CN-121980953-A - Moving body cruising motion performance prediction model, method, device and system based on large model and parameterization technology

CN121980953ACN 121980953 ACN121980953 ACN 121980953ACN-121980953-A

Abstract

The invention discloses a model, a method, a device and a system for predicting cruising motion performance of a moving body based on a large model and a parameterization technology, wherein the moving body comprises a ship, an airplane and a vehicle. The core is that a multitask model for predicting cruising motion performance of a motion body is built through cloud pre-training (general+specific motion body samples, fusion simulation, real and test data) and edge end fine adjustment deployment. In the pre-training process, a task mask mechanism is adopted to realize the prediction of the multi-task performance parameters, and the Newton first law constraint guarantee is embedded to accord with the physical law. The scheme can be used for bi-directionally predicting the single-value relationship, is suitable for navigation guidance of specific moving bodies, route planning and concept design of non-specific moving bodies, solves the problems that the traditional method is difficult to fit a high-dimensional nonlinear relationship, the conventional AI technology is poor in generalization and the like, and can improve prediction accuracy and application efficiency.

Inventors

FENG BO

Assignees

上海交策源科技发展有限公司
上海太科舟智能科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260205

Claims (10)

1. A method for training and deploying a large model for predicting cruising motion performance of a moving body is characterized by comprising the following steps: The cloud end acquires a general moving body training sample set, which comprises the steps of respectively acquiring a motion simulation sample, a test sample and a real motion sample from a general moving body motion simulation system, a moving body model test and an actual navigation scene of a real moving body, and cleaning, sorting and classifying the acquired samples to form the general moving body training sample set; The cloud pre-training a model for predicting and learning the cruising motion thrust resistance of a moving body comprises the steps of pre-training the model for predicting and learning the cruising motion thrust resistance of the moving body by using a neural network based on a training sample set of a general moving body motion simulation system comprising a thrust value and a resistance value, and establishing a constraint frame for meeting Newton's first law for a general moving body cruising motion performance prediction learning big model and a specific moving body cruising motion performance prediction learning big model prediction result; the cloud pre-training general moving body cruising movement performance prediction learning big model comprises the steps of pre-training a designed and constructed general moving body cruising movement performance prediction learning big model based on a general moving body training sample set, wherein the learning big model comprises a prediction network and a physical constraint network, multi-task prediction is realized through a task masking mechanism, and a prediction result is ensured to meet Newton first law constraint through the physical constraint network and a corresponding loss function in the training process; The cloud enhanced specific moving body cruising movement training sample set comprises a real movement sparse sample of a specific moving body based on a diffusion model technology, a denoising sampling model is trained, an enhanced sample close to a real scene is generated, and the enhanced sample is stored in a real training sample set buffer of the specific moving body; The cloud pre-training a specific moving body cruising motion performance prediction learning large model comprises the steps of customizing a specific moving body cruising motion performance prediction learning large model by freezing parameters related to specific moving body characteristics in the general moving body cruising motion performance prediction learning large model, and pre-training the specific moving body learning large model based on a training sample set in a specific moving body real training sample set buffer; The cloud verification of the specific moving body cruising motion performance prediction big model comprises the steps of adjusting the pre-trained specific moving body cruising motion performance prediction learning big model structure to be a specific moving body cruising motion performance prediction big model, and using a real sample for verification; Migrating the edge end model, namely migrating the specific moving body cruising motion performance prediction learning large model and the specific moving body cruising motion performance prediction large model which pass verification to a navigation brain system of the specific moving body edge end to construct a moving body edge end cruising motion performance prediction system; the edge end real-time prediction comprises a prediction system for the cruising motion performance of an edge end moving body, wherein the prediction system calls a specific moving body cruising motion performance prediction big model to predict cruising motion performance parameters according to an input or transmitted prediction task, power parameter information and real-time information of the moving body and environment; The method comprises the steps of monitoring a cruising motion sample newly collected by an edge, when a fine tuning condition is met, fine tuning the migrated specific moving body cruising motion performance prediction learning large model by using the new sample, and obtaining an updated prediction large model based on the fine-tuned learning large model structure, replacing the original prediction large model of the edge, and realizing continuous optimization of the model.
2. A method for predicting cruising motion performance of a moving body, comprising: acquiring environmental parameters, dynamic parameters and characteristic parameters of a moving body; And executing a plurality of prediction tasks by using a pre-trained moving body cruising motion performance prediction big model based on the environmental parameters, the power parameters and the moving body characteristic parameters, wherein the moving body cruising motion performance prediction big model distinguishes different prediction tasks through a task masking mechanism, the moving body cruising motion performance prediction big model is a neural network model designed according to the requirements of the invention by referring to a hybrid expert MoE architecture and is used for completing a plurality of prediction tasks through the same model structure, the prediction tasks are the predicted moving body cruising motion performance parameters, the moving body cruising motion performance parameters comprise a horizontal drift speed, a navigational speed and a power parameter affecting the horizontal drift speed and the navigational speed, the power parameter comprises an energy consumption rate and a power device rotating speed, the performance parameters predicted by the prediction tasks comprise one or more parameter combinations of the horizontal drift speed, the navigational speed, the energy consumption rate and the rotating speed, the environmental parameters are data perceived in real time by a moving body 'hydrological weather system' and a sensor transmitted by a 'sensor router', or the moving body 'is called by an automatic party' the moving body's power line' and the power parameter is called by the dynamic system or the dynamic parameter is acquired by the dynamic parameter or the dynamic parameter is input by the dynamic parameter.
3. The method for predicting cruise performance of a moving body according to claim 2, wherein the performing a plurality of prediction tasks using the pre-trained moving body cruise performance prediction big model comprises: the input information comprises environmental parameters, power parameters, moving body characteristic parameters and prediction task information; Inputting input information into an input layer, the input layer identifying each parameter by a physical quantity and representing the parameter predicted by the prediction task by a mask; the input preprocessing module normalizes the information output by the input layer and performs linear projection or feature extraction; The feature embedding module performs embedding projection on the output information input into the preprocessing module to obtain a feature embedding body; The multi-head self-attention layer performs normalization processing according to the output information of the feature embedding module, calculates multi-head attention, performs residual connection, and captures the coupling relation between multiple variables; the router of the feedforward network layer identifies a prediction task according to the mask representation, normalizes the multi-head attention layer output information according to the prediction task, routes the multi-head attention layer output information to a corresponding expert module feedforward network to carry out nonlinear mapping and residual connection, wherein each expert module corresponds to one prediction task; Predicting a multi-layer stacking structure designed by a large model according to cruising motion performance of a moving body, and completing multi-layer circulating multi-head attention calculation and feedforward network mapping; And the regression output layer carries out regression output on the output of the expert module to output the cruise motion performance parameters of the moving body of the prediction task.
4. A method for training a large model for predicting and learning the cruising motion performance of a moving body is characterized by comprising the following steps: Acquiring a training sample set of a moving body, wherein the training sample set comprises a motion simulation sample, a real motion sample and a model test sample; Based on the training sample set, a moving body cruising motion performance prediction large model is pre-trained, wherein the moving body cruising motion performance prediction large model comprises a prediction network and a physical constraint network, the prediction network is designed according to the requirements of the invention by referring to a hybrid expert MoE architecture, the physical constraint network is a moving body cruising motion thrust resistance prediction learning model for completing the pre-training, and the moving body cruising motion thrust resistance prediction learning model is designed according to the requirements of the invention by referring to neural network models such as MLP (not limited to MLP); Finishing a plurality of prediction tasks through the prediction network, and ensuring that a prediction result meets Newton first law constraint through the physical constraint network; optimizing the moving body cruise motion performance prediction learning large model parameters based on a plurality of loss functions, wherein the loss functions comprise a prediction task loss function and a physical constraint loss function.
5. The moving body cruising motion performance prediction learning large model training method according to claim 4, characterized in that the pre-training moving body cruising motion performance prediction learning large model based on the training sample set includes: inputting a batch of training samples into an input layer of the large learning model, wherein the input layer carries out physical quantity identification on each parameter in a sample set; the input preprocessing module normalizes the information output by the input layer and performs linear projection or feature extraction; the feature embedding module performs embedding mapping on the information output by the input preprocessing module to obtain a feature embedded body; The multi-head attention layer carries out normalization processing according to the output information of the feature embedding module, calculates multi-head attention and carries out residual connection so as to learn the complex coupling relation among multiple variables; the prediction task traversing module traverses all the prediction tasks, carries out mask processing on vectors related to the prediction tasks from the multi-head attention layer output, and embeds the prediction tasks into the input of the feedforward network layer; The router performs normalization processing on the data blocks after mask processing according to the prediction task, guides the data blocks to the corresponding expert modules to map nonlinear relations and performs residual connection; according to the multi-layer stacking structure designed by the prediction learning large model, multi-layer circulating multi-head attention calculation and feedforward network mapping are completed; the regression output layer outputs the feedforward network layer to regress and output the cruise motion performance parameters of the moving body of the prediction task; the input updating layer of the physical constraint network updates items corresponding to the training samples by using the performance values predicted by the prediction network, and the physical constraint network predicts the thrust resistance regression values based on the training samples updated by the input updating layer; Calculating a total loss function, and optimizing model parameters based on a random gradient descent method until the training ending condition is met; After the general moving body cruising movement performance prediction learning big model is trained, a general moving body cruising movement performance prediction big model is built, a specific moving body cruising movement performance prediction learning big model is built according to the characteristic parameter freezing part parameter of the specific moving body, and after the specific moving body cruising movement performance prediction learning big model is trained, the specific moving body cruising movement performance prediction big model is built.
6. The method for training a model for predicting cruising motion performance of a moving body as claimed in claim 5, wherein the loss function comprises: a first loss function based on a mean square error calculation between the predicted task output value and the sample value; a second loss function for calculating a mean square error between the thrust predictor and the drag predictor based on the Newton first law constraint; the third loss function is based on a thrust monotonicity penalty function, so that thrust monotonicity change along with the energy consumption rate or the rotating speed of the propeller is ensured; a fourth loss function, based on a resistance monotonicity penalty function, ensuring that the resistance monotonically changes along with the speed of the moving body; a fifth loss function, based on thrust boundary conditions, ensuring zero thrust at zero energy consumption rate or zero propeller rotational speed; A sixth loss function that ensures zero drag at zero speed based on drag boundary conditions; the total loss function is a weighted sum of the first loss function, the second loss function, the third loss function, the fourth loss function, the fifth loss function, and the sixth loss function.
7. The model and method for pre-training the model according to claim 1, wherein the model comprises an input layer, a feature extractor, a thrust predictor and a resistance predictor, and the method comprises: Reading batch training samples to an input layer, and inputting sample data to a feature extractor by the input layer; The feature extractor extracts features from the training sample data; The thrust predictor and the resistance predictor respectively predict the thrust and the resistance in a regression way; And calculating a loss function, and optimizing model parameters based on a random gradient descent method until the training ending condition is met.
8. A cruise motion performance prediction model fine tuning and distillation method, comprising: The motion performance parameter monitoring module in the cruising motion performance predictor detects that a latest collected cruising motion sample exists in the navigation data storage medium; The motion performance parameter monitoring module carries out mask processing on corresponding components of the latest cruising motion sample according to a prediction task, then inputs the corresponding components into a motion performance prediction big model, and the motion performance prediction big model outputs a motion performance parameter prediction result; Verifying the motion performance parameter prediction result, and if the motion performance parameter prediction result meets the preset fine adjustment condition, adopting a fine adjustment technology mature by a third party to carry out fine adjustment on the cruising motion performance prediction learning large model; And after verification, covering the motion performance prediction big model in the cruise motion performance predictor by using the new cruise motion performance prediction big model to realize updating and optimizing of the model.
9. A moving body edge cruise performance prediction system according to claim 1, comprising: The system comprises a cruise motion performance predictor, a navigation data storage medium, a fine-tuning distillation module, a cruise motion performance prediction learning large model, a cruise motion performance prediction large model and a navigation data man-machine interaction page, and is connected with an Internet access device, a sensor router, a motion body route automatic planning system, a hydrological weather recognition system, a reporting data subsystem and a navigation brain server side display; the cruising motion performance predictor is used for predicting motion performance parameters according to input or transmitted prediction task and power parameter information and read moving bodies and environment information thereof; the navigation data storage medium is used for storing the hydrological meteorological elements collected, identified and predicted by the system and real-time and historical data of the navigation of the moving body; The fine adjustment distillation module is used for fine adjustment of the cruising motion performance prediction learning large model and customizing the cruising motion performance prediction large model; The cruising motion performance prediction learning large model is transferred to an edge end by a specific moving body cruising motion performance prediction learning large model pre-trained by a cloud end for fine adjustment based on a real environment sample; The cruising motion performance prediction large model is customized by the cruising motion performance prediction learning large model adjustment structure with fine adjustment completed and is used for replacing a motion performance parameter prediction module in the cruising motion performance predictor; and the navigation data man-machine interaction page is used for inputting manual parameters and prediction tasks and checking predicted cruising motion performance parameters.
10. The moving body cruise motion performance prediction system according to claim 9, characterized in that the cruise motion performance predictor includes: the motion performance parameter monitoring module is used for sensing the real-time motion performance of the moving body through the sensor, and calling the prediction module to perform performance prediction according to the current environmental condition so as to monitor the accuracy of the motion performance prediction in real time; the motion performance interaction module is used for responding to the prediction task information input through a human-computer interaction interface or transmitted by an external calling party, calling the prediction module to predict the motion performance parameters and returning or displaying a prediction result; The motion performance parameter prediction module is used for performing prediction of the cruising motion performance parameters of the moving body according to the input task information and the environmental condition information, wherein the motion performance parameter prediction module is constructed by a cruising motion performance prediction large model which completes fine adjustment and distillation; The system management module is used for managing system parameters such as super parameters, wherein the super parameters comprise minimum navigational speed, maximum rotational speed, capsizing condition parameters and the like.

Description

Moving body cruising motion performance prediction model, method, device and system based on large model and parameterization technology Technical Field The invention relates to the technical field of artificial intelligence and motion performance prediction of a moving body, in particular to a model, a method, a device and a system for predicting cruising motion performance of a moving body based on a large model and a parameterization technology. Background Cruising ability of moving bodies (including ships, airplanes, vehicles) is an important performance index-the field of ships is called rapidity and the field of vehicles is called dynamic. For a long time, the industry mainly establishes the relation between cruising motion performance indexes and other physical quantities through methods such as tests, empirical formulas or computational fluid dynamics, for example, the corresponding relation between the rotating speed and the navigational speed of a ship host is measured in a speed measuring field of specific hydro-meteorological conditions, the empirical formulas of the relation between waves, flows and navigational speeds are established according to navigational statistics data, and a simulation model of the nonlinear relation among elements such as cruising motion performance, fuel consumption rate, host rotating speed, a moving body, hydro-meteorological environment and the like is established by utilizing the computational fluid dynamics method. However, the traditional method has the obvious limitations that the speed measuring field cannot reproduce complex and changeable ship conditions, load conditions and hydrometeorological conditions in actual sailing, the experimental formula deviates from a true value due to large manual observation errors, simple regression statistics tools and no consideration of specific moving bodies and other reasons, and the computational fluid mechanics is used for reference, so that a plurality of factors are needed to be abandoned for establishing an abstract mechanical model, and therefore, the high-dimensional nonlinear relation is difficult to accurately fit, and the limitations lead to larger cruising performance prediction result errors of the moving bodies and become bottlenecks for limiting technical breakthroughs of industries. With the development of artificial intelligence technology, the technical methods such as deep learning, large models and the like provide a new path for fitting complex nonlinear relations, and related application attempts also appear in the field of motion body cruising motion performance prediction. The method is characterized in that the vehicle field has patents for determining elements influencing the vehicle performance through a neural network, the ship field part adopts a convolutional neural network proxy model for predicting ship resistance and hydrostatic force values, but the method does not consider environment high-dimensional complexity and ship self condition change, relies on traditional software and tests to obtain samples to cause result distortion, has limited generalization and is not suitable for operating ships, the aircraft field has patents for predicting flight states based on long-short time memory network modeling, but the network is more suitable for short-time domain sequence modeling, and environmental interference and physical law constraint are not considered, so that the prediction drift risk exists. The existing artificial intelligence technologies can not effectively solve the core problems of sample sparseness, out-of-distribution prediction, commonality knowledge learning and the like in the prediction of the cruising motion performance of a moving body. In general, the cruising performance of a moving body is influenced by multi-factor coupling, and all factors are in the same motion state space, so that a technical scheme capable of learning common knowledge and meeting the requirement of multiple purposes is needed. The traditional method and the AI related technology cannot establish a complex high-dimensional nonlinear relation of cruising movement of a moving body, cannot learn common knowledge, and cannot meet the requirements of scenes such as actual navigation guidance, route planning or concept design of the moving body. Therefore, the high-simulation-degree cruise motion simulation environment of the moving body, which is based on parameterization and large model technology and accords with the physical law, is constructed, the cruise motion performance is predicted, and the method has important industrial application value. Disclosure of Invention The invention provides a large model training deployment method for predicting cruising motion performance of a moving body, which comprises the steps of collecting a general moving body training sample set by a cloud end, collecting a motion simulation sample, a test sample and a real motion sample from a general moving body motion s