Deep-sea mining hoisting system coupling motion prediction method, system, medium and product based on physical operator and multi-coding

Abstract

The application discloses a deep-sea mining lifting system coupling motion prediction method and system based on physical operators and multi-coding, a medium and a product. First, a high-fidelity dynamic model of the system is established, and multi-working-condition motion response data is obtained as a training set. In the Transform network, a trainable second-order dynamic physical operator is introduced, the environmental working condition and continuous time are coded into feature input encoders, and the output is a key cross-section bending moment and an equivalent excitation prediction sequence. By calculating the physical consistency residual and the data residual, a multi-domain loss function containing time domain error, frequency domain error, dynamic residual and spatial smoothing constraint is designed. The model parameters and the physical operator are jointly optimized to obtain a trained model, and the target working condition is input to output the prediction result. The application considers the prediction accuracy and the dynamic characteristic description, improves the prediction stability and the physical interpretability, and is suitable for deep-sea mining system design verification, operation monitoring and risk warning scenes.

Description

Technical Field

[0001] This invention belongs to the field of dynamic modeling and intelligent prediction technology for deep-sea mining equipment, specifically involving a coupled motion prediction method, system, medium, and product for deep-sea mining hoisting systems based on physical operators and multi-coding. Background Technology

[0002] Deep-sea mining hoisting systems typically consist of multiple subsystems, including surface support vessels, ultra-long hoisting risers, relay stations, flexible hoses, and mining vehicles. Under the combined effects of complex marine environmental loads and operational disturbances, they exhibit significant multi-scale, strongly coupled, and non-stationary dynamic characteristics. Under random environmental loads such as wind, waves, and currents, as well as the transport of internal slurry, the system is prone to significant longitudinal, lateral, and bending coupled motion responses. The bending moment response along the hoisting riser is a crucial indicator for structural safety assessment; excessive local bending can lead to fatigue or buckling failure.

[0003] Existing high-fidelity numerical simulation methods based on finite element method (FEM) or multibody dynamics can accurately describe the dynamic behavior of systems, but their modeling process is complex and computationally expensive, making it difficult to meet the needs of rapid prediction and online monitoring. Deep learning prediction methods, which have emerged in recent years, have advantages in computational efficiency by learning from historical response data; however, most methods do not explicitly incorporate physical constraints of system dynamics, easily leading to problems such as prediction results violating physical laws, insufficient generalization ability, and strong dependence on training data. Especially at critical locations such as the ship-riser interface, where the response is dominated by six-degree-of-freedom ship motion and exhibits high non-stationarity, traditional neural network methods struggle to achieve acceptable prediction accuracy. Therefore, there is an urgent need for a coupled motion prediction method that integrates deep learning modeling capabilities with physical constraints of system dynamics, to improve the physical consistency and engineering reliability of prediction results while ensuring prediction accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide a method, system, medium, and product for predicting coupled motion of deep-sea mining hoisting systems based on physical operators and multi-encoding. By introducing trainable dynamic physical operators into deep learning models, the invention achieves collaborative modeling of data-driven prediction and system dynamic mechanism constraints, thereby improving the accuracy, stability, and physical interpretability of coupled motion prediction for deep-sea mining hoisting systems.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A motion prediction method for deep-sea mining hoisting systems based on physical operators and multi-encoding includes the following steps:

[0007] Step 1: Establish a dynamic model of the lifting system, obtain time-series data of coupled motion response under multiple working conditions, and construct a training set;

[0008] Step 2: Introduce a trainable second-order dynamic physics operator into the output segment of the Transformer sequence prediction network to build a prediction model; encode the environmental conditions and continuous time position as features, input them into the encoder of the Transformer sequence prediction network, and output the predicted sequence of bending moment and equivalent excitation of key monitoring sections.

[0009] Step 3: Calculate the physical consistency residual of the predicted sequence based on the second-order dynamic physical operator;

[0010] Step 4: Construct a multi-domain loss function that includes time-domain prediction error, frequency-domain feature error, dynamic residuals, and spatial smoothing constraints;

[0011] Step 5: Using the training set, with the goal of minimizing the multi-domain loss function, jointly optimize the encoder parameters and trainable second-order dynamic physical operators in the prediction model to obtain the trained prediction model.

[0012] Step 6: Input the target working condition parameters into the trained prediction model and output the prediction results of the coupled motion of the lifting system.

[0013] Furthermore, the lifting system includes a surface support vessel, a lifting riser, a booster pump, a relay station, a flexible hose, and a mining vehicle; the dynamic model is modeled using the lumped mass method, the hydrodynamic loads are calculated using the Morison formula, and the environmental conditions include wave, current, and wind loads.

[0014] Furthermore, the environmental operating condition coding includes encoding the meaningful waveform height. Spectral peak period Ship speed The direction angle of waves, wind, and ocean waves Min-Max normalization is performed and projected to the conditional feature vector obtained at time step. ;

[0015] The continuous time position encoding uses a method based on a family of sine functions to generate time feature vectors. ,

[0016] The conditional feature vector is added to the temporal code and used as the input to the Transformer encoder:

[0017]

[0018] Stack the input vectors from all time steps in chronological order to form the input sequence. .

[0019] Furthermore, the Transformer encoder is characterized by employing a sequence coding structure based on a multi-head self-attention mechanism.

[0020] For the first Layer input sequence Projected as a query ,key Sum Calculate the scaled dot product attention:

[0021]

[0022] The multi-head attention output, after residual connections and layer normalization, is then passed through a position feedforward network:

[0023]

[0024] through After layer encoding, the bending moment prediction result is obtained through the linear output head. And equivalent incentive prediction .

[0025] in, It is the dimension of each attention head. For FFN input features, and The weight matrix is ​​a learnable matrix. and This is a learnable bias vector.

[0026] Furthermore, the equivalent second-order dynamic equation of the trainable second-order dynamic physical operator is:

[0027]

[0028] in, , and Do not use the equivalent mass matrix, damping matrix, and stiffness matrix to be learned. and These are the first and second time derivatives of the predicted bending moment, respectively.

[0029] Cholesky decomposition is used to parameterize the quality matrix. and stiffness matrix Symmetric positive definiteness:

[0030]

[0031]

[0032] in, and For a trainable lower triangular matrix, Here is the regularization constant. for An identity matrix of order 1.

[0033] Damping matrix Rayleigh damping is used:

[0034]

[0035] in, Here is the Rayleigh damping coefficient, ensuring non-negativity.

[0036] Furthermore, the physical consistency residual of the predicted sequence for,

[0037]

[0038] in, For the first Discrete time moments The velocity term is approximated using the central difference. The acceleration term is approximated using the central difference. The equivalent incentive for prediction.

[0039] Furthermore, the multi-domain loss function is,

[0040]

[0041] in, For fitting time-domain data, channel-weighted Huber loss is used; This is the physical residual loss; For spatial smoothing loss based on Euler-Bernoulli beam theory; This represents the loss of amplitude consistency in the frequency domain. Amplitude statistical loss; To provide an equivalent incentive for regularized loss, These are the weighting coefficients;

[0042] The weighting coefficients are dynamically adjusted using a linear annealing strategy.

[0043]

[0044] in, For the current training round, For the annealing cycle, and These are the initial and final weights, respectively.

[0045] A computer system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of a motion prediction method for a deep-sea mining hoisting system based on physical operators and multiple codes.

[0046] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a motion prediction method for a deep-sea mining hoisting system based on physical operators and multiple codes.

[0047] A computer program product includes a computer program that, when executed by a processor, implements the steps of a motion prediction method for a deep-sea mining hoisting system based on physical operators and multiple codes.

[0048] The beneficial effects of this invention are as follows:

[0049] By introducing trainable dynamic physics operators, the equivalent mass, damping, and stiffness characteristics of the system are explicitly characterized, improving the physical consistency of the prediction results. Utilizing the self-attention mechanism of the Transformer network, long-term dependencies and non-stationary features in the coupled motion response of the deep-sea mining hoisting system are effectively extracted. At critical locations where traditional methods struggle to predict accurately, such as the ship-riseer interface, the prediction accuracy R² can reach over 0.95. Inference time is on the order of milliseconds, representing a speedup of approximately 10 times compared to high-fidelity simulation. 5 It can be multiplied by 10 times, making it suitable for real-time monitoring; it can simultaneously identify the equivalent dynamic parameters of the system, providing a basis for structural safety assessment. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0051] Figure 2 A schematic diagram of the Transformer network structure with physical operator regularization;

[0052] Figure 3 This is a schematic diagram showing the components and monitoring locations of a deep-sea mining hoisting system.

[0053] Figure 4 Comparison of bending moment time history predictions between simulation and PORT-Net model (10.5m monitoring point);

[0054] Figure 5 Comparison of bending moment time history predictions between simulation and PORT-Net model (533m monitoring points);

[0055] Figure 6 Comparison of bending moment time history predictions between simulation and PORT-Net model (1043m monitoring point);

[0056] Figure 7 Comparison of bending moment time history predictions between simulation and PORT-Net model (1533m monitoring point);

[0057] Figure 8 Comparison of bending moment time history predictions between simulation and PORT-Net model (2063m monitoring point). Detailed Implementation

[0058] The present invention will now be further described with reference to the accompanying drawings.

[0059] This invention relates to a motion prediction method, system, medium, and product based on physical operators and multi-encoding for deep-sea mining hoisting systems, such as... Figure 1 As shown, the method of the present invention includes four steps: S1 system dynamics modeling and data acquisition, S2 conditional encoding and Transformer sequence prediction, S3 physical operator construction and dynamic constraints, and S4 multi-domain loss function construction and joint optimization.

[0060] S1. System dynamics modeling and data acquisition;

[0061] like Figure 3 As shown, the deep-sea mining hoisting system includes a surface support vessel, a hoisting riser, booster pumps, a relay station, flexible hoses, and mining vehicles. In this embodiment, the system operates at a water depth of 6000m, with a total hoisting riser length of 5490m. It is equipped with four booster pumps connected in series, installed at depths of 533.0m, 1043.0m, 1553.0m, and 2063.0m, respectively.

[0062] S11, Modeling of surface support vessel;

[0063] The hydrodynamic coefficients are solved in the frequency domain using linear potential flow theory to obtain frequency-dependent additional mass, radiation damping, and first-order wave excitation force.

[0064] The JONSWAP spectral density function is:

[0065]

[0066] in, Phillips constant, It is the acceleration due to gravity. The peak angular frequency, Angular frequency, Peak enhancement factor The spectral width parameter is used to synthesize the ship's six-degree-of-freedom motion through spectral superposition under irregular sea conditions, which serves as the boundary condition for the motion of the top of the riser.

[0067] S12, Modeling of riser and flexible hose;

[0068] The lumped mass method is used to perform time-domain discretization modeling of the hoisting riser, relay station, flexible hose, and mining vehicle. The semi-discrete equilibrium equation at point is:

[0069]

[0070] in, , , These are the nodal mass matrix, damping matrix, and stiffness matrix, respectively. The nodal displacement vector; and They are respectively The first and second derivatives with respect to time, i.e., the velocity and acceleration vectors; Axial tensile force, bending moment, and torsional internal force transmitted between adjacent units; It is an external load that includes hydrodynamic, gravity, buoyancy and internal flow effects.

[0071] The hydrodynamic loads on slender structures are calculated using the Morison formula:

[0072]

[0073] in, The density of seawater, and These are the inertia coefficient and the drag coefficient, respectively. outer diameter, For the projected area, To displace the volume, This represents the relative displacement between the structure and the water particles. and They are respectively The first and second derivatives with respect to time, i.e., the velocity and acceleration vectors.

[0074] S13, Data Acquisition;

[0075] The coupled motion equations of the overall system are:

[0076]

[0077] , and These are the mass matrix, damping matrix, and stiffness matrix of the overall system, respectively. This is the displacement vector of the entire system; and They are respectively The first and second time derivatives are the velocity and acceleration vectors, respectively; Let be the external load vector, and be the time vector. With environmental parameters The function.

[0078] The implicit Newmark scheme was used for integration with a time step of 0.01s, and the simulation duration was 3600s. Environmental conditions included a meaningful wave height of 2.5m, a spectral peak period of 7.0s, a surface flow velocity of 1m / s, and wind, wave and flow directions varying at 15° intervals, generating a total of 195 operating conditions.

[0079] Five key monitoring sections were selected, with arc lengths of 10.5m, 533.0m, 1043.0m, 1553.0m, and 2063.0m, respectively. The data were divided into 156 training conditions and 39 test conditions in an 8:2 ratio.

[0080] S2, Conditional Coding and Transformer Sequence Prediction;

[0081] like Figure 2 As shown, the inputs to PORT-Net include environmental-condition coding and continuous-time location coding.

[0082] S21, Conditional coding vector;

[0083] Define the environment-condition description vector:

[0084]

[0085] in, For the sake of righteousness, the waves rise high. For the spectral peak period, For ship speed, These represent the directional angles of waves, wind, and ocean waves, respectively.

[0086] Perform Min-Max normalization on each component:

[0087]

[0088] in, It is the first of the input vectors One characteristic, and They are the first The minimum and maximum values ​​of each feature across all samples. These are normalized eigenvalues.

[0089] Here, a conditional projection layer is introduced to... Mapped to conditional coding vector :

[0090]

[0091] in, For learnable weight matrix, It is a learnable bias vector.

[0092] set up For high-fidelity models in time The reference bending moment vector at a given location is used to predict the mapping relationship described below.

[0093]

[0094] in, This represents the predicted load; the mapping is performed using trainable parameters. Conditional transformer network accomplish.

[0095] S22, Continuous Time Position Encoding;

[0096] Normalize the time:

[0097]

[0098] in, No. The time value at each discrete moment. It is the time value of the first moment. It is the time value at the last moment. It represents the total number of time steps in the sequence.

[0099] Subsequently, the normalized time is encoded using a set of sine functions, whose angular frequencies are logarithmically spaced to construct sine-cosine codes:

[0100]

[0101] in, It is the first angular frequency, This represents the number of angular frequencies used.

[0102] The continuous time position encoding is constructed as follows:

[0103]

[0104]

[0105] in, The sinusoidal part of the vector is encoded for continuous time position. The cosine portion vector encoded for continuous time positions. Connected encoded vector. The temporal code is obtained by projecting it onto the model dimension through a learnable linear transformation. :

[0106]

[0107] in, For the first Normalized time at each point in time, For connecting the encoding vector, For learnable weight matrix, This is the offset vector.

[0108] The conditional feature vector is added to the temporal code and used as the input to the Transformer encoder:

[0109]

[0110] in, For conditional coding vectors, It is a continuous time position encoding vector.

[0111] Stack the input vectors of all time steps in chronological order to form the input sequence. .

[0112] S23, Transformer encoder; each layer applies a multi-head self-attention mechanism, followed by a position feedforward network.

[0113] Multi-head self-attention mechanism: for the first Layer, input sequence Projected as a query ,key Sum ,

[0114]

[0115] in, , and The projection matrix is ​​learnable;

[0116] Calculate the scaled dot product attention:

[0117]

[0118] in, It is the dimension of each attention head.

[0119] The multi-head attention output, after residual connection and layer normalization, is passed through a position feedforward network, which consists of two linear transformations and uses the ReLU activation function.

[0120]

[0121] in, For FFN input features, and The weight matrix is ​​a learnable matrix. and Learnable bias vectors. After layer encoding, the bending moment prediction result is obtained through the linear output head. And equivalent incentive prediction .

[0122] S3, Construction of physical operators and dynamic constraints;

[0123] Establish the equivalent second-order dynamic equations acting on the predicted bending moment sequence:

[0124]

[0125] in, , and Do not use the equivalent mass matrix, damping matrix, and stiffness matrix to be learned. The predicted load sequence is the output of the network. and These are the first and second time derivatives of the predicted bending moment, respectively. This is the prediction of the equivalent incentive output of the network.

[0126] Cholesky decomposition parameterization is used to ensure the symmetric positive definiteness of the mass and stiffness matrices:

[0127]

[0128]

[0129] in, and For a trainable lower triangular matrix, Here is the regularization constant. for An identity matrix of order 1.

[0130] S32, Damping matrix;

[0131] The damping matrix adopts the Rayleigh damping form:

[0132]

[0133] in, The Rayleigh damping coefficient is used to ensure non-negativity. Equivalent quality matrix Equivalent stiffness matrix.

[0134] S33, Physical residual calculation;

[0135] For internal time points The central difference approximation is used for velocity and acceleration:

[0136]

[0137] in, For the first Discrete time moments Time step , , The predicted bending moment is at three adjacent time points. The velocity term is approximated using the central difference. This is the acceleration term approximated by the central difference.

[0138] exist The physical residual vector at time step is:

[0139]

[0140] in, The equivalent incentive for prediction.

[0141] S4. Construction and joint optimization of multi-domain loss functions;

[0142] The total loss function is:

[0143]

[0144] in, For fitting time-domain data, channel-weighted Huber loss is used; This is the physical residual loss; For spatial smoothing loss based on Euler-Bernoulli beam theory; This represents the loss of amplitude consistency in the frequency domain. Amplitude statistical loss; To equalize the regularization loss These are the weighting coefficients.

[0145] S41, Time-domain data fitting loss ;

[0146] The main regression loss function uses the channel-weighted Huber loss function to measure the difference between the model's predicted value and the reference signal in the time domain. This loss function is robust to outliers.

[0147]

[0148] in, For batch size, For time steps, For predicted values, For the true value, yes Channel-specific weights for each monitoring segment, Huber's loss function:

[0149]

[0150] in, This is the conversion threshold. The channel weights are set to... To highlight the challenging ship-rise interface and first pump station, This is the error.

[0151] S42, Frequency domain amplitude loss;

[0152] To ensure that PORT-Net can capture dynamic features across frequency bands, the frequency domain loss function penalizes the difference between the predicted amplitude spectrum and the reference amplitude spectrum:

[0153]

[0154] in, This represents the discrete Fourier transform along the time axis. It is the number of significantly low-frequency bands that are retained. These are channel-specific weights; frequency weights are set to... To particularly enhance the spectral matching of the 10.5m and 533.0m bands.

[0155] S43, Physical residual loss;

[0156] Explicitly penalize the residuals of a trainable second-order system:

[0157]

[0158] in, See formula for definition , This represents the channel weight.

[0159] S44, Spatial Smoothing Loss;

[0160] Indicates the time step and arc length Monitoring section The predicted bending moment. The spatial smoothing loss is constructed based on Euler-Bernoulli beam theory, calculating the discrete spatial second derivative of the bending moment on a non-uniform spatial grid:

[0161]

[0162] Collect these values ​​into a vector In the mean square norm It can penalize non-smooth, oscillatory spatial patterns that are inconsistent with the bending behavior of Euler-Bernoulli beams under distributed loads. This term serves as a physics-driven spatial smoothing prior for predicting the bending moment distribution along the riser.

[0163] The spatial smoothing loss is:

[0164]

[0165] in, It is the discrete spatial second derivative at the internal monitoring point.

[0166] S45, Amplitude statistical loss;

[0167] To further stabilize the prediction results of the upper section, an amplitude statistical loss is introduced into the loss function to ensure the consistency between the predicted time series statistics and the reference time series statistics.

[0168]

[0169] in, and These represent the time mean and standard deviation of the predicted response, respectively. This loss function is applied only to selected channels (10.5m and 533.0m) to enhance the model's performance at the ship-riser interface.

[0170] S46, Spatial Smoothing Loss;

[0171] To prevent non-physical outlier excitations and stabilize the training process, the predicted equivalent excitations are regularized.

[0172]

[0173] in, The equivalent incentive for prediction.

[0174] Introducing a physical regularization term , , and And assign it a smaller initial weight.

[0175] S47, Linear Annealing Strategy;

[0176] As training progresses, the weight coefficients are dynamically adjusted using a linear annealing strategy:

[0177]

[0178] in, For the current training round, For the annealing cycle, and These are the initial and final weights, respectively. This strategy allows training to smoothly transition from early data-driven learning to later physical constraint optimization.

[0179] In this embodiment, Annealing from 1.0 to 0.55, Annealing from 0.03 to 0.40, Annealing from 0.02 to 0.03, Annealing from 0 to 0.25, The annealing cycle from 0 to 0.15 is 400 rounds.

[0180] Training was performed using the Adam optimizer with an initial learning rate of 10. -3 Gradient clipping norm 1.0, batch size 4.

[0181] S5. Verification results;

[0182] The prediction performance of PORT-Net at each monitoring section is shown in Table 1:

[0183] Table 1 Predictive performance of each monitoring section

[0184]

[0185] At the highest monitoring section (10.5m), PORT-Net achieved an R² value of 0.9544, an MSE value of 0.9268 (kN·m)², and a MAE value of 0.6320 kN·m. This location, situated at the water inlet below the ship-riseer connection, is directly governed by the six-DOF ship motion and exhibits significant non-stationarity. The improved prediction accuracy at this section is attributed to the dynamic constraints of the physical operator and the channel-weighted loss strategy. In the intermediate and deep sections (1043.0–2063.0m), the prediction accuracy R² exceeded 0.99. These sections exhibit characteristic flexible beam behavior, with the response dominated by the first few bending modes, closely matching the form of the trainable second-order dynamic operator.

[0186] Despite the mismatch between the second-order operator formula and the kinematically driven boundary conditions, PORT-Net still achieved a convergence coefficient of R² > 0.95, indicating that the conditional transformer played a complementary role in capturing external driving dynamics. This superior performance in this challenging river section is attributed to three design features: (1) the channel-weighted Huber loss function assigns the highest weight of 4.0 to the 10.5-meter channel; (2) the frequency domain consistency loss function employs a higher channel weight of 7.0, ensuring spectral matching of the dominant wave frequency; and (3) the amplitude statistics loss function is specifically optimized for mean and standard deviation consistency in the 10.5-meter and 533.0-meter channels.

[0187] At the first pumping station (533.0 meters), the model's R² value was 0.8926, the mean square error (MSE) was 0.0449 (kN·m)², and the mean absolute error (MAE) was 0.1509 kN·m. The decrease in R² value at deeper depths is due to the local stiffness discontinuities introduced by the pumping units, which produce complex wave reflection and transmission phenomena.

[0188] At the middle and deeper monitoring points, prediction accuracy is significantly improved. The PORT-Net model achieves R²=0.9904 at 1043.0 m, R²=0.9939 at 1553.0 m, and R²=0.9947 at 2063.0 m. These sections exhibit typical flexible beam behavior under distributed hydrodynamic loads, with their response primarily determined by the first few bending modes.

[0189] like Figure 4 As shown, at a depth of 10.5 meters, the bending moment fluctuates between 5 and 26 kN·m, exhibiting a typical irregular pattern caused by random wave excitation. The predicted time history curves are in high agreement with the simulation results, successfully capturing the low-frequency envelope changes associated with wave group effects and the high-frequency oscillations driven by individual wave components. The model accurately reproduces the time and amplitude of most peak events, including maximum values ​​exceeding 25 kN·m.

[0190] like Figure 5 As shown, at 533.0 meters, the PORT-Net model captures the baseline oscillation pattern while slightly smoothing out extreme peaks. This behavior is consistent with the Huber loss function, which reduces sensitivity to outliers by transitioning from quadratic to linear penalties above a threshold.

[0191] In a deeper section ( Figure 6-8 The predicted time histories matched well with the simulation results at depths of 1043.0m, 1553.0m, and 2063.0m, respectively. The error remained within a narrow range, confirming the model's reliability at medium depths. In the initial transient phase at 1043.0m (…),… s) Some biases were observed, which may be due to the model learning quasi-steady-state dynamics rather than transient startup behavior. The quasi-steady-state response dominates for most of the simulation time, and local errors have a small impact on the overall prediction quality.

[0192] PORT-Net's computational efficiency meets the requirements of real-time applications. The average training time for 400 epochs is 45 minutes. The total number of trainable parameters is 70,000. The inference time for predicting a complete 3600-second time series (240 time steps) is in the millisecond range, representing an speedup of approximately 10% compared to high-fidelity simulation. 5 This is a multiple of the standard and is suitable for real-time monitoring.

[0193] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A motion prediction method for deep-sea mining hoisting systems based on physical operators and multi-encoding, characterized in that, Includes the following steps: Step 1: Establish a dynamic model of the lifting system, acquire time-series data of coupled motion response under multiple working conditions, and construct a training set; the lifting system includes a surface support vessel, lifting riser, booster pump, relay station, flexible hose and mining vehicle; the dynamic model is modeled using the lumped mass method, the hydrodynamic load is calculated using the Morison formula, and the environmental conditions include wave, ocean current and wind loads. Step 2: Introduce a trainable second-order dynamic physics operator at the output of the Transformer sequence prediction network to build a prediction model; encode the environmental conditions and continuous time position as features, input them into the encoder of the Transformer sequence prediction network, and output the predicted sequence of bending moment and equivalent excitation of key monitoring sections. The equivalent second-order dynamic equation of the trainable second-order dynamic physical operator is: in, , and These are the equivalent mass matrix, damping matrix, and stiffness matrix to be learned, respectively. and These are the first and second time derivatives of the predicted bending moment, respectively; The bending moment prediction results; Cholesky decomposition is used to parameterize the quality matrix. and stiffness matrix Symmetric positive definiteness: in, and For a trainable lower triangular matrix, Here is the regularization constant. for An identity matrix of order 1; Damping matrix Rayleigh damping is used: in, The Rayleigh damping coefficient is used to ensure non-negativity; Step 3: Based on the second-order dynamic physical operator, calculate the physical consistency residual of the predicted sequence; The physical consistency residual of the predicted sequence for, in, For the first Discrete time moments The velocity term is approximated using the central difference. The acceleration term is approximated using the central difference. For the predicted equivalent incentive; Step 4: Construct a multi-domain loss function that includes time-domain prediction error, frequency-domain feature error, dynamic residuals, and spatial smoothing constraints; Step 5: Using the training set, with the goal of minimizing the multi-domain loss function, jointly optimize the encoder parameters and trainable second-order dynamic physical operators in the prediction model to obtain the trained prediction model. Step 6: Input the target working condition parameters into the trained prediction model and output the prediction results of the coupled motion of the lifting system.

2. The motion prediction method for deep-sea mining hoisting systems based on physical operators and multi-encoding as described in claim 1, characterized in that, The environmental operating condition coding includes the encoding of meaningful wave height. Spectral peak period Ship speed The direction angle of waves, wind, and ocean waves Min-Max normalization is performed and projected to the conditional feature vector obtained at time step. ; The continuous time position encoding uses a method based on a family of sine functions to generate time feature vectors. , The conditional feature vector is added to the temporal code and used as the input to the Transformer encoder: Stack the input vectors of all time steps in chronological order to form the input sequence. .

3. The motion prediction method for deep-sea mining hoisting systems based on physical operators and multi-encoding as described in claim 2, characterized in that, The Transformer encoder employs a sequence coding structure based on a multi-head self-attention mechanism. For the Layer input sequence Projected as a query ,key Sum Calculate the scaled dot product attention: The multi-head attention output, after residual connections and layer normalization, is then passed through a position feedforward network: through After layer encoding, the bending moment prediction result is obtained through the linear output head. And equivalent incentive prediction ; in, It is the dimension of each attention head. For FFN input features, and The weight matrix is ​​a learnable matrix. and It is a learnable bias vector.

4. The motion prediction method for deep-sea mining hoisting systems based on physical operators and multi-encoding as described in claim 1, characterized in that, The multi-domain loss function is, in, For fitting time-domain data, channel-weighted Huber loss is used; This is the physical residual loss; For spatial smoothing loss based on Euler-Bernoulli beam theory; This represents the loss of amplitude consistency in the frequency domain. Amplitude statistical loss; To provide an equivalent incentive for regularized loss, These are the weighting coefficients; The weighting coefficients are dynamically adjusted using a linear annealing strategy. in, For the current training round, For the annealing cycle, and These are the initial and final weights, respectively.

5. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1 to 4.

7. A computer program product, comprising a computer program, characterized in that: When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 4.

Images