The critical components are the universal joint rubber pad and the rake head sling. The third section is a shallow water condition, and the judgment condition is... and At this point, the universal joint rubber pad and the rake head rubber pad are both located above or flush with the bottom of the ship, and the critical components are the universal joint rubber pad and the rake head rubber pad.

[0057] S5: Construct constraint equations and use the homotopy extension iteration method to solve for the limiting horizontal angle.

[0058] Vertical angle determined by S4 and The risk-critical component judgment result is used as input, and the horizontal angle of the upper rake arm is respectively... and the horizontal angle of the lower rake arm Nonlinear constraint equations are constructed and solved using the homotopy extension iterative method. Simultaneously, the prediction model in S3 is... The output dynamic safety distance prediction value is used as the modulation parameter for the initial solution of the auxiliary equation in the homotopy extension method. The specific modulation sequence is as follows: S3 completes the prediction model. After curing, first input the current environmental parameters in S5. Obtain the minimum safe distance between the rake arms Minimum safe distance from the rake head Subsequently, the initial angle estimates of the homotopy extension method are obtained through the inverse geometric mapping of the upper and lower rake arms, respectively. and Finally, homotopy extension iterations are performed to obtain the limiting horizontal angles of the upper and lower rake arms.

[0059] For the extreme horizontal angle of the upper rake arm To solve this problem, first calculate the horizontal distance from the midpoint of the universal joint to the side of the hull. The height of the midpoint of the universal joint. The value calculated in S4 is used and will not be calculated again. The calculation expression is: ; in The horizontal distance from the rake arm bend shaft to the side of the hull, as defined in S4. The length of the upper rake arm is defined in S4.

[0060] Calculate the coordinates of the upper rake arm lifting point : ; ; in The distance from the upper rake arm lifting point to the horizontal hinge, as defined in S4; The meter is the vertical distance of the eccentricity of the suspension point as defined in S4.

[0061] The hoisting cable passes through the A-frame lifting point of the intermediate pipe. and the upper rake arm lifting point The nearest point on the side is taken as Minimum distance from the hoisting cable to the hull. The distance from a point to a line is calculated using the formula: ; in and The coordinates of the A-frame suspension points for the intermediate tube defined in S4.

[0062] Based on the working condition type determined by S4, the constraint equations for the upper rake arm are constructed respectively. and lower rake arm constraint equations .in This represents the final value of the limiting horizontal angle of the upper drag arm, which has already been solved. For the upper drag arm, if the current working condition is deep water and the universal joint rubber pad is below the bottom of the ship, the upper drag arm uses the sling constraint equation. If the current operating condition is in medium or shallow water, and the universal joint rubber pad is located above the bottom of the vessel, the upper rake arm uses the rubber pad constraint equation. For the lowered hook arm, if the current operating condition is deep or medium water depth, and the hook head rubber pad is below the bottom of the vessel, the lowered hook arm uses the sling constraint equation. If the current operating condition is shallow water, and the rake head rubber pad is above the bottom of the boat, the lower rake arm uses the rubber pad constraint equation. .

[0063] Before solving the equation, first determine its condition. exist Does the solution exist in the given condition? Calculate. Minimum value on the interval and maximum value .like If the equation has a solution, then the homotopy extension method can be continued. If This indicates that even when the horizontal angle of the rake arm is adjusted to the mechanical limit position... The distance between the sling and the hull is still greater than the preset safety clearance. At this time The report is and output the current preset minimum safety distance. The maximum safe clearance of the rake arm under mechanical limit is greater than the limit horizontal angle reported as the mechanical limit angle. Tip. If This indicates that the distance from the sling to the hull is less than [a certain value] at any horizontal angle. This means that the safety distance requirement cannot be met under the current working conditions, so work is stopped and a warning that the safety distance cannot be reached is issued. Simultaneously, the homotopy continuation method is applied... The last convergence node corresponding angle As a suggestion to avoid directional output, the rake operator is advised to adjust the rake arm posture in the direction that reduces the absolute value of the horizontal angle.

[0064] The above nonlinear equations are solved using the homotopy extension iterative method. Construct homotopy functions , the objective equation With auxiliary equations Perform homotopy join: ; in For homotopy parameters; This is the estimated initial angle of the upper rake arm, in degrees. This initial angle estimate is derived from the prediction model in S3. The output dynamic safety clearance is obtained through inverse geometric mapping. During inverse geometric mapping, the corresponding constraint function must be selected based on the working condition type determined by S4. For deep-water conditions, cable constraint equations are used; for shallow or medium-depth conditions, rubber pad constraint equations are used. The prediction model will be... Minimum safe distance of the output rake arm Use the target distance value to be matched. Define the inverse mapping function. To solve the equation The root is found. A strategy combining global search and local binary search is employed: the interval is... Divided into equal parts There are several sub-intervals, with a step size of [number]. ,in This is the floor function. When hour , Calculate the residuals at the endpoints of each sub-interval. If the residual at any subinterval endpoint Then directly take that endpoint. for Starting from the left end of the interval, check adjacent endpoints sequentially to find the first one that satisfies the condition. subintervals Set the left endpoint of the sub-interval as The right endpoint is set to The initial search interval for the bisection method is set as follows: If no pair of endpoints with opposite signs exists after traversing all sub-intervals, the endpoint with the smallest absolute residual value is taken as the approximate solution, and the bisection method is not executed again. If the residuals of all sub-interval endpoints have the same sign, the endpoint with the smallest absolute residual value is taken as the approximate solution. The iterative steps of the bisection method are: initially, the left boundary of the search interval is set. and right boundary The endpoints of the sub-intervals with residual differences. In the... The midpoint is calculated in the next iteration. Calculate the residual .like Then it converges; if and If the signs are different, then let Otherwise The iterative convergence criterion is: The degree may reach the maximum number of iterations (100) and converge. That is This strategy does not rely on the global monotonicity assumption, making it more robust.

[0065] Number of subintervals in homotopy extension Determined through preliminary experiments. Tests were conducted under typical operating conditions. , , , , Five configurations were used to record the convergence success rate and average computation time of the homotopy extension method. When... The convergence success rate was 94%. The convergence success rate reached 99%, and the average computation time was approximately 0.8 seconds. The convergence success rate is 99.5%, but the average computation time increases to 1.6 seconds. Considering both convergence reliability and real-time computation requirements, [the following option is selected]. As the optimal value.

[0066] The 1% failure rate in the preliminary experiment all occurred in exist In cases where the path has multiple local extrema, the homotopy extension method may get trapped in local optima and fail to converge. For such cases, after the standard homotopy extension method fails (i.e., Newton's iteration fails in...), a solution is needed. The condition is still not satisfied after 100 iterations. Miqie (Any convergence condition in radians) automatically switches to a global fine-grained search strategy: [The interval is then used for further analysis]. by A dense scan is performed with a step size consistent with the measurement accuracy of the shipborne angle sensor to ensure that the scan resolution of the backup strategy matches the sensor accuracy. The residual of each scan point is calculated, and the point with the smallest absolute residual value is taken as the approximate solution. The computation time of this backup strategy is about three times that of the standard homotopy continuation method (about 2.4 seconds), but it is only triggered when the standard method fails, so the impact on overall real-time performance is acceptable.

[0067] Hotopism parameter range Divided into equal parts The number of subintervals and the node sequence satisfy the following conditions: ,in Step length That is, the node sequence is For the current homotopy parameter node , The solution is obtained using Newton's iteration method. : ; in For the first The angle value after one Newton iteration, in degrees; ; right The partial derivatives are: ; Let be the derivative of the constraint equation with respect to the horizontal angle. For the rubber pad constraint case... Its derivative is For sling-constrained situations Its derivative is calculated numerically using the central difference method: ; Iterative convergence criterion is Meters (hotopy function value close to zero, corresponding to a distance error of less than 0.1 millimeters) or Convergence is defined as the change in radians (with sufficiently small angular variations) being satisfied with any of these conditions. Current homotopy parameter node. convergent solution As the next node Initial guess value: Proceeding sequentially to The objective equation is obtained. Solution .

[0068] The lower rake arm's limit horizontal angle The solution process is symmetrical to that of the upper rake arm. First, define... This is the estimated initial angle of the lower rake arm. Coordinates of the lower rake arm lifting point. for: ; ; in The final value of the limiting horizontal angle of the upper rake arm, obtained from the solution, is expressed in degrees. The distance from the lower rake arm lifting point to the universal joint, as defined in S4; This is the height value of the gimbal midpoint calculated in S4; The meter is the vertical distance of the eccentricity of the suspension point as defined in S4.

[0069] The rake head sling passes through the rake head A-frame lifting point. and the lower rake arm lifting point Minimum distance from the rake head sling to the hull for: ; in and The coordinates of the lifting points of the A-frame rake head defined in S4.

[0070] Based on the working condition type determined by S4, construct the lower rake arm constraint equation. If the current operating condition is in deep or medium water (the sling head rubber pad is below the hull bottom, and the risk lies in the slings), the constraint equations are: ; If the current operating condition is shallow water (the rake head rubber pad is located above the bottom of the ship, and the risk lies in the rubber pad), the constraint equations are: ; in This is the horizontal distance from the center of the rake head to the rubber pad, as defined in S4.

[0071] For the ultimate horizontal angle of the lower rake arm The solution is obtained by constructing the constraint equations. The same no-solution check is performed. Calculation. In the interval minimum value on and maximum value Processed according to the same three-stage logic as the upper rake arm: If Then there is a solution, and we continue with the homotopy extension method; if Then The report is And output the corresponding prompt; if The output will show a warning that the safety clearance cannot be reached, and the homotopy continuation method will be applied. The last convergence node corresponding angle This is an output that suggests ways to avoid certain directions.

[0072] Initial angle estimate of the lower rake arm Obtained through inverse geometric mapping. In the inverse geometric mapping of the lower rake arm, the prediction model is... Minimum safe distance of the output rake head The target distance value to be matched is used as the lower rake arm constraint equation. Construction residuals Using the same global search and local binary search strategy as the upper rake arm, the initial angle estimate is obtained. When a solution exists, the same homotopy extension iteration method as the upper rake arm is used for solving. The homotopy function of the lower rake arm... The expression is: ; The same Newton-Raphson iteration method as the upper rake arm is used to solve the problem. ,get .

[0073] S6: Construction of safety criteria for the fusion of dynamic prediction and static limits and closed-loop feedback.

[0074] Integrating security criteria and closed-loop feedback processes, such as Figure 4 As shown. Predictive model for S3 solidification. The output dynamic safety clearance prediction value is fused with the static limit horizontal angle solved by S5 to construct a unified real-time safety criterion system, and a weighting coefficient from the safety criterion to the S2 physical constraint is established. and A closed-loop feedback mechanism.

[0075] For the current operating conditions, input environmental parameters. The prediction model using S3 After performing dynamic prediction, standardization, forward propagation, and destandardization, six dynamic safety distance prediction values ​​are obtained: ; The static limit horizontal angle obtained by solving S5 and Substituting back into the geometric relationship of S4, calculate the corresponding static safety distance under this extreme attitude. and Because the constraint equations in S5 are based on the preset minimum safety clearance. Established for the threshold, therefore and .

[0076] Define the minimum prediction spacing Define the safety margin for fusion. The calculation of the fusion safety margin adopts a piecewise approach. When the minimum predicted spacing... At the same time, integrate safety margin Take directly as Preserve information about the degree of negative values, so that Rice meters. When predicting the minimum spacing. At the same time, integrate safety margin Defined as the weighted geometric mean of the dynamically predicted minimum spacing and the static limit spacing, minus the safety spacing threshold: ; in To dynamically predict the minimum spacing, To predetermine the minimum safety distance, this scheme uses the geometric mean instead of the arithmetic mean or harmonic mean for the following reasons: *Geometric Mean Having physical boundary consistency—when hour, It precisely corresponds to the safety boundary state (the dynamic prediction spacing is exactly equal to the safety threshold); when hour, This precisely corresponds to the extreme danger state (when the safety margin is completely exhausted). Within the interval, the geometric mean is generated It exhibits a non-linear, accelerating decay trend, consistent with the physical law that the safety margin is rapidly depleted as the spacing decreases. (Arithmetic mean) exist Only at time It cannot reflect the extreme danger state, and at the boundary it is incompatible with... The dimensional relationships are inconsistent; the harmonic mean, on the other hand, is in... This can produce singular values ​​and is not suitable for this scenario.

[0077] when At that time, the current working condition is determined to be safe, and both dynamic risk and static geometry meet the safety requirements. At that time, it was determined that there was a risk of collision.

[0078] Establish from the perspective of security margin Weight coefficients for physical constraints and The adaptive closed-loop feedback mechanism. Due to and They work together within physical constraints, employing a unified adjustment method to... and Synchronous scaling. Define the scaling factor. : ; in To adjust the sensitivity coefficients, the updated first-order and second-order physical constraint weights are as follows: ; when hour Physical constraints are strengthened; when When it is a large positive value Physical constraints are weakened.

[0079] set up The effective range is , The effective range is When scaled or When exceeding the corresponding boundary, truncation is performed to the boundary value. Simultaneously, the fusion safety margin for 10 consecutive operating conditions is considered. At that time, and Reset to initial value , This resets the continuous deviation counter used to trigger incremental training to prevent strong physical constraints accumulated from historical adverse operating conditions from persisting in subsequent safe operating conditions and affecting the model's ability to fit data details. The 30% threshold corresponds to a safety margin of 30% of the preset safety interval. At this point, the safety margin not only meets the safety requirements of the current operating condition but also provides sufficient redundancy to cope with short-term environmental fluctuations (such as gusts of wind, sudden surges, etc.). It can be determined that the operating condition has been continuously stable and in a safe state. Set a margin of 3 times that of the green safety zone boundary (20%) to avoid unnecessary weight resets triggered by small fluctuations in environmental parameters near the threshold.

[0080] Updated physical constraint weight coefficients and This is fed back into the composite loss function definition of S2. When new field-measured data is collected, or when the predictive model of S3 is found to be... When the deviation between the predicted value and the actual monitored value exceeds a preset threshold, an incremental training process is triggered. The on-site measured data comes from a sensor system installed on the ship, specifically including: an ultrasonic distance sensor (sampling frequency 10Hz, range 0~5m, accuracy ±0.01m) installed near the connection point between the rake arm and the rake head, which collects the distance between the rake arm and the hull and the distance between the rake head and the hull in real time. The raw sampled data is first preprocessed by a 5-point moving average filter to eliminate high-frequency wave noise, and then averaged every 10 seconds to form measured sample pairs for comparison with the model's predicted values. The preset threshold is the mean absolute error. The threshold is determined based on the mean absolute error of leave-one-out cross-validation in S3. The meter value, approximately six times the actual value, is sufficient to exclude normal prediction fluctuations. Incremental training is triggered when the average absolute error between the predicted and measured values ​​for five consecutive operating conditions exceeds 0.05 meters. After incremental training is triggered, the consecutive deviation count is reset to zero, and accumulation begins again. This is done using the current optimal parameters. Using the initial values, the Adam optimization algorithm described in S3 is applied to the updated physical constraint weight coefficients. and The neural network undergoes 50 fine-tuning steps to gradually adapt to the new operating condition data distribution while maintaining its original generalization ability. After incremental training, the new model enters a stabilization period, defined as the period during which the mean absolute error (MAE) of predictions for 20 consecutive operating conditions is below MAE_threshold = 0.05 meters. During the stabilization period, consecutive out-of-tolerance operating conditions are still counted, but incremental training is not triggered. When all 20 consecutive operating conditions are below the threshold, the stabilization period ends. If the accumulated consecutive out-of-tolerance count has reached 5, incremental training is immediately triggered and the counter is reset to zero; if the accumulated out-of-tolerance count is less than 5, it continues to accumulate until it reaches 5 before triggering incremental training. During the stabilization period, consecutive out-of-tolerance operating conditions are still counted, but incremental training is not triggered.

[0081] All the above information is integrated into operating commands that can be directly received by the driver's console. These operating commands include: dynamic safety clearance prediction values. Displayed in numerical and color-coded form on the software interface, when or A red warning is displayed when the condition is active; otherwise, a green safe status is displayed. Static limit horizontal angle. and The red warning line displayed on the angle gauge indicates the real-time angle from the ship's onboard angle sensor. and The indicator is displayed as a pointer, and an audible and visual alarm is triggered when the pointer touches the red warning line.

[0082] Integrating safety margin Presented as a single, comprehensive health index. Health Index The calculation formula is: ; In the health formula The operation sets the minimum health level to 0% to represent the physical state of zero safety margin. When At that time, health was truncated to 0%, and 0% here has two physical meanings: Critical state and The dangerous state. The key to distinguishing between the two states lies in The symbol refers to the health level itself, not a numerical health value. Health levels range from 0% to 5%. In response to the emergency rake lifting command, This corresponds to a red danger signal that necessitates a forced stop. Therefore, the triggering conditions for a forced stop must simultaneously include... and Two sub-conditions.

[0083] The health index employs a multi-level early warning system. The operational meanings of each instruction are as follows: Green corresponds to safety. With ample safety margin, no warning will be issued; a yellow warning indicates a need to lift the rake accordingly. The safety margin is 10% to 20% requires the operator to complete the preparation and operation of lifting the harrow within 60 seconds; orange indicates immediate harrow lifting. The safety margin is 5% to 10%, requiring the operator to begin winch operation within 10 seconds; this time window is sufficient to complete the standard operating procedure from identification to winch activation; emergency rake lifting corresponds to... and Insufficient safety margin 5% requires the operator to start the winch within 3 seconds. The time window only allows the simplest operational sequence: directly pressing the emergency stop or rake lifting button. No parameter adjustments or secondary confirmation are permitted. A red forced stop corresponds to... and Output a command to force the operation to stop.

[0084] The 20% threshold is set based on simulation data from 10 typical crossflow conditions (flow velocity 0.5–1.5 m / s). The lateral drift rate of the rake arm is 0.02–0.05 m / s, and the average reaction time for the operator from warning recognition to winch activation is 22 seconds, with a maximum of 35 seconds. Within 35 seconds, the maximum rake arm drift is approximately 1.75 meters. Considering the suppressive effect of the rake lifting operation, the required warning safety margin is approximately... ( The accuracy is approximately 0.2 meters (equivalent to a health level of 20%). A 10% threshold corresponds to a safety margin of 0.1 meters, which is approximately the measurement accuracy of the shipborne angle sensor (approximately...). This is equivalent to 2 to 3 times the horizontal distance (approximately 0.03 to 0.05 meters), ensuring that the early warning trigger is not dominated by sensor noise. The 5% threshold corresponds to a 0.05-meter safety margin, serving as the trigger line for emergency rake lifting, ensuring reliable triggering even within the sensor noise range.

[0085] In the control panel interface, the health value must be accompanied by a status text: When When the display shows 'critical - emergency rake lifting', when The system displays 'Danger - Forced Stop' to ensure that operators do not need to understand it. The meaning of the symbols indicates the correct execution of the operation instructions.

[0086] Example 1: Example 1 provides a specific implementation of a method for calculating the safe angle of the trailing suction hopper dredger based on neural networks and geometric analysis. The target trailing suction hopper dredger is the Xin Hai Xun, the operating area is the Yangtze River estuary channel, and the dredging depth is... Set to 12.5 meters, with a preset minimum safety distance. Set to 1.0 meter. Environmental parameter values ​​are: significant wave height. Meters, spectral peak period Seconds, crossflow velocity m / s, wave direction is 90°.

[0087] Perform a sequential experimental design based on the S1 physical perception D-optimal criterion. Generate 1000 candidate sampling points in the three-dimensional environmental parameter space. Use a sequential forward selection algorithm to filter the sample points. After selecting at least 10 samples, calculate the information coverage rate after each iteration. After 21 iterations A total of 21 sample points were selected. The 21 selected sets of input parameters were input into the time-domain coupled dynamic simulation model of the Xinhaihun vessel. Each working condition was simulated for 10,800 seconds in the time domain. The safe distance time history data in the interval of 600 seconds to 11,400 seconds was extracted, and 21 sets of input-output sample pairs were obtained through statistical analysis.

[0088] Perform data standardization and neural network construction as described in S2. Z-score standardization was applied to 21 sets of samples. A three-layer feedforward neural network was constructed, with 3 neurons in the input layer, 10 neurons in each of the two hidden layers, and 6 neurons in the output layer. The network weights were initialized using a Xavier normal distribution. A composite loss function containing first-order and second-order physical monotonicity constraints was defined.

[0089] Perform hyperparameter selection and two-stage training as described in S3. Hyperparameters are determined using a sequential selection strategy employing five-fold cross-validation. , , In the first stage, Levenberg-Marquardt optimization, an augmented residual vector was constructed, and the damping factor was initialized to 0.01. It converged after 387 iterations. In the second stage, Adam physics constraint fine-tuning, the learning rate was 0.001, and the loss function stably converged after 500 iterations. Leave-one-out cross-validation was performed, and the average root mean square error of 21 validations was... Meters, mean absolute error Meters, meeting the accuracy requirements. The optimal parameters obtained from training on all 21 samples will be used. Solidification, together with standardized parameters, forms the prediction model. .

[0090] Determine the vertical angle of the rake arm in S4. The digging depth is 12.5 meters. First, assume... Calculations revealed Therefore, a system of simultaneous equations needs to be solved. The solution obtained using the Newton-Raphson method is... , Calculate the height of the midpoint of the universal joint. rice, Meters > 0, If the depth is less than 0 meters, it is classified as a medium-depth working condition, and the critical components are the universal joint rubber pad and the rake head sling.

[0091] Solve the nonlinear equation for the limiting horizontal angle of S5. For the limiting horizontal angle of the upper rake arm, construct the rubber pad constraint equation: Solving for For the limiting horizontal angle of the lower rake arm, calculate... Meters, constructing cable constraint equations based on medium water depth conditions. By solving the homotopy extension iterative problem, we obtain .

[0092] Execute the S6 fusion security criteria and closed-loop feedback. Input the current environmental parameters. The predicted dynamic safety distance is obtained as follows: rice, Meters. Predicted minimum spacing. Meters > 0, with a safety margin. Meters, a collision risk has been identified. Health level and Output a command to force stop the job. Update the scaling factor. , Updated to 0.801. Updated to version 0.320, with enhanced physical constraints.

[0093] Example 2: Example 2 is based on Example 1, but with adjustments made to the crossflow velocity. The speed is meters per second, and all other parameters remain unchanged. The static calculation results for S1-S5 are exactly the same as those in Example 1. , .

[0094] Execute S6 to input environment parameters The dynamic predicted value is obtained: rice, rice. Meters < 0, Rice. Health and Output a forced stop command. Update the scaling factor. , Updated to 0.589. Updated to 0.235. Compared with Example 1, the increased flow velocity leads to a sharp reduction in the dynamic safety distance, making the safety criteria more stringent.

[0095] Example 3: Example 3 adjusts the excavation depth based on Example 1. The environmental parameters are the same as in Example 2. In S4, it is first assumed that... Seeking , The requirements are met. Calculation. rice, Meters < 0, If the distance is less than 0 meters, it is determined to be a deep-water working condition, and the critical risk components are the middle hoisting cable and the head hoisting cable.

[0096] In S5, the structural cable constraint equations under deep-water conditions The solution obtained through homotopy extension is The constraint solution of the lower rake arm sling is obtained. .

[0097] In S6, the dynamically predicted value is: rice, rice. Meters > 0, Rice. Health The corresponding orange alert should be immediately raised. Update the scaling factor. , Updated to 0.482. Updated to 0.193. A comparison of Example 3 and Example 2 shows that as the excavation depth increases, the critical risk components shift, the static limit horizontal angle increases, the dynamic safety margin improves, and the warning level decreases from mandatory stop to immediate harrowing, achieving differentiated decision-making.

[0098] Comparative Example 1: Comparative Example 1 provides a traditional evaluation method based on a fixed angle threshold. A safe horizontal angle threshold for the upper rake arm is set. Lower rake arm .

[0099] Under the operating condition of Example 1, the measured angle is approximately and None of the values ​​exceeded the threshold, and traditional methods determine safety, while the method of this invention forces the output to stop. This reduces the risk of underreporting compared to the comparison.

[0100] Under the operating condition of Example 2, the measured angle is approximately and If the flow rate is still below the threshold, the traditional method determines it is safe, but the method of this invention outputs a forced stop again. The traditional method misses the report because it does not consider the attenuation of the dynamic spacing due to the flow rate.

[0101] Under the operating condition of Example 3, the measured angle is approximately... and If the safety margin is below a certain threshold, traditional methods determine it as safe, while the method of this invention outputs an immediate warning. Traditional methods cannot identify critical states with insufficient safety margins and cannot provide quantitative suggestions.

Claims

1. A method for calculating the safety angle of the trailing suction hopper dredger's boom based on neural networks and geometric analysis, characterized in that, Includes the following steps: S1: Obtain environmental parameters of the sea area where the trailing suction hopper dredging vessel is operating, construct a candidate sample pool based on the range of variation of environmental parameters and select the optimal training sample set, then input it into the time-domain coupled dynamic simulation model for simulation, extract the safety distance statistics from the simulation results as output parameters, and form a training dataset. S2: Construct a physical information neural network, perform standardization processing on the training dataset, and train the physical information neural network using the standardized training dataset and preset physical constraints to obtain the trained physical information neural network. S3: Perform cross-validation on the trained physical information neural network to solidify the physical information neural network that meets the accuracy requirements into a prediction model; S4: Obtain the inherent geometric parameters and adjustable parameters of the trailing suction hopper dredging vessel, establish the geometric constraint relationship between the drag arm system and the hull based on the inherent geometric parameters and adjustable parameters, solve the vertical angle of the drag arm system according to the geometric constraint relationship, and determine the water depth range to which the current working condition belongs based on the vertical angle to identify the critical risk components. S5: Input environmental parameters into the prediction model and output the predicted dynamic safety distance between the rake arm system and the hull. Based on the predicted dynamic safety distance, critical risk components, vertical angle and preset minimum safety distance in the adjustable operation parameters, construct the rake arm horizontal angle constraint equation and solve it using the homotopy extension iteration method to obtain the static limit horizontal angle. S6: The predicted dynamic safety distance is fused with the static limit horizontal angle to generate a fused safety criterion. The safety angle of the rake arm is output based on the fused safety criterion, and the weight coefficients of the physical constraints are adjusted according to the fused safety criterion. The results are then fed back to S2 for subsequent incremental training.

2. The method for calculating the safety angle of the trailing suction hopper dredger arm based on neural networks and geometric analysis according to claim 1, characterized in that, The environmental parameters in S1 include at least meaningful wave height, spectral peak period, and crossflow velocity. The specific process of constructing a candidate sample pool and selecting the optimal training sample set is as follows: construct a three-dimensional environmental parameter space covering the range of environmental parameter variations, discretize the three-dimensional environmental parameter space along three dimensions into multiple candidate sampling points to form a candidate sample pool, and use the D-optimal criterion to select sampling points from the candidate sample pool whose information coverage meets the preset threshold to form the optimal training sample set.

3. The method for calculating the safety angle of the trailing suction hopper dredger arm based on neural networks and geometric analysis according to claim 2, characterized in that, The preset physical constraints in S2 include first-order physical monotonicity constraints and second-order physical monotonicity constraints. The first-order physical monotonicity constraints are used to constrain the statistical value of the safety clearance to decrease monotonically as the meaningful wave height, spectral peak period, and crossflow velocity increase. The second-order physical monotonicity constraints are used to constrain the second-order difference sign of the statistical value of the safety clearance to conform to the physical law of downward convex monotonically decreasing as the crossflow velocity increases.

4. The method for calculating the safety angle of the trailing suction hopper dredger arm based on neural networks and geometric analysis according to claim 1, characterized in that, The specific process of training the physical information neural network in S2 is as follows: a composite loss function is constructed to constrain the physical information neural network. The composite loss function is composed of a weighted average of a data fitting loss term, a Bayesian regularization loss term, a first-order physical monotonicity constraint loss term, and a second-order physical monotonicity constraint loss term. A two-stage training strategy of cascading optimization of the Levenberg-Marquardt algorithm and the Adam algorithm is used to train the physical information neural network. In the first stage, supervised learning is performed using the data fitting loss term and the Bayesian regularization loss term. In the second stage, the complete composite loss function is used for fine-tuning of physical constraints.

5. The method for calculating the safety angle of the trailing suction hopper dredger arm based on neural networks and geometric analysis according to claim 1, characterized in that, In S3, cross-validation is leave-one-out cross-validation. Each sample in the training dataset is used as the validation set, and the remaining samples are used as the training set. The training process of the physical information neural network is repeated for each validation. The prediction error of the validation set is calculated. If the global average error of all validation sets meets the preset accuracy threshold, the physical information neural network trained based on all training samples is solidified as the prediction model.

6. The method for calculating the safety angle of the trailing suction hopper dredger arm based on neural networks and geometric analysis according to claim 1, characterized in that, The specific process of solving the vertical angle of the rake arm system based on geometric constraints in S4 is as follows: Solve the vertical angles of the upper and lower rake arms based on the dredging depth and the rake head's contact with the seabed geometry; calculate the height of the universal joint midpoint and the height of the rake head rubber pad based on the vertical angles of the upper and lower rake arms; determine the water depth range to which the current working condition belongs based on the height of the universal joint midpoint and the height of the rake head rubber pad; the specific process of solving the vertical angles of the upper and lower rake arms based on the dredging depth and the rake head's contact with the seabed geometry is as follows: Calculate the target depth of the rake head nozzle based on the dredging depth, the ship's draft, and the distance from the mud hopper inlet to the bottom of the ship; establish the geometric equation for the contact between the rake head nozzle and the seabed based on the lengths of each section of the rake arm, the vertical distance from the rake head nozzle to the rake head connection point, and the preset optimal vertical angle of the lower rake arm; solve the geometric equation using the Newton-Raphson iterative method to obtain the vertical angles of the upper and lower rake arms.

7. The method for calculating the safety angle of the trailing suction hopper dredger arm based on neural networks and geometric analysis according to claim 1, characterized in that, The S4 water depth range includes deep water conditions, medium water depth conditions, and shallow water conditions; the critical risk components include the harrow center sling, harrow head sling, universal joint rubber pad, and harrow head rubber pad; in deep water conditions, the critical risk components are the harrow center sling and harrow head sling; in medium water depth conditions, the critical risk components are the universal joint rubber pad and harrow head sling; in shallow water conditions, the critical risk components are the universal joint rubber pad and harrow head rubber pad.

8. The method for calculating the safety angle of the trailing suction hopper dredger arm based on neural networks and geometric analysis according to claim 1, characterized in that, The specific process of obtaining the static limit horizontal angle after solving the rake arm horizontal angle constraint equation in S5 is as follows: take the minimum safe distance of the rake arm in the dynamic safe distance prediction value as the target distance value, perform inverse geometric mapping based on the rake arm horizontal angle constraint equation to obtain the initial angle estimate value of the homotopy extension method; construct the homotopy function to connect the rake arm horizontal angle constraint equation and the auxiliary equation, wherein the auxiliary equation is solved with the initial angle estimate value. The homotopy parameter interval is discretized into multiple sub-intervals. Newton's iteration method is applied to each sub-interval to solve the homotopy function. The homotopy parameter is gradually advanced until convergence, and the upper and lower rake arm limit horizontal angles are obtained as static limit horizontal angles.

9. The method for calculating the safety angle of the trailing suction hopper boom based on neural networks and geometric analysis according to claim 1, characterized in that, The specific process of generating the fusion safety criterion in S6 is as follows: calculate the static safety distance based on the static limit horizontal angle; define the predicted minimum distance as the smaller value between the minimum safety distance of the rake arm and the minimum safety distance of the rake head in the predicted dynamic safety distance; when the predicted minimum distance is less than zero, the fusion safety margin is directly taken as the predicted minimum distance; when the predicted minimum distance is greater than or equal to zero, the fusion safety margin is defined as the weighted geometric average of the predicted minimum distance and the static safety distance minus the preset minimum safety distance. Fusion security criteria are generated based on fusion security margin.

10. The method for calculating the safety angle of the trailing suction hopper boom based on neural networks and geometric analysis according to claim 1, characterized in that, S6 also includes: calculating a health index based on the fusion safety margin. The health index is calculated by converting the ratio of the fusion safety margin to the preset minimum safety distance into a percentage. Based on the numerical range of the health index, corresponding warning instructions are triggered. The warning instructions include five levels: green (safe), yellow (suggest lifting the rake), orange (immediate lifting of the rake), emergency lifting of the rake, and red (forced stop).