Data prediction and early warning method, device and equipment suitable for long-distance transportation of deep sea mining and medium

Abstract

This application discloses a data prediction and early warning method, apparatus, equipment, and medium applicable to long-distance transportation in deep-sea mining, relating to the field of computer technology. The method includes: performing feature engineering processing on physical entity data and prior engineering data; fusing multi-fidelity data based on dimensionless numbers and target coupling indices; training a target model using the fused data; transferring the knowledge of the trained teacher model to a preset student model to obtain a target student model; optimizing the target student model to obtain an initial prediction model; if real-time operational data drift is detected, performing feature distribution alignment and small-step fine-tuning on the initial prediction model; performing meta-training on the processed prediction model to obtain a target prediction model; inputting a target vector constructed based on dimensionless numbers into the target prediction model to obtain prediction data and confidence intervals; and issuing an early warning based on the conditional value of risk and confidence intervals of the prediction data. This achieves high-precision prediction and early warning.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to data prediction and early warning methods, devices, equipment and media suitable for long-distance transportation in deep-sea mining. Background Technology

[0002] Two-phase flow in deep-water risers mainly faces the following unique and severe challenges: 1. Severe Flow Pattern Changes and Instabilities: A huge pressure gradient exists inside the riser, from the high-pressure environment of the seabed to the low-pressure environment of the sea surface. This causes drastic phase changes (gas expansion) and flow pattern transformations (from bubbly flow to slug flow, annular flow, etc.) in the oil-gas mixture as it rises, easily forming huge and severe slug flows. This causes periodic and massive impacts on the production system and riser structure, which is one of the biggest technical challenges.

[0003] 2. Risk of hydrate formation: Under the low temperature and high pressure conditions of the deep sea, natural gas and water are very likely to form solid natural gas hydrates, which can cause blockages in risers, valves and pipelines, leading to serious safety accidents and production interruptions.

[0004] 3. Solid deposition problem: As pressure and temperature change, solid substances such as asphalt and paraffin may precipitate and deposit on the inner wall of the riser, reducing the flow area, increasing pressure drop, and even causing blockage.

[0005] 4. Deep-sea high pressure and long distance: Deep-water risers can be thousands of meters long. The huge hydrostatic pressure and flow resistance make it very difficult to accurately predict parameters such as pressure drop and liquid holdup, but they are crucial for production optimization and safety management.

[0006] Current technologies often employ mathematical modeling of multiphase fluid dynamics to predict flow. The core of these high-fidelity simulation systems lies in simulating complex transient processes within risers by solving the conservation equations (mass, momentum, energy) describing multiphase flow. The specific simulation steps can be summarized as follows: First, preprocessing and system modeling are performed. Based on the riser's geometry, tilt angle, internal roughness, and the PVT properties (pressure-volume-temperature relationship) and phase behavior of the fluid components, an accurate physical model is constructed, and initial and boundary conditions (such as the output, pressure, and temperature at the seabed inlet and the back pressure of the surface platform) are defined. Next, the numerical solution and transient simulation phase begins. The software discretizes the entire system into a series of control volumes. Within each time step, it solves a set of governing equations (typically using a two-fluid model, treating each phase as a continuous, interpenetrating medium). Built-in empirical correlations for flow pattern transitions, interphase drag, heat and mass transfer are used to close the equations, thereby calculating the evolution of key parameters such as pressure, temperature, velocity, and phase holdup for each grid cell, reproducing unstable flow phenomena such as severe slug flow. Finally, results analysis and post-processing are performed. The transient data output from the simulation (such as pressure fluctuations at the bottom and top of the riser, changes in liquid holdup, and temperature distribution) are visualized and analyzed to assess flow stability, predict hydrate formation risk, estimate pressure drop and system energy consumption, and provide a basis for developing control strategies. The entire process relies heavily on a deep understanding of the underlying physicochemical mechanisms. However, this technology involves computationally intensive numerical solutions, failing to meet the timeliness requirements of real-time early warning and rapid optimization. Secondly, the fixed model parameters lack adaptability, making it difficult to cope with operational drift caused by reservoir changes and pipe wall scaling, necessitating frequent manual calibration by experts. Most critically, its predictions are single, deterministic values, lacking inherent uncertainty quantification and failing to provide a confidence level for the predictions, resulting in insufficient basis for risk decision-making. Furthermore, modeling extreme physical phenomena such as erosion and complex deposition still relies on empirical formulas, limiting accuracy. These shortcomings make it difficult for existing systems to achieve reliable early warning and proactive optimization.

[0007] Therefore, how to achieve high-precision prediction and early warning for long-distance transportation is an urgent problem to be solved. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide a data prediction and early warning method, device, equipment, and medium suitable for long-distance transportation in deep-sea mining, capable of achieving high-precision prediction and early warning of key indicators such as pressure drop, solid phase distribution, energy consumption, blockage, and erosion, thereby reducing energy consumption and downtime risks. The specific solution is as follows: In the first aspect, this application discloses a data prediction and early warning method applicable to long-distance transportation in deep-sea mining, including: The physical entity data and engineering prior data of the pipeline are acquired, and the physical entity data and engineering prior data are synchronized in time to obtain synchronized data; the physical entity data includes pressure, flow rate, solid content and vibration; the engineering prior data includes sea state, pipeline geometry and pump station or valve status. The synchronized data is subjected to feature engineering to extract the corresponding dimensionless numbers and target coupling indices. Multi-fidelity data fusion is then performed based on the dimensionless numbers and target coupling indices to obtain the fused data. The dimensionless numbers include Reynolds numbers, Froude numbers, Stokes numbers, Shields numbers, and Begno numbers. The target model is trained using the fused data to obtain a trained teacher model. The knowledge of the trained teacher model is then transferred to a preset student model based on knowledge distillation to obtain the target student model. The target student model is optimized based on the loss function determined according to the preset physical constraints and the fused data to obtain an initial prediction model. Concept drift detection is performed on the initial prediction model. If real-time working condition data drift is detected, feature distribution alignment and small step size fine-tuning are performed on the initial prediction model to obtain a processed prediction model. Meta-training is performed on the processed prediction model to obtain the target prediction model. A target vector is constructed based on the dimensionless number, and the target vector is input into the target prediction model to obtain the corresponding prediction data and the confidence intervals corresponding to the prediction data. The prediction data includes pressure drop along the pump shaft, solid volume distribution, pump shaft power, erosion rate, and blockage probability. Determine the conditional risk value for each of the predicted data, and issue corresponding early warnings based on the conditional risk value and the confidence interval corresponding to each of the predicted data.

[0009] Optionally, the feature engineering processing of the synchronized data includes: Determine the observation vector of the synchronized data; The Kalman filter algorithm is used to repair missing data in the physical entity data based on the observation vector, and to correct abnormal observations in the observation vector to obtain corresponding processed data, which can then be used for feature engineering.

[0010] Optionally, the process of determining the target coupling index includes: The contact-hydraulic coupling index is determined by a preset formula; the preset formula for determining the contact-hydraulic coupling index is as follows: ; Wherein, TCCI is the contact-hydraulic coupling index; The modulus of contact force; For fluid force model; It is a constant; This represents the volume fraction of the solid phase. For reference volume fraction; The radial force-volume fraction coupling index is determined by a preset formula; the preset formula for determining the radial force-volume fraction coupling index is as follows: ; Wherein, RFVF is the radial force-volume fraction coupling index; This represents the average value of the radial force. This represents the average radial velocity. This represents the rate of change of radial force with volume fraction. This represents the radial gradient of the radial velocity.

[0011] Optionally, before optimizing the target student model based on the loss function determined according to preset physical constraints and the fused data, the method further includes: The overall loss function for multi-task learning is determined by a pre-defined formula; the pre-defined formula for determining the overall loss function for multi-task learning is as follows: ; in, is the overall loss function for multi-task learning; k represents the different learning tasks; These are the true value vector and the model's predicted value vector for the k-th task, respectively. The heteroscedasticity uncertainty for the predicted k-th task; The conserved residual term is determined by a pre-defined formula; the pre-defined formula for determining the conserved residual term is as follows: ; in, For the conserved residual term; For divergence operators; The density of the mixture; The velocity vector field of the mixture; The square of the L2 norm; The monotonic residual term is determined by a preset formula; the preset formula for determining the monotonic residual term is as follows: ; in, The term is the monotonic residual; ReLU is the linear rectified function; The partial derivative of the pressure drop with respect to the flow rate Q; The loss function for multi-fidelity learning is determined by a pre-defined formula; the pre-defined formula for determining the loss function for multi-fidelity learning is as follows: ; in, The loss function for multi-fidelity learning; A mapping function represented by a neural network; This is the output of the high-fidelity model; The square of the L2 norm; This is the output of the low-fidelity model; For hyperparameters; For mapping functions The gradient.

[0012] Optionally, the optimization of the target student model based on a loss function determined according to preset physical constraints and the fused data to obtain an initial prediction model includes: The loss function is determined based on preset physical constraints; the formula for determining the loss function is: ; in, The loss function is... Let be the overall loss function for the multi-task learning; For weak-form PDE residual constraints; For the consistency residuals of the submerged boundary method; For the conserved residual term; For the monotonic residual term; Θ represents the L2 regularization term; Θ represents all trainable parameters of the model. These are the weighting hyperparameters corresponding to each loss or residual; The loss function for multi-fidelity learning; The target student model is optimized based on the loss function and the fused data to obtain an initial prediction model.

[0013] Optionally, the step of issuing corresponding early warnings based on the conditional value of risk and the confidence intervals corresponding to each of the predicted data includes: If the conditional risk value does not meet the confidence interval corresponding to the predicted data, then a risk is determined to exist, and a warning light is used to issue an early warning.

[0014] Optionally, the method further includes: Based on the model predictive control algorithm, the corresponding confidence intervals of the predictive data and the current risk situation, the corresponding data adjustment suggestions are output. The predicted data and the suggested adjustments are visualized through a human-computer interface.

[0015] Secondly, this application discloses a data prediction and early warning device suitable for long-distance transportation in deep-sea mining, comprising: The synchronized data acquisition module is used to acquire the physical entity data and engineering prior data of the pipeline, and synchronize the physical entity data and engineering prior data in time to obtain synchronized data; the physical entity data includes pressure, flow rate, solid content and vibration; the engineering prior data includes sea state, pipeline geometry and pump station or valve status. The data fusion module is used to perform feature engineering processing on the synchronized data to extract the corresponding dimensionless numbers and target coupling indices, and to perform multi-fidelity data fusion based on the dimensionless numbers and target coupling indices to obtain fused data; the dimensionless numbers include Reynolds numbers, Froude numbers, Stokes numbers, Shields numbers, and Begno numbers; The knowledge transfer module is used to train the target model using the fused data to obtain a trained teacher model, and to transfer the knowledge of the trained teacher model to a preset student model based on knowledge distillation to obtain the target student model. The post-processing prediction model acquisition module is used to optimize the target student model based on the loss function determined according to the preset physical constraints and the fused data to obtain an initial prediction model. Concept drift detection is performed on the initial prediction model. If real-time working condition data drift is detected, feature distribution alignment and small step size fine-tuning are performed on the initial prediction model to obtain the post-processing prediction model. The prediction data acquisition module is used to perform meta-training on the processed prediction model to obtain the target prediction model, construct a target vector based on the dimensionless number, and input the target vector into the target prediction model to obtain the corresponding prediction data and the confidence intervals corresponding to the prediction data; the prediction data includes pressure drop along the friction, solid volume distribution, pump shaft power, erosion rate, and blockage probability; The early warning module is used to determine the conditional risk value of each of the predicted data, and to issue corresponding early warnings based on the conditional risk value and the confidence interval corresponding to each of the predicted data.

[0016] Thirdly, this application discloses an electronic device, including: Memory, used to store computer programs; A processor is used to execute computer programs to implement data prediction and early warning methods applicable to long-distance transportation in deep-sea mining, as described above.

[0017] Fourthly, this application discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned data prediction and early warning method for long-distance transportation in deep-sea mining.

[0018] This application first acquires the physical entity data and prior engineering data of the pipeline, and then synchronizes the physical entity data and prior engineering data in time to obtain synchronized data. The physical entity data includes pressure, flow rate, solid content, and vibration. The prior engineering data includes sea state, pipeline geometry, and pump station or valve status. Feature engineering processing is performed on the synchronized data to extract corresponding dimensionless numbers and target coupling indices. Multi-fidelity data fusion is then performed based on the dimensionless numbers and target coupling indices to obtain fused data. The dimensionless numbers include Reynolds number, Froude number, Stokes number, Shields number, and Begno number. The fused data is used to train a target model to obtain a trained teacher model. Knowledge distillation is then used to transfer the knowledge of the trained teacher model to a preset student model to obtain a target student model. The target student model is optimized based on the loss function determined by preset physical constraints and the fused data to obtain an initial prediction model. Concept drift detection is performed on the initial prediction model. If real-time operating data drift is detected, feature distribution alignment and small-step fine-tuning are performed on the initial prediction model to obtain a processed prediction model. Meta-training is performed on the processed prediction model to obtain a target prediction model. A target vector is constructed based on the dimensionless number and input into the target prediction model to obtain corresponding prediction data and confidence intervals for each prediction data. The prediction data includes pressure drop along the pump shaft, solid volume distribution, pump shaft power, erosion rate, and blockage probability. The conditional risk value of each prediction data is determined, and corresponding early warnings are issued based on the conditional risk value and the confidence intervals corresponding to each prediction data. As can be seen, this application ensures the rationality of predictions and the stability of extrapolation by embedding the physical laws governing multiphase flow as fundamental constraints into a deep learning framework; it utilizes real-time drift detection and adaptive learning mechanisms to enable the model to dynamically track changes in operating conditions and maintain long-term effectiveness; and it employs uncertainty quantification technology to provide confidence intervals for the prediction results of key indicators such as pressure drop, solid phase distribution, energy consumption, erosion rate, and blockage probability, thereby achieving an upgrade from "point prediction" to "probabilistic prediction." Ultimately, through high-precision prediction and risk-based early warning, the system provides decision support for optimizing operations, reducing energy consumption, and preventing unplanned downtime, thereby improving the safety and economy of the conveying process. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This application discloses a flowchart of a data prediction and early warning method for long-distance transportation in deep-sea mining. Figure 2 This is a schematic diagram of a multi-task loss and operator network structure disclosed in this application; Figure 3 This is a schematic diagram of a concept drift detection and online update disclosed in this application; Figure 4 This is a schematic diagram of a data prediction and early warning device for long-distance transportation in deep-sea mining, as disclosed in this application. Figure 5 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Deep-sea polymetallic nodule / sulfide mining requires transporting coarse-grained seawater mixtures over long distances to the surface via risers, bends, and subsea pipelines. This process presents challenges such as wide particle size distribution (e.g., 10-50 mm), solids content fluctuations, sea state disturbances, cascade pumping, and the coexistence of erosion and blockage risks. Traditional pure mechanistic models (such as one-dimensional pressure drop correlations) require cumbersome calibration and incur high computational costs when rapidly generalized to new sea areas / slurries; pure data-driven models lack physical consistency and interpretability, and their performance degrades sharply under domain shifts. To address these technical problems, this application discloses a data prediction and early warning method, device, equipment, and medium suitable for long-distance transport in deep-sea mining. This method enables high-precision prediction and early warning of key indicators such as pressure drop, solids distribution, energy consumption, blockage, and erosion, reducing energy consumption and downtime risks.

[0023] See Figure 1 As shown in the figure, this invention discloses a data prediction and early warning method suitable for long-distance transportation in deep-sea mining, including: Step S11: Obtain the physical entity data and engineering prior data of the pipeline, and synchronize the physical entity data and engineering prior data in time to obtain synchronized data; the physical entity data includes pressure, flow rate, solid content and vibration; the engineering prior data includes sea state, pipeline geometry and pump station or valve status.

[0024] In this embodiment, the core objective of the current step is to acquire comprehensive, high-quality, and standardized data to remove data obstacles for feature extraction and model training, ensuring the reliability and validity of the input data. Therefore, real-time signals such as pressure, flow rate, solid content, and vibration are collected through a sensor array. Environmental and design parameters are also acquired, including prior engineering data such as sea state, pipeline geometry, and pump / valve status. Since these data are multimodal, heterogeneous time-series data, they must meet time synchronization requirements. Therefore, the timestamps of all connected devices are uniformly calibrated to the same standard time base. This involves synchronizing the physical entity data and the prior engineering data to obtain synchronized data.

[0025] Step S12: Perform feature engineering on the synchronized data to extract the corresponding dimensionless numbers and target coupling indices. Perform multi-fidelity data fusion based on the dimensionless numbers and target coupling indices to obtain fused data. The dimensionless numbers include Reynolds numbers, Froude numbers, Stokes numbers, Shields numbers, and Begno numbers.

[0026] In this embodiment, the observation vector of the synchronized data is determined; the Kalman filter algorithm is used to repair missing data in the physical entity data based on the observation vector, and abnormal observations in the observation vector are corrected to obtain corresponding processed data, which is then used for feature engineering processing. Specifically, observation data modeling: ; in, H(...) represents the vector of observations at time t, typically derived from sensor or experimental measurements. t As an observation operator, it takes the true state of the model (such as the flow field u, particle phase fraction) (etc.) is mapped to the observation space. U is the velocity field vector of the fluid; The observation noise vector represents the measurement error. This indicates that the expression follows a pattern with a mean of 0 and a covariance matrix of... It follows a multivariate normal distribution. The covariance matrix of the observation noise describes the magnitude of the noise and the correlation between the noise levels of each observation. Furthermore, by modeling the observation data, the mapping relationship between sensor observations and the actual system state is quantified, including observation noise modeling, providing a foundation for data assimilation.

[0027] Additionally, regarding the reconstruction of missing data: ; in, Given all observations up to time t, this is the posterior estimate of the system state. K represents the prior estimate (i.e., the predicted value) of the system state after all observations up to time t-1. t Let y be the Kalman gain matrix at time t, which balances the uncertainty of the predicted and observed values. t Let be the actual observation value at time t. H is the observation matrix, which maps the state space to the observation space. P t|t-1 Let be the prior estimated error covariance matrix. T is the matrix transpose operator. R is the covariance matrix of the observation noise. Then, Kalman filtering is used to estimate the system state from noisy observations, repair missing or outlier data, and improve data quality.

[0028] Feature engineering is then performed on the processed data to extract the corresponding dimensionless numbers and target coupling indices. The dimensionless numbers are defined as follows: ; ; ; ; ; Where Re is the Reynolds number, which represents the ratio of fluid inertial force to viscous force; U is the fluid density; U is the characteristic velocity (such as incoming flow velocity, average flow velocity in the pipe); D is the characteristic length (such as pipe diameter, plate length, particle diameter). ρ is the fluid dynamic viscosity. Fr is the Froude number, representing the ratio of inertial force to gravity. g is the gravitational acceleration; Stk is the Stokes number, representing the ratio of the characteristic time of particle response to flow change to the characteristic time of flow. Particle density; θ is the particle diameter; θ is the Shield number, which represents the ratio of the drag force of the water flow on the sediment at the bed surface to the underwater weight of the sediment, and is used to determine the initiation of sediment movement. Ba is the bed shear stress; Ba is the Begno number, which characterizes the ratio of the shear stress generated by particle collision to the shear stress generated by the fluid, and is used to determine the fluidization state of particles. denoted as shear rate.

[0029] In addition, the process of determining the target coupling index includes: The contact-hydraulic coupling index is determined by a preset formula; the preset formula for determining the contact-hydraulic coupling index is as follows: ; Wherein, TCCI is the contact-hydraulic coupling index; The modulus of contact force; For fluid force model; It is a constant; This represents the volume fraction of the solid phase. For reference volume fraction; The radial force-volume fraction coupling index is determined by a preset formula; the preset formula for determining the radial force-volume fraction coupling index is as follows: ; Wherein, RFVF is the radial force-volume fraction coupling index; This represents the average value of the radial force. This represents the average radial velocity. This represents the rate of change of radial force with volume fraction. This represents the radial gradient of the radial velocity.

[0030] After extracting the dimensionless number and the target coupling index, multi-fidelity data fusion is performed based on these parameters. Specifically, the bias of data at different fidelity levels is modeled using a Gaussian process; low-fidelity data is used to expand sample coverage, and high-fidelity data is used to calibrate accuracy. Data types include: low-cost low-fidelity data (rapid simulations, empirical formula results) and high-cost high-fidelity data (simulation data, experimental data). Deep multi-fidelity mapping: ; in, The loss function for multi-fidelity learning; Let y be a mapping function represented by a neural network, with parameters θ. The goal is to learn how to generate a low-fidelity output y. L To high-fidelity output y H Complex mappings. The square of the L2 norm; For mapping functions The gradient. β is a hyperparameter used to balance the weights of the two terms in the loss function. Thus, through the neural network... The low-fidelity to high-fidelity mapping is learned, and gradient regularization ensures the smoothness of the mapping.

[0031] When data in a critical area (such as high-precision experiments or complex simulations) is extremely scarce or prohibitively expensive, this application instead utilizes more readily available "knowledge" from other sources to aid learning. Multi-fidelity data fusion: This involves simultaneously using a large amount of low-precision, low-cost data (such as rapid but coarse simulations or empirical formula results) and a small amount of high-precision, high-cost data (such as detailed simulations or real experimental data). The model learns general patterns from the massive amount of low-precision data, and then uses a small amount of high-precision data for calibration and correction, thereby approximating high-precision results without overly relying on expensive data.

[0032] Step S13: Use the fused data to train the target model to obtain a trained teacher model. Based on knowledge distillation, transfer the knowledge of the trained teacher model to a preset student model to obtain the target student model.

[0033] In this embodiment, the target model is trained using the fused data to obtain a trained teacher model. Then, based on knowledge distillation, the knowledge of the trained teacher model is transferred to a preset student model to obtain the target student model. The core logic is to transfer the knowledge of a complex, high-fidelity model (teacher model) to a lightweight model (student model) to meet the low-latency requirements of the edge. During knowledge distillation, the knowledge of a large, complex model (teacher) is transferred to a small, efficient model (student). Distillation loss: ; in, For knowledge distillation loss, is the temperature parameter used to soften the probability distributions of the teacher and student model outputs. KL is the KL divergence, used to measure the difference between the two probability distributions. The Softmax function transforms the output into a probability distribution. T S S These represent the logits (output values ​​before softmax) of the teacher model and the student model, respectively. γ is the weight used to balance the two losses. T y S These represent the final predicted output values ​​of the teacher model and the student model, respectively.

[0034] Knowledge distillation: Knowledge distillation is a model compression and knowledge transfer technique. Its core objective is to transfer knowledge learned from a pre-trained, high-performance but often complex and computationally expensive model (referred to as the "teacher model") to a simpler, smaller, and more efficient model (referred to as the "student model"). In this way, this application obtains fused, high-quality data and a lightweight student model architecture, providing core support for the subsequent construction of physically consistent multi-task models. The fused data contains effective information at different fidelities, more comprehensively reflecting the physical laws of the transmission process; the lightweight model architecture reserves computational space for subsequent embedding of complex physical constraints and achieving multi-task prediction, avoiding training difficulties or excessive inference latency due to overly complex models, while ensuring the model has a sufficient accuracy foundation.

[0035] Step S14: Optimize the target student model based on the loss function determined according to the preset physical constraints and the fused data to obtain an initial prediction model. Perform concept drift detection on the initial prediction model. If real-time working condition data drift is detected, perform feature distribution alignment and small step size fine-tuning on the initial prediction model to obtain a processed prediction model.

[0036] In this embodiment, to overcome the pain point of traditional data-driven models lacking physical rationality, this application ensures that the model's prediction results conform to basic physical laws by embedding physical constraints. Simultaneously, a multi-task architecture is used to achieve simultaneous prediction of multiple key indicators, improving the model's practicality and comprehensiveness. Firstly, the structure of the multi-task loss and operator network in this application is as follows: Figure 2 As shown, the core network consists of a Fourier Neural Operator (FNO) shared encoder and a multi-task head. Before optimizing the target student model based on the loss function determined according to preset physical constraints and the fused data, the overall loss function for multi-task learning is determined by a preset formula for determining the overall loss function for multi-task learning. The preset formula for determining the overall loss function for multi-task learning is as follows: ; in, is the overall loss function for multi-task learning; k represents the different learning tasks; These are the true value vector and the model's predicted value vector for the k-th task, respectively. The heteroscedasticity uncertainty for the predicted k-th task; The conserved residual term is determined by a pre-defined formula; the pre-defined formula for determining the conserved residual term is as follows: ; in, For the conserved residual term; For divergence operators; The density of the mixture; The velocity vector field of the mixture; The square of the L2 norm; The monotonic residual term is determined by a preset formula; the preset formula for determining the monotonic residual term is as follows: ; in, The term is the monotonic residual; ReLU is the linear rectified function; The partial derivative of the pressure drop with respect to the flow rate Q; The loss function for multi-fidelity learning is determined by a pre-defined formula; the pre-defined formula for determining the loss function for multi-fidelity learning is as follows: ; in, The loss function for multi-fidelity learning; A mapping function represented by a neural network; This is the output of the high-fidelity model; The square of the L2 norm; This is the output of the low-fidelity model; For hyperparameters; For mapping functions The gradient.

[0037] Finally, the loss function is determined based on preset physical constraints; the formula for determining the loss function is: ; in, The loss function is... Let be the overall loss function for the multi-task learning; For weak-form PDE residual constraints; For the consistency residuals of the submerged boundary method; For the conserved residual term; For the monotonic residual term; Θ represents the L2 regularization term; Θ represents all trainable parameters of the model. These are the weighting hyperparameters corresponding to each loss or residual; The loss function for multi-fidelity learning; The target student model is optimized based on the loss function and the fused data to obtain an initial prediction model. In this way, by balancing the noise levels of different tasks through heteroscedasticity weighting, and fusing physical constraints, multi-fidelity, distillation, and domain adaptive losses, the model's fitting accuracy, physical rationality, and cross-domain generalization ability are ensured. This application embeds physical laws as fundamental constraints into the deep learning model. This method mainly includes two levels of physical constraints: First, it uses the residuals (PDEs) of partial differential equations as basic constraints, requiring the model's prediction results to conform to the governing equations as closely as possible in either a "strong form" (pointwise satisfied) or a more flexible "weak form" (satisfied in the integral sense), thereby ensuring that its predictions follow basic physical conservation laws; second, it introduces higher-order monotonicity and convexity constraints, which, derived from the mathematical properties of thermodynamics and stability, ensure that the model's input-output relationship maintains a physically reasonable overall trend. The constructed physically consistent multi-task model is the core object for subsequent domain adaptation and online updates. The model's physical consistency ensures that its prediction logic still conforms to basic physical laws in cross-domain scenarios (such as different sea conditions and different pipeline configurations), providing a stable foundation for domain adaptation. Meanwhile, the multi-task output capability enables the simultaneous monitoring of the drift of multiple indicators during domain adaptation, ensuring that the updated model maintains high performance on all key tasks and avoiding the decrease in accuracy of other tasks caused by single-task adaptation.

[0038] In this application, to address the "concept drift" problem in practical applications—that is, the inconsistency between training data and actual data distribution caused by changes in operating conditions (such as new sea areas, new ship schedules, and changes in particle size ratios), leading to a decline in model performance—domain adaptation is used to achieve cross-domain generalization, and online updates ensure that the model maintains high accuracy under dynamic operating conditions. Firstly, as... Figure 3 As shown, the monitoring indicators are: prediction error within the sliding window (e.g., mean squared error) and feature embedding distribution statistics (mean, covariance). The detection algorithm is Page-Hinkley. ; in, Here is the Page-Hinkley statistic at time t; Let be the error at time i (e.g., prediction error). This is the cumulative mean of the error up to time t; λ represents the allowable drift amplitude; λ is the threshold for determining whether drift has occurred. It should be noted that the trigger condition for drift detection is: the statistic exceeds the historical 95th percentile threshold.

[0039] If real-time operating data drift is detected, the initial prediction model is subjected to feature distribution alignment and small-step fine-tuning to obtain a processed prediction model. Specifically, the distribution alignment method is as follows: ; in, It is the square of the maximum mean difference, used to measure the distance between two distributions. , Sample features from the source domain and the target domain, respectively; This is a function that maps features to the Hilbert space. , These represent the number of samples in the source domain and the target domain, respectively. Let be the norm in Hilbert space.

[0040] Wasserstein distance: ; in, For source domain distribution and target domain distribution The Wasserstein distance between them. inf is the infim (minimum). The joint distribution γ belonging to the coupled set Π has the following marginal distribution: and . This represents the distance between sample pairs.

[0041] The online update strategy is: small-step fine-tuning (based on the online model update rule of weighted loss): ; in, This is the parameter vector of the model (such as the weights and biases of a neural network). This is the assignment operator, indicating an update. The learning rate controls the step size for each parameter update. Let be the weight of the i-th sample. Let be the gradient operator with respect to the model parameter Θ. For the i-th sample (input is...) The loss function is calculated on .

[0042] Model rollback mechanism: If the metric decreases after an update, it reverts to the stable version in the model library. The final processed predictive model is then obtained.

[0043] Step S15: Perform meta-training on the processed prediction model to obtain the target prediction model. Construct a target vector based on the dimensionless number and input the target vector into the target prediction model to obtain the corresponding prediction data and the confidence intervals corresponding to the prediction data. The prediction data includes pressure drop along the flow path, solid volume distribution, pump shaft power, erosion rate, and blockage probability.

[0044] In this embodiment, after obtaining the target prediction model, meta-training is performed on the processed prediction model to obtain the target prediction model. The meta-learning framework is as follows: ; in, These are the model parameters after internal gradient updates for a specific task. A is the internal learning rate, used for task-specific rapid adaptation. This is the parameter vector of the model; This is the loss calculated on the support set for a given task. The support set is used for the internal adaptation of the task. Let be the loss calculated on the query set for a given task, where the model parameters are the adapted values. The query set is used to evaluate the performance of the adapted model and optimize the meta-objective. The goal of meta-learning is to optimize the initial model parameters. This allows the model to perform well on new tasks after only a few updates. This involves summing all meta-training tasks. Meta-training then enables the model to quickly adapt to new shipping schedules / new sea areas (cross-domain scenarios), reducing the need for new domain data.

[0045] In this way, the reliability and dynamic adaptability of the prediction results of the domain-adaptive and online-updated model are significantly improved, providing a more stable foundation for uncertainty quantification. Only when the model can maintain stable prediction performance in different domains and dynamic operating conditions can the uncertainty of the prediction results be accurately quantified. At the same time, information such as error changes and distribution shifts monitored during the online update process can also provide supplementary evidence for uncertainty assessment, making the calculation of confidence intervals more consistent with actual operating conditions.

[0046] Furthermore, model predictions inherently contain uncertainties (such as data noise and model approximation errors). By quantifying these uncertainties and constructing confidence intervals, the system provides users with an output of "predicted value + reliable range," avoiding decision-making errors due to deviations in a single predicted value and improving the credibility and practicality of the results. This application first integrates dimensionless features, particle characteristics, geometric and environmental parameters to form the model input. The target vector is then input into the target prediction model, simultaneously predicting multiple physical quantities: pressure drop along the flow path, solid phase volume distribution, pump shaft power, erosion rate, and blockage probability. The output results after uncertainty quantification (predicted value + confidence interval) form the core data foundation for edge-cloud collaboration and federated optimization. The edge device needs to quickly determine the risk level of the current transport process (such as whether the confidence interval of the blockage probability exceeds the safety threshold) based on the prediction results with confidence intervals, and trigger real-time inference or drift detection. The cloud device can analyze the commonalities and differences in global operating conditions based on the uncertainty data uploaded from multiple edge devices, providing a basis for global model training and parameter aggregation, while ensuring that the direction of parameter updates during federated optimization is more in line with actual needs.

[0047] Step S16: Determine the conditional risk value of each predicted data, and issue a corresponding early warning based on the conditional risk value and the confidence interval corresponding to each predicted data.

[0048] In this embodiment, model predictions are transformed into actionable engineering decisions. Potential risks under extreme conditions are assessed through risk quantification. Combining model predictions with uncertainties, optimal parameter tuning suggestions are output, achieving a closed loop of "prediction-risk assessment-decision optimization." When performing risk quantification: ; in, α is the conditional value of risk for a random variable Z (e.g., loss). It calculates the average loss under the worst-case scenario of 1-α. Z represents the random variable of loss. α is the confidence level (e.g., 0.95 or 0.99). For about variables Find the infimum (minimum value). Let be a threshold variable, typically corresponding to the value at risk. E is the expected value operator. It is a positive part function.

[0049] If the conditional risk value does not meet the confidence interval corresponding to the predicted data, a risk is determined to exist, and an early warning light is used. Furthermore, based on the model predictive control algorithm, corresponding data adjustment suggestions are output according to the predicted data, the confidence intervals corresponding to each predicted data, and the current risk situation; the predicted data and the data adjustment suggestions are visualized through a human-machine interface. Specifically, the prediction results in this application include point predictions and confidence intervals for pressure drop, solid phase distribution, pump shaft power, erosion rate, and blockage probability. Visualization: Prediction curves, risk heatmaps, and operational suggestions are displayed through a human-machine interface. The core achievements of the system (prediction results and decision suggestions) are presented to the user in an intuitive way, while experimental verification of system performance provides a basis for subsequent system iteration and optimization, ensuring that the system truly meets engineering requirements.

[0050] In this way, this application can be directly applied to scenarios such as deep-sea hoisting systems, long-distance subsea pipelines, and tailings return transportation in mineral processing, supporting design selection, energy consumption optimization, and online early warning of blockage and erosion risks, demonstrating significant economic value and engineering feasibility. This invention significantly improves prediction accuracy and robustness under wide particle size ratios and high solids content scenarios, reducing calibration costs; it maintains accuracy during domain shifts (new mining areas, equipment upgrades), and outputs actionable parameter tuning suggestions in real time.

[0051] In summary, this application first acquires the physical entity data and prior engineering data of the pipeline, and then synchronizes the physical entity data and prior engineering data in time to obtain synchronized data. The physical entity data includes pressure, flow rate, solid content, and vibration. The prior engineering data includes sea state, pipeline geometry, and pump station or valve status. Feature engineering processing is performed on the synchronized data to extract the corresponding dimensionless numbers and target coupling indices. Multi-fidelity data fusion is then performed based on the dimensionless numbers and target coupling indices to obtain fused data. The dimensionless numbers include Reynolds number, Froude number, Stokes number, Shields number, and Begno number. The target model is trained using the fused data to obtain a trained teacher model. Knowledge distillation is then used to transfer the knowledge of the trained teacher model to a preset student model to obtain the target student model. The target student model is optimized based on a loss function determined according to preset physical constraints and the fused data to obtain an initial prediction model. Concept drift detection is performed on the initial prediction model. If real-time operating data drift is detected, feature distribution alignment and small-step fine-tuning are performed on the initial prediction model to obtain a processed prediction model. Meta-training is performed on the processed prediction model to obtain a target prediction model. A target vector is constructed based on the dimensionless number and input into the target prediction model to obtain corresponding prediction data and confidence intervals for each prediction data. The prediction data includes pressure drop along the pump shaft, solid volume distribution, pump shaft power, erosion rate, and blockage probability. The conditional risk value of each prediction data is determined, and corresponding early warnings are issued based on the conditional risk value and the confidence intervals corresponding to each prediction data. As can be seen, this application ensures the rationality of predictions and the stability of extrapolation by embedding the physical laws governing multiphase flow as fundamental constraints into a deep learning framework; it utilizes real-time drift detection and adaptive learning mechanisms to enable the model to dynamically track changes in operating conditions and maintain long-term effectiveness; and it employs uncertainty quantification technology to provide confidence intervals for the prediction results of key indicators such as pressure drop, solid phase distribution, energy consumption, erosion rate, and blockage probability, thereby achieving an upgrade from "point prediction" to "probabilistic prediction." Ultimately, through high-precision prediction and risk-based early warning, the system provides decision support for optimizing operations, reducing energy consumption, and preventing unplanned downtime, thereby improving the safety and economy of the conveying process.

[0052] Furthermore, to balance real-time performance with global optimization needs while protecting data privacy, this application employs a multi-tiered approach: edge processing handles local real-time processing to meet low latency requirements; the cloud handles global model management and optimization to improve overall performance; and federated learning enables collaborative value across multiple data sources without leaking local data.

[0053] The division of labor in the architecture is as follows: Edge computing: data preprocessing, feature engineering, real-time inference, drift detection, small-step updates (low latency requirement ≤10ms, throughput ≥1000 samples / second, deployment recommendations); Cloud-based: Global model training, model library management, and parameter aggregation.

[0054] Federated learning process: Cloud initialization: Distribute initial model parameters ; Local training at the edge: Train using local data, calculate gradients or update parameters (raw data is not sent to the cloud to protect privacy). Parameter aggregation: ; This formula describes the core aggregation algorithm in federated learning.

[0055] The parameters of the global model after aggregation in round t+1 are given. K is the total number of clients participating in federated learning. K is the client index, k=1,2,...,K. This represents the number of local data samples on the k-th client. This represents the total number of data samples across all clients. This refers to the local model parameters for the k-th client after local training and before uploading. The global model parameters after aggregation in round t are used as the initial values ​​for local training. This is the learning rate. For model parameters The gradient operator. L k The loss function is calculated on the local data of the k-th client.

[0056] Privacy protection: Gradient cropping + differential privacy noise. ; This formula demonstrates how to add noise to the gradient to achieve differential privacy protection. This is a gradient estimate that satisfies differential privacy after adding noise. This is the gradient clipping function. It clips the gradient vector. The norm of L is restricted to within a threshold C; that is, if the gradient norm exceeds C, it is scaled down proportionally. L is the original gradient vector. C is the maximum norm threshold for gradient clipping. With a mean of 0 and a covariance matrix of σ 2 C 2 The gradient I follows a Gaussian (normal) distribution. This noise is added to the clipped gradient. σ is a parameter that controls the magnitude of the noise and determines the strength of privacy protection.

[0057] The edge-cloud collaborative architecture ensures efficient system deployment and privacy security, providing a stable operating environment and comprehensive data support for the seventh step of risk measurement and strategy optimization. Real-time inference results at the edge and global optimization models in the cloud provide multi-dimensional and high-precision input data for risk measurement (such as pressure drop prediction and congestion probability at each edge). Federated optimization, while protecting data privacy, maximizes the collaborative value of multi-end data, making risk assessment more comprehensive and strategy optimization more universal, ensuring that the output decision recommendations can be adapted to the specific operating conditions of different edge devices.

[0058] Furthermore, this application can combine predicted values, uncertainties, and risks to output feasible parameter tuning suggestions. Specifically, it predicts the optimization objective function of the controller using a robust model: ; Where min is the minimization operator; st is the sequence of control or state variables, which are the decision variables to be solved in the optimization problem. t=t0 is the starting time of the summation, the current time. t0+H is the ending time of the summation, the end point of the prediction time domain. H represents the length (number of steps) of the prediction time domain. This is the weighted quadratic norm. Q is a weight matrix used to measure the importance of the tracking error. Let be the model's predicted pressure drop along the friction path at time t. The desired or optimal target pressure drop along the process. This is the energy consumption weighting coefficient, used to balance the importance of energy consumption targets in the overall objective function. Let be the predicted specific energy consumption at time t. This is the risk weighting coefficient, used to balance the importance of risk cost in the overall objective function. At time t, the loss function Z t Alpha conditional value at risk is a financial risk measure used to assess the average loss under the worst-case 1-α scenario, making the controller more robust. To optimize the constraints of the problem, let represent the control or state variable s at each time t. t It must belong to a closed set U, which typically represents physical or operational constraints.

[0059] Finally, by combining predicted values, uncertainties, and risks, executable parameter tuning suggestions (pump speed, concentration window, sand flushing cycle) are output, with a self-training confidence threshold of 0.8-0.95 (deployment suggestion) to assist in strategy generation.

[0060] See Figure 4 As shown, this embodiment of the invention discloses a data prediction and early warning device suitable for long-distance transportation in deep-sea mining, comprising: The synchronized data acquisition module 11 is used to acquire the physical entity data and engineering prior data of the pipeline, and synchronize the physical entity data and engineering prior data in time to obtain synchronized data; the physical entity data includes pressure, flow rate, solid content and vibration; the engineering prior data includes sea state, pipeline geometry and pump station or valve status. The data fusion module 12 is used to perform feature engineering processing on the synchronized data to extract the corresponding dimensionless numbers and target coupling indices, and to perform multi-fidelity data fusion based on the dimensionless numbers and target coupling indices to obtain fused data; the dimensionless numbers include Reynolds numbers, Froude numbers, Stokes numbers, Shields numbers, and Begno numbers; The knowledge transfer module 13 is used to train the target model using the fused data to obtain a trained teacher model, and to transfer the knowledge of the trained teacher model to a preset student model based on knowledge distillation to obtain the target student model. The post-processed prediction model acquisition module 14 is used to optimize the target student model based on the loss function determined according to the preset physical constraints and the fused data to obtain an initial prediction model, perform concept drift detection on the initial prediction model, and if real-time working condition data drift is detected, perform feature distribution alignment and small step size fine-tuning on the initial prediction model to obtain the post-processed prediction model. The prediction data acquisition module 15 is used to perform meta-training on the processed prediction model to obtain the target prediction model, construct a target vector based on the dimensionless number, and input the target vector into the target prediction model to obtain the corresponding prediction data and the confidence intervals corresponding to the prediction data; the prediction data includes pressure drop along the friction, solid volume distribution, pump shaft power, erosion rate and blockage probability. The early warning module 16 is used to determine the conditional risk value of each of the predicted data and to issue a corresponding early warning based on the conditional risk value and the confidence interval corresponding to each of the predicted data.

[0061] Since the embodiments of the device part correspond to the embodiments described above, please refer to the embodiments described in the method part for the embodiments of the device part, and will not be repeated here.

[0062] In summary, this application first acquires the physical entity data and prior engineering data of the pipeline, and then synchronizes the physical entity data and prior engineering data in time to obtain synchronized data. The physical entity data includes pressure, flow rate, solid content, and vibration. The prior engineering data includes sea state, pipeline geometry, and pump station or valve status. Feature engineering processing is performed on the synchronized data to extract the corresponding dimensionless numbers and target coupling indices. Multi-fidelity data fusion is then performed based on the dimensionless numbers and target coupling indices to obtain fused data. The dimensionless numbers include Reynolds number, Froude number, Stokes number, Shields number, and Begno number. The target model is trained using the fused data to obtain a trained teacher model. Knowledge distillation is then used to transfer the knowledge of the trained teacher model to a preset student model to obtain the target student model. The target student model is optimized based on a loss function determined according to preset physical constraints and the fused data to obtain an initial prediction model. Concept drift detection is performed on the initial prediction model. If real-time operating data drift is detected, feature distribution alignment and small-step fine-tuning are performed on the initial prediction model to obtain a processed prediction model. Meta-training is performed on the processed prediction model to obtain a target prediction model. A target vector is constructed based on the dimensionless number and input into the target prediction model to obtain corresponding prediction data and confidence intervals for each prediction data. The prediction data includes pressure drop along the pump shaft, solid volume distribution, pump shaft power, erosion rate, and blockage probability. The conditional risk value of each prediction data is determined, and corresponding early warnings are issued based on the conditional risk value and the confidence intervals corresponding to each prediction data. As can be seen, this application ensures the rationality of predictions and the stability of extrapolation by embedding the physical laws governing multiphase flow as fundamental constraints into a deep learning framework; it utilizes real-time drift detection and adaptive learning mechanisms to enable the model to dynamically track changes in operating conditions and maintain long-term effectiveness; and it employs uncertainty quantification technology to provide confidence intervals for the prediction results of key indicators such as pressure drop, solid phase distribution, energy consumption, erosion rate, and blockage probability, thereby achieving an upgrade from "point prediction" to "probabilistic prediction." Ultimately, through high-precision prediction and risk-based early warning, the system provides decision support for optimizing operations, reducing energy consumption, and preventing unplanned downtime, thereby improving the safety and economy of the conveying process.

[0063] Furthermore, embodiments of this application also disclose an electronic device, Figure 5 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0064] Figure 5This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the data prediction and early warning method for long-distance transportation in deep-sea mining disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0065] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0066] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0067] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the data prediction and early warning method for long-distance deep-sea mining transportation disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0068] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned data prediction and early warning method for long-distance transportation in deep-sea mining. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0069] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0070] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0071] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0072] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0073] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A data prediction and early warning method suitable for long-distance transportation in deep-sea mining, characterized in that, include: The physical entity data and engineering prior data of the pipeline are acquired, and the physical entity data and engineering prior data are synchronized in time to obtain synchronized data; the physical entity data includes pressure, flow rate, solid content and vibration; the engineering prior data includes sea state, pipeline geometry and pump station or valve status. The synchronized data is subjected to feature engineering to extract the corresponding dimensionless numbers and target coupling indices. Multi-fidelity data fusion is then performed based on the dimensionless numbers and target coupling indices to obtain the fused data. The dimensionless numbers include Reynolds numbers, Froude numbers, Stokes numbers, Shields numbers, and Begno numbers. The target model is trained using the fused data to obtain a trained teacher model. The knowledge of the trained teacher model is then transferred to a preset student model based on knowledge distillation to obtain the target student model. The target student model is optimized based on the loss function determined according to the preset physical constraints and the fused data to obtain an initial prediction model. Concept drift detection is performed on the initial prediction model. If real-time working condition data drift is detected, feature distribution alignment and small step size fine-tuning are performed on the initial prediction model to obtain a processed prediction model. Meta-training is performed on the processed prediction model to obtain the target prediction model. A target vector is constructed based on the dimensionless number, and the target vector is input into the target prediction model to obtain the corresponding prediction data and the confidence intervals corresponding to the prediction data. The prediction data includes pressure drop along the pump shaft, solid volume distribution, pump shaft power, erosion rate, and blockage probability. Determine the conditional risk value for each of the predicted data, and issue corresponding early warnings based on the conditional risk value and the confidence interval corresponding to each of the predicted data.

2. The data prediction and early warning method for long-distance transportation in deep-sea mining according to claim 1, characterized in that, The feature engineering process performed on the synchronized data includes: Determine the observation vector of the synchronized data; The Kalman filter algorithm is used to repair missing data in the physical entity data based on the observation vector, and to correct abnormal observations in the observation vector to obtain corresponding processed data, which can then be used for feature engineering.

3. The data prediction and early warning method for long-distance transportation in deep-sea mining according to claim 1, characterized in that, The process of determining the target coupling index includes: The contact-hydraulic coupling index is determined by a preset formula; the preset formula for determining the contact-hydraulic coupling index is as follows: ； Wherein, TCCI is the contact-hydraulic coupling index; The modulus of contact force; For fluid force model; It is a constant; This represents the volume fraction of the solid phase. For reference volume fraction; The radial force-volume fraction coupling index is determined by a preset formula; the preset formula for determining the radial force-volume fraction coupling index is as follows: ； Wherein, RFVF is the radial force-volume fraction coupling index; This represents the average value of the radial force. This represents the average radial velocity. This represents the rate of change of radial force with volume fraction; This represents the radial gradient of the radial velocity.

4. The data prediction and early warning method for long-distance transportation in deep-sea mining according to claim 1, characterized in that, Before optimizing the target student model based on the loss function determined according to preset physical constraints and the fused data, the method further includes: The overall loss function for multi-task learning is determined by a pre-defined formula; the pre-defined formula for determining the overall loss function for multi-task learning is as follows: ； in, Let be the overall loss function for multi-task learning; k represents the different learning tasks. These are the true value vector and the model's predicted value vector for the k-th task, respectively. The heteroscedasticity uncertainty for the predicted k-th task; The conserved residual term is determined by a pre-defined formula; the pre-defined formula for determining the conserved residual term is as follows: ； in, For the conserved residual term; For divergence operators; The density of the mixture; The velocity vector field of the mixture; The square of the L2 norm; The monotonic residual term is determined by a preset formula; the preset formula for determining the monotonic residual term is as follows: ； in, The term is the monotonic residual; ReLU is the linear rectified function; The partial derivative of the pressure drop with respect to the flow rate Q; The loss function for multi-fidelity learning is determined by a pre-defined formula; the pre-defined formula for determining the loss function for multi-fidelity learning is as follows: ； in, The loss function for multi-fidelity learning; A mapping function represented by a neural network; This is the output of the high-fidelity model; The square of the L2 norm; This is the output of the low-fidelity model; For hyperparameters; For mapping functions The gradient.

5. The data prediction and early warning method for long-distance transportation in deep-sea mining according to claim 4, characterized in that, The optimization of the target student model based on a loss function determined according to preset physical constraints and the fused data to obtain an initial prediction model includes: The loss function is determined based on preset physical constraints; the formula for determining the loss function is: ； in, The loss function is... Let be the overall loss function for the multi-task learning; For weak-form PDE residual constraints; For the consistency residuals of the submerged boundary method; For the conserved residual term; For the monotonic residual term; Θ represents the L2 regularization term; Θ represents all trainable parameters of the model. These are the weighting hyperparameters corresponding to each loss or residual. The loss function for multi-fidelity learning; The target student model is optimized based on the loss function and the fused data to obtain an initial prediction model.

6. The data prediction and early warning method for long-distance transportation in deep-sea mining according to any one of claims 1 to 5, characterized in that, The provision of corresponding early warnings based on the conditional value of risk and the confidence intervals corresponding to each of the predicted data includes: If the conditional risk value does not meet the confidence interval corresponding to the predicted data, then a risk is determined to exist, and a warning light is used to issue an early warning.

7. The data prediction and early warning method for long-distance transportation in deep-sea mining according to claim 6, characterized in that, Also includes: Based on the model predictive control algorithm, the corresponding confidence intervals of the predictive data and the current risk situation, the corresponding data adjustment suggestions are output. The predicted data and the suggested adjustments are visualized through a human-computer interface.

8. A data prediction and early warning device suitable for long-distance transportation in deep-sea mining, characterized in that, include: The synchronized data acquisition module is used to acquire the physical entity data and engineering prior data of the pipeline, and synchronize the physical entity data and engineering prior data in time to obtain synchronized data; the physical entity data includes pressure, flow rate, solid content and vibration; the engineering prior data includes sea state, pipeline geometry and pump station or valve status. The data fusion module is used to perform feature engineering processing on the synchronized data to extract the corresponding dimensionless numbers and target coupling indices, and to perform multi-fidelity data fusion based on the dimensionless numbers and target coupling indices to obtain fused data; the dimensionless numbers include Reynolds numbers, Froude numbers, Stokes numbers, Shields numbers, and Begno numbers; The knowledge transfer module is used to train the target model using the fused data to obtain a trained teacher model, and to transfer the knowledge of the trained teacher model to a preset student model based on knowledge distillation to obtain the target student model. The post-processing prediction model acquisition module is used to optimize the target student model based on the loss function determined according to the preset physical constraints and the fused data to obtain an initial prediction model. Concept drift detection is performed on the initial prediction model. If real-time working condition data drift is detected, feature distribution alignment and small step size fine-tuning are performed on the initial prediction model to obtain the post-processing prediction model. The prediction data acquisition module is used to perform meta-training on the processed prediction model to obtain the target prediction model, construct a target vector based on the dimensionless number, and input the target vector into the target prediction model to obtain the corresponding prediction data and the confidence intervals corresponding to the prediction data; the prediction data includes pressure drop along the friction, solid volume distribution, pump shaft power, erosion rate, and blockage probability; The early warning module is used to determine the conditional risk value of each of the predicted data, and to issue corresponding early warnings based on the conditional risk value and the confidence interval corresponding to each of the predicted data.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing a computer program to implement the data prediction and early warning method for long-distance transportation in deep-sea mining as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program that, when executed by a processor, implements the data prediction and early warning method for long-distance transportation in deep-sea mining as described in any one of claims 1 to 7.

Images