[go: up one dir, main page]

CN119442921A - A marine digital twin optimization method and system based on multi-scale feature fusion - Google Patents

A marine digital twin optimization method and system based on multi-scale feature fusion Download PDF

Info

Publication number
CN119442921A
CN119442921A CN202510031039.8A CN202510031039A CN119442921A CN 119442921 A CN119442921 A CN 119442921A CN 202510031039 A CN202510031039 A CN 202510031039A CN 119442921 A CN119442921 A CN 119442921A
Authority
CN
China
Prior art keywords
digital twin
data
features
marine
feature
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202510031039.8A
Other languages
Chinese (zh)
Inventor
马元庆
苏博
由丽萍
李淑君
李志林
刘昕
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shandong Marine Resource and Environment Research Institute
Original Assignee
Shandong Marine Resource and Environment Research Institute
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shandong Marine Resource and Environment Research Institute filed Critical Shandong Marine Resource and Environment Research Institute
Priority to CN202510031039.8A priority Critical patent/CN119442921A/en
Publication of CN119442921A publication Critical patent/CN119442921A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/10—Pre-processing; Data cleansing
    • G06F18/15—Statistical pre-processing, e.g. techniques for normalisation or restoring missing data
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00—Details relating to CAD techniques
    • G06F2111/10—Numerical modelling
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Image Processing (AREA)

Abstract

The invention relates to the technical field of ocean data monitoring, in particular to an ocean digital twin optimization method and system based on multi-scale feature fusion. The method comprises the steps of obtaining marine environment multi-source data, carrying out data preprocessing on the obtained marine environment multi-source data, carrying out feature extraction and feature fusion on the multi-source data based on a digital twin model, optimizing the digital twin model based on the fused features, and carrying out marine environment multi-source data prediction by utilizing the optimized digital twin model. According to the invention, the digital twin closed loop system covering the whole life cycle is constructed by carrying out fusion modeling on multi-mode and multi-scale marine data, the defects existing in the prior art are overcome, the modeling precision, instantaneity and adaptability of marine digital twin are improved, and innovative technical support is provided for marine equipment and environment management.

Description

Ocean digital twin optimization method and system based on multi-scale feature fusion
Technical Field
The invention relates to the technical field of ocean data monitoring, in particular to an ocean digital twin optimization method and system based on multi-scale feature fusion.
Background
The ocean environment is complex and changeable, and the traditional test and development mode has the problems of low efficiency, high cost and high risk. Particularly in the research and development and test of marine equipment, the marine equipment is difficult to dynamically adapt to real sea conditions and environmental changes due to long-term dependence on a traditional physical model test, so that the development period is long and the iteration efficiency is low. In recent years, the rise of digital twin technology provides a new idea for solving the problem. Digital twinning supports full life cycle optimization management of equipment from design, manufacturing to use and maintenance by mapping the behavior and state of physical objects with virtual environments in real time. In particular, in a marine scene, digital twinning can realize efficient modeling and accurate simulation of a complex marine environment and equipment running states by integrating multi-source heterogeneous data and multi-scale information. The existing method can solve the problems to a certain extent, but still has the following defects that 1. The technical bottleneck of multi-scale feature fusion is shown, the data dimension in the marine environment is complex, the features span multiple scales such as a time domain, a space domain and the like, and the traditional method is difficult to effectively fuse the multi-scale features. 2. The real-time performance and the robustness are insufficient, the real-time performance of the existing model is low when large-scale ocean data are processed, the robustness to noise and abnormal environments is poor, and the model cannot adapt to dynamic changes in practical application. 3. The full life cycle integration is insufficient, the digital twin technology is applied to many researches in the design, manufacturing and operation stages, but the full-flow integration solution from data acquisition to feedback optimization is lacking.
At present, the marine digital twin data optimization has the following defects that 1. The multi-scale feature modeling is insufficient, the expression and fusion capability of the existing method on the multi-scale features of a time domain, a space domain and a semantic domain in a marine environment are insufficient, and a complex marine dynamic environment cannot be comprehensively and accurately described. 2. The real-time performance and the robustness are poor, and when facing complex ocean environment changes, the traditional model is difficult to ensure high-efficiency processing and has enough robustness, and is easy to be influenced by environmental noise and data loss. 3. The full life cycle integration is limited, the current technology is focused on a specific link, lacks a full life cycle integration framework covering the data acquisition to feedback optimization, and is difficult to provide closed-loop support for the design, manufacture, operation and maintenance of marine equipment. 4. The system has weak adaptability, and the optimization method aiming at different task demands lacks dynamic adjustment capability, so that the universality and adaptability of the system are poor, and the universality and flexibility of practical application are limited.
Disclosure of Invention
In order to solve the problems, the invention provides a marine digital twin optimization method and a system based on multi-scale feature fusion.
In a first aspect, the marine digital twin optimization method based on multi-scale feature fusion provided by the invention adopts the following technical scheme:
a marine digital twin optimization method based on multi-scale feature fusion comprises the following steps:
Acquiring marine environment multisource data;
carrying out data preprocessing on the acquired marine environment multisource data;
Carrying out feature extraction and feature fusion on the multi-source data based on the digital twin model;
Converting the fused features into uniform feature representations through dimension alignment;
Optimizing the digital twin model based on the fused features;
And predicting the marine environment multisource data by using the optimized digital twin model.
Further, the acquiring the marine environment multisource data includes performing multi-mode and multi-scale data acquisition by utilizing a mode of collaborative operation of a multi-sensor network, a fixed sensor, an unmanned aerial vehicle and a satellite multi-platform system, and time sequence data generated by the sensor is expressed as:
Wherein, Is the firstThe output of each sensor at time t, m being the characteristic dimension;
acquiring multi-channel images from multiple angles by using the unmanned aerial vehicle and the satellite, wherein the image data are expressed as follows:
where H is the image height, W is the image width, and C is the channel number.
Further, the data preprocessing of the acquired marine environment multisource data comprises the steps of respectively adopting a differential optimization strategy to time series data and image data so as to eliminate the difference between the data, wherein for the time series data, a linear interpolation method is applied to complement missing values, the continuity and the integrity of the data are ensured, then the characteristic values are adjusted to a uniform range through normalization processing, and for the preprocessing of the image data, firstly, the noise is removed through Gaussian filtering, and the definition and the stability of the image are improved.
Further, the method comprises the steps of extracting dynamic characteristics of time sequence data through an improved space-time convolution network ST-CNN to capture the change trend of the time sequence data in the time dimension, extracting multi-scale characteristics of image data by utilizing a median enhanced spatial channel attention network MECS to strengthen the expression capacity of spatial distribution, wherein the characteristics after enhancement are decomposed and extracted into characteristics with different resolutions, the characteristics are grouped according to channels to obtain characteristics of different channels, and finally, the time sequence characteristics and the image characteristics of different channels are weighted and fused through a double-time fusion network BFM to integrate the multi-mode data into a unified characteristic representation for modeling.
Further, the step of converting the fused features into unified feature representation through dimension alignment comprises the steps of mapping two modes into a unified feature space through a linear projection and broadcasting mechanism, wherein projection and expansion of time sequence features are performed first, image feature projection is performed, attention distribution is generated after feature stitching, the features of the two modes are subjected to weighted fusion according to attention weights, multi-scale convolution optimization is performed on fusion results, and local and global feature information is extracted.
Further, the method comprises optimizing a digital twin model based on the fused features, training the model by using a loss function consisting of a prediction error loss and a regularization term according to the fused features, wherein a Mean Square Error (MSE) is adopted by a prediction error part to evaluate the accuracy of a prediction result, the regularization term is introduced into the loss function to restrict the complexity of model parameters in order to prevent the model from losing generalization performance due to overfitting in the training process, and meanwhile, an Adam optimization algorithm is used to optimize the model parameters in a non-convex optimization problem by dynamically adjusting the learning rate and combining the accumulated information of first-order and second-order momentum, and gradient clipping is introduced on the basis of Adam to prevent the disturbance of gradient explosion to the training process.
Further, the method for predicting marine environment multisource data by using the optimized digital twin model comprises the steps of deploying the optimized digital twin model into an actual scene, monitoring and simulating in real time, and calculating errors by comparing target variables and observed values of model prediction, wherein for time series data, a threshold value is set to be 0.05 based on a mean square error MSE between the predicted values and the observed values, namely, an error square average value between the predicted values and the actual values is smaller than 5%, for image feature prediction, a threshold value is set to be 3% based on an average absolute difference MAE of pixel point errors, namely, the proportion of the pixel value errors to an image gray value range is smaller than 3%, and the calculated errors are expressed as:
wherein the target variable Observations of。
In a second aspect, a marine digital twin optimization system based on multi-scale feature fusion, comprising:
The data acquisition module is configured to acquire marine environment multi-source data;
The preprocessing module is configured to perform data preprocessing on the acquired marine environment multi-source data;
The feature module is configured to perform feature extraction and feature fusion on the multi-source data based on the digital twin model;
the conversion module is configured to convert the fused features into unified feature representations through dimension alignment;
An optimization module configured to optimize the digital twin model based on the fused features;
And the prediction module is configured to predict marine environment multisource data by using the optimized digital twin model.
In a third aspect, the present invention provides a computer readable storage medium having stored therein a plurality of instructions adapted to be loaded by a processor of a terminal device and to perform the method of marine digital twin optimization based on multi-scale feature fusion.
In a fourth aspect, the invention provides a terminal device, which comprises a processor and a computer readable storage medium, wherein the processor is used for realizing instructions, and the computer readable storage medium is used for storing a plurality of instructions, and the instructions are suitable for being loaded by the processor and executing the marine digital twin optimization method based on multi-scale feature fusion.
In summary, the invention has the following beneficial technical effects:
according to the invention, the digital twin closed loop system covering the whole life cycle is constructed by carrying out fusion modeling on multi-mode and multi-scale marine data, the defects existing in the prior art are overcome, the modeling precision, instantaneity and adaptability of marine digital twin are improved, and innovative technical support is provided for marine equipment and environment management.
Drawings
FIG. 1 is a schematic representation of the process of example 1 of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings.
Example 1
Referring to fig. 1, an ocean digital twin optimization method based on multi-scale feature fusion of the present embodiment includes:
Acquiring marine environment multisource data;
carrying out data preprocessing on the acquired marine environment multisource data;
Carrying out feature extraction and feature fusion on the multi-source data based on the digital twin model;
Converting the fused features into uniform feature representations through dimension alignment;
Optimizing the digital twin model based on the fused features;
And predicting the marine environment multisource data by using the optimized digital twin model.
Specifically, the method comprises the following steps:
S1, collecting and sensing multi-source data,
In a complex marine environment, multi-modal, multi-scale data are acquired using sensors, unmanned aerial vehicles and satellite remote sensing equipment. The data comprise temperature, flow rate, salinity physical time sequence data, high-definition local image data and global remote sensing image data, and the data represent dynamic marine environments and provide rich input data for subsequent steps.
S2, preprocessing and standardizing the data,
Preprocessing the collected data to ensure the consistency and quality of the multi-mode data. The time sequence data eliminates noise and abnormal value through missing value complement and normalization processing, and the image data improves the quality and diversity through denoising, standardization and data enhancement technology. This step establishes a reliable basis for subsequent feature modeling.
S3, constructing a digital twin model,
A digital twin model is constructed using a time space convolution network as a backbone network to simulate dynamic changes in marine environments and equipment operating conditions. Dynamic characteristics in time sequence data are extracted through a space-time convolution network, and the multi-mode data are converted into unified characteristic representation by combining multi-scale space characteristics of image data and adopting a fusion technology. The model provides core data support for training and optimization.
S4, training and optimizing the digital twin model,
Based on the fused characteristic data, training and optimizing the digital twin model. The error between the predicted value and the true value is measured by defining the loss function, and the model parameters are continuously adjusted by combining an optimization algorithm, so that the prediction precision and the robustness of the model are improved. The trained model can be efficiently adapted to complex dynamic environments.
S5, twin system application and feedback optimization,
And deploying the optimized digital twin model into an actual system for real-time monitoring and dynamic simulation. And starting a feedback mechanism through the error between the prediction result and the actual observation value output by the monitoring system, collecting new data, returning to the step S1, updating the model, and realizing closed-loop optimization of the data and the model. This step ensures high adaptability and real-time performance of the system to the dynamic environment.
The method comprises the following steps:
S1, time sequence data acquisition and image data acquisition are carried out by utilizing a multi-sensor network and a mode of collaborative operation of a fixed sensor, an unmanned aerial vehicle and a satellite multi-platform system to carry out multi-mode and multi-scale data acquisition. The fixed sensor is used for providing continuous and stable time sequence data, ensuring that the time sequence change of a dynamic environment is accurately captured, simultaneously, acquiring a high-resolution image from a local angle by using the unmanned aerial vehicle, supplementing the detailed description of a key area, and the satellite provides global background data through large-range remote sensing to make up the limitation of a single sensor. The design of the combination of multiple platforms ensures the diversity and the space coverage rate of the data. The time series data generated by the sensor can be expressed as:
Wherein, Is the firstThe output of each sensor at time t, m is the characteristic dimension.
Acquiring multi-channel images from multiple angles by using the unmanned aerial vehicle and the satellite, wherein the image data are expressed as follows:
where H is the image height, W is the image width, and C is the channel number.
And S2, preprocessing the acquired data, and respectively adopting different optimization strategies for time series data and image data to eliminate the difference between the data and improve the data quality.
For time series data, firstly, a linear interpolation method is applied to complement the missing value, so that the continuity and the integrity of the data are ensured:
Wherein, Is the firstFeatures at timeAnd (3) interpolation results.AndIs a known point in timeAndUpper firstAnd characteristic values.AndIs two known points in time, and t 1<t<t2.
And then, the characteristic values are adjusted to a uniform range through normalization processing, so that the subsequent model processing is facilitated:
Wherein, AndThe mean and standard deviation of the features, respectively.
For preprocessing of image data, firstly, noise is removed through Gaussian filtering, and definition and stability of the image are improved. The specific formula of Gaussian filtering is:
Wherein, Is the weight of the gaussian filter,,Is the abscissa offset of the pixel,Is the standard deviation of the gaussian distribution, which determines the degree of smoothing of the filter.
Next, to further enhance the image features, a spatial module of a median enhanced channel attention Module (MECS) is employed. First, the global feature of each channel is enhanced by the channel attention, and the calculation formula is as follows:
Wherein, For a1 x1 convolution operation,AndRepresenting global average pooling and maximum pooling operations respectively,The function is activated for Sigmoid. The formula enables dynamic weight allocation for channel importance.
Then, the spatial information expression capability of the image is enhanced by a spatial attention module, and the specific formula is as follows:
Wherein, Represents a channel dimension concatenation of the results of the global averaging and max pooling operations of the input feature map F,Is a 7 x 7 convolution operation for capturing local spatial context information.
The enhanced image features are expressed as:
wherein, Representing a pixel-by-pixel weighting operation,For enhanced image features.
Time sequence data after pretreatmentEnhancing image dataAs a next step input.
The multi-scale feature modeling and fusion of the S3 is implemented by firstly extracting dynamic features of time sequence data through an improved space-time convolution network (ST-CNN) so as to capture the change trend of the dynamic features in the time dimension:
Wherein, Is a convolution kernel of different scales, the present invention uses convolution kernels of 3 xc, 5 xc, 7 xc, C being the dimension of the feature, each convolution kernel being capable of capturing a dynamic feature of a particular time scale,Is a dynamic feature extracted by a time gating mechanism or a cyclic convolution layer that can guide the network to focus on long-term dependent features of the time series.
For enhanced image dataAnd carrying out multi-scale decomposition, and extracting features with different spatial resolutions. Will beGrouping according to channels:
Each group of features is then downsampled:
and then upsampling and feature recovery are carried out, and features with different resolutions are recovered to the original resolution:
Wherein, Representing a depth convolution, for extracting local features,Representing upsampling for recovering feature sizes.
Then the characteristics of different channels are spliced and fused, and then the characteristics of the different channels are passed throughConvolution reduces the number of channels and integrates different scale features:
Is the feature of the fused image.
And then will beThe features after multi-scale fusion are further optimized through the MECS module, namely, channel attention optimization is firstly carried out, and global information of feature channels is extracted through global average pooling, maximum pooling and median pooling:
Wherein, Representing a channel attention map, generating channel optimized features through global pooling and MLP is:
Wherein, Representing element-by-element multiplication for attention weighting.
And then carrying out space attention optimization, and capturing a space dependency relationship by using a multi-scale depth convolution module of MECS:
then pass through Convolution generates a spatial attention map:
Output characteristics As a final result of multi-scale feature extraction, the method has stronger robustness and expression capability:
Wherein, Representing element-by-element multiplication for attention weighting. Representation ofAnd finally fusing the optimized image features.
S3, converting the multi-mode data into uniform characteristic representation by adopting a fusion technology. Since the time series features and the representation forms of the image features are different (the time series features are vector forms, the image features are tensor forms), dimension alignment is required. Both modalities are mapped to a unified feature space by linear projection and broadcast mechanisms.
Firstly, performing projection and expansion of time sequence characteristics:
Wherein, The time sequence characteristics are represented by a sequence,The projection matrix is represented by a matrix of projections,The term of the bias is indicated,Timing characteristics mapped to a uniform dimension.
And then will beExpanded to the same spatial dimension as the image features:
Wherein, Representing broadcast operations, timing characteristicsBroadcast to spatial dimensions with image featuresConsistent, expanded into a three-dimensional tensor.
Then image feature projection is performed:
Wherein, The projection matrix is represented by a matrix of projections,The offset is indicated as being a function of the offset,Representing image features mapped to a uniform dimension.
And introducing an attention weight mechanism for dynamically balancing the contributions of the two modal characteristics, and calculating the weight distribution of the two modal characteristics.
The feature fusion weight is calculated, the attention distribution is generated after feature splicing,
Wherein, The fusion weight matrix is represented as a matrix of fusion weights,The term of the bias is indicated,Attention weights representing timing characteristics.
The weights of the image features are: ,
the features of the two modes are weighted and fused according to the attention weight:
Wherein, The characteristics after the fusion are represented,Representing an element-by-element multiplication operation.
To further enhance the expression of fusion features, the fusion results are subjected toAnd (3) performing multi-scale convolution optimization, and extracting local and global characteristic information.
Multi-scale convolution operation:
Where k=11 is used to extract global features and k=3 is used to extract local features.
Fusing global and local features:
Wherein, The operation of the splice is indicated and,A linear transformation matrix is represented and is represented,The term of the bias is indicated,Representing the final fusion characteristics.
The training and optimized core data information is obtained through the operation, and the obtained core data information is obtainedAnd optimizing by using a training loss function and an optimization algorithm.
In the model training and optimization of S4, in order to effectively improve the prediction precision and generalization capability of the digital twin model, a multi-level optimization strategy based on combination of a loss function design and an optimization algorithm is adopted.
First, in order to measure the deviation between the model predicted value and the true value, the present invention designs a comprehensive loss function. The loss function consists of prediction error loss and regularization terms. The prediction error part adopts a Mean Square Error (MSE) to evaluate the accuracy of a prediction result, and the formula is as follows:
Wherein, Is a predicted value of the current value,Is a true value of the code,Is the number of samples.
The predictive power of the model can be continuously optimized by calculating the minimum value of MSE. Meanwhile, in order to prevent the model from losing generalization performance due to overfitting in the training process, regularization terms are introduced into the loss function to constrain the complexity of model parameters. Specifically, L2 regularization is adopted, and the formula is as follows:
Wherein, As a parameter of the model, it is possible to provide,The regularization weight is used for controlling the influence degree of the regularization term on the loss.
The final loss function combines the two parts, and ensures that the model achieves the best effect between accuracy and robustness by balancing the prediction error and the model complexity.
In optimizing model parameters, adam optimization algorithm is used. Adam can show excellent performance in non-convex optimization problems by dynamically adjusting the learning rate in combination with the accumulated information of first and second order momentums. The parameter updating formula is as follows:
Wherein, As a parameter of the t-th step,In order for the rate of learning to be high,AndThe first order momentum and the second order momentum of the gradient respectively,To prevent a small constant with zero denominator.
In order to further enhance the training stability, a gradient clipping technology is introduced on the basis of Adam, so that the interference of gradient explosion to the training process is prevented. Gradient clipping is achieved by limiting the maximum norm of the gradient, which is formulated as:
Wherein, A threshold value is tailored for the pre-set gradient,Is the gradient norm. By clipping, the model can still maintain stable parameter updates in extreme cases.
In the training process, in order to ensure that the model can be converged to the optimal solution efficiently, the invention adopts a dynamic learning rate adjustment strategy. The learning rate is gradually reduced along with the training by using a learning rate attenuation (LEARNING RATE DECAY) method, and the formula is as follows:
Wherein, For the initial rate of learning to be the same,And t is the current training step number for the attenuation factor. Along with the training, the reduction of the learning rate can effectively prevent the model from vibrating when approaching to the optimal solution, and meanwhile, the convergence accuracy is improved.
Through the design, the optimized model not only has higher prediction precision, but also can keep robustness in a complex dynamic environment. After training, the model enters an actual deployment stage, and core support is provided for real-time monitoring and feedback optimization.
And S5, performing real-time monitoring and dynamic feedback optimization, and deploying the optimized digital twin model into an actual scene to perform monitoring and simulation in real time. Target variable predicted by system through modelAnd observed valueIs calculated error:
Feedback mechanism and closed loop optimization for time series data, a threshold value of 0.05 is set based on the Mean Square Error (MSE) between the predicted and observed values (i.e., the mean square error between the predicted and actual values is less than 5%). For image feature prediction, a threshold value of 3% (i.e., the ratio of pixel value error to image gray value range is less than 3%) is set based on the mean absolute difference (MAE) of pixel point errors. The system can dynamically adjust the threshold according to the quality of data collected in real time and the environment change condition in the running process. For example, in the case of high data noise levels, the threshold is suitably relaxed, while in mission critical scenarios (e.g., extreme weather warnings), the threshold is tightened to improve the accuracy of the prediction. When the error is When the threshold value is exceeded, the system starts a feedback mechanism, and the newly acquired data is returned to S1 for retraining, so that the model performance is continuously optimized, and the real-time response capability and adaptability of the model to environmental changes are ensured.
Example 2
The embodiment provides a marine digital twin optimizing system based on multi-scale feature fusion, which comprises the following components:
The data acquisition module is configured to acquire marine environment multi-source data;
The preprocessing module is configured to perform data preprocessing on the acquired marine environment multi-source data;
The feature module is configured to perform feature extraction and feature fusion on the multi-source data based on the digital twin model;
the conversion module is configured to convert the fused features into unified feature representations through dimension alignment;
An optimization module configured to optimize the digital twin model based on the fused features;
And the prediction module is configured to predict marine environment multisource data by using the optimized digital twin model.
A computer readable storage medium having stored therein a plurality of instructions adapted to be loaded by a processor of a terminal device and to perform the method of marine digital twin optimization based on multi-scale feature fusion.
A terminal device comprises a processor and a computer readable storage medium, wherein the processor is used for realizing instructions, and the computer readable storage medium is used for storing a plurality of instructions which are suitable for being loaded by the processor and executing the marine digital twin optimization method based on multi-scale feature fusion.
The above embodiments are not intended to limit the scope of the invention, so that the equivalent changes of the structure, shape and principle of the invention are covered by the scope of the invention.

Claims (10)

1.一种基于多尺度特征融合的海洋数字孪生优化方法，其特征在于，包括：1. A marine digital twin optimization method based on multi-scale feature fusion, characterized by comprising: 获取海洋环境多源数据；Obtain multi-source data on the marine environment; 对获取的海洋环境多源数据进行数据预处理；Carry out data preprocessing on the acquired multi-source data of marine environment; 基于数字孪生模型对多源数据进行特征提取和特征融合；Extract and fuse features from multi-source data based on the digital twin model; 将融合后的特征通过维度对齐转化为统一的特征表示；The fused features are transformed into a unified feature representation through dimension alignment; 基于融合的特征对数字孪生模型进行优化；Optimize the digital twin model based on the fused features; 利用优化后的数字孪生模型进行海洋环境多源数据预测。Use the optimized digital twin model to predict multi-source data of the marine environment. 2.根据权利要求1所述的一种基于多尺度特征融合的海洋数字孪生优化方法，其特征在于，所述获取海洋环境多源数据，包括利用多传感器网络和固定传感器、无人机、卫星多平台系统协同作业的方式进行多模态、多尺度数据采集，传感器生成的时间序列数据表示为：2. According to claim 1, a method for optimizing ocean digital twins based on multi-scale feature fusion is characterized in that the acquisition of multi-source data of the ocean environment includes multi-modal and multi-scale data acquisition by using a multi-sensor network and a collaborative operation of fixed sensors, unmanned aerial vehicles, and satellite multi-platform systems, and the time series data generated by the sensor is expressed as: 其中，为第个传感器在时间t的输出，m为特征维度；in, For the The output of a sensor at time t, m is the feature dimension; 利用无人机和卫星从多角度采集多通道图像，图像数据表示为：Using drones and satellites to collect multi-channel images from multiple angles, the image data is represented as: 其中，H为图像高度，W为图像宽度，C为通道数。Among them, H is the image height, W is the image width, and C is the number of channels. 3.根据权利要求2所述的一种基于多尺度特征融合的海洋数字孪生优化方法，其特征在于，所述对获取的海洋环境多源数据进行数据预处理，包括分别对时间序列数据和图像数据采用差异性优化策略，以消除数据之间的差异性，其中，对于时间序列数据，应用线性插值法对缺失值进行补全，确保数据的连续性和完整性，再通过归一化处理将特征值调整到统一范围；对于图像数据的预处理，首先通过高斯滤波去除噪声，提升图像的清晰度和稳定性。3. According to claim 2, a marine digital twin optimization method based on multi-scale feature fusion is characterized in that the acquired marine environment multi-source data is preprocessed, including adopting difference optimization strategies for time series data and image data respectively to eliminate the differences between the data, wherein, for time series data, linear interpolation is used to fill in missing values to ensure the continuity and integrity of the data, and then the eigenvalues are adjusted to a uniform range through normalization; for the preprocessing of image data, Gaussian filtering is first used to remove noise to improve the clarity and stability of the image. 4.根据权利要求3所述的一种基于多尺度特征融合的海洋数字孪生优化方法，其特征在于，所述基于数字孪生模型对多源数据进行特征提取和特征融合，包括通过改进的时空卷积网络ST-CNN提取时间序列数据的动态特征，以捕捉其在时间维度上的变化趋势，利用中值增强空间通道注意力网络MECS提取图像数据的多尺度特征，以强化其对空间分布的表达能力，其中，对增强后的特征进行分解提取不同分辨率的特征，并按照通道进行分组，得到不同通道的特征，最后通过双时融合网络BFM对不同通道的时序特征和图像特征进行加权融合，将多模态数据整合为统一的特征表示，用于建模。4. According to claim 3, a marine digital twin optimization method based on multi-scale feature fusion is characterized in that the feature extraction and feature fusion of multi-source data are performed based on the digital twin model, including extracting the dynamic features of time series data through the improved spatiotemporal convolutional network ST-CNN to capture its changing trend in the time dimension, and using the median enhanced spatial channel attention network MECS to extract the multi-scale features of image data to enhance its expression ability of spatial distribution, wherein the enhanced features are decomposed to extract features of different resolutions, and grouped according to channels to obtain features of different channels, and finally the time series features and image features of different channels are weightedly fused through the dual-time fusion network BFM, and the multimodal data are integrated into a unified feature representation for modeling. 5.根据权利要求4所述的一种基于多尺度特征融合的海洋数字孪生优化方法，其特征在于，所述将融合后的特征通过维度对齐转化为统一的特征表示，包括通过线性投影和广播机制将两种模态映射到统一的特征空间，其中，先进行时序特征的投影与扩展，再进行图像特征投影，通过特征拼接后生成注意力分布，将两种模态的特征按注意力权重进行加权融合，并对融合结果进行多尺度卷积优化，提取局部和全局的特征信息。5. According to claim 4, a marine digital twin optimization method based on multi-scale feature fusion is characterized in that the fused features are converted into a unified feature representation through dimensional alignment, including mapping the two modalities to a unified feature space through linear projection and broadcasting mechanism, wherein the time series features are projected and expanded first, and then the image features are projected, and the attention distribution is generated after feature splicing, the features of the two modalities are weightedly fused according to the attention weights, and the fusion results are subjected to multi-scale convolution optimization to extract local and global feature information. 6.根据权利要求5所述的一种基于多尺度特征融合的海洋数字孪生优化方法，其特征在于，所述基于融合的特征对数字孪生模型进行优化，包括根据融合后的特征，利用由预测误差损失和正则化项组成的损失函数训练模型，其中，预测误差部分采用均方误差MSE来评估预测结果的准确性，为了防止模型在训练过程中因过度拟合而丧失泛化性能，引入正则化项到损失函数中，以对模型参数的复杂度进行约束；同时使用Adam优化算法，通过动态调整学习率，结合一阶和二阶动量的累积信息，在非凸优化问题中优化模型参数，并在Adam的基础上引入了梯度裁剪，防止梯度爆炸对训练过程的干扰。6. According to claim 5, a marine digital twin optimization method based on multi-scale feature fusion is characterized in that the digital twin model is optimized based on the fused features, including training the model according to the fused features using a loss function consisting of a prediction error loss and a regularization term, wherein the prediction error part uses the mean square error MSE to evaluate the accuracy of the prediction result, and in order to prevent the model from losing generalization performance due to overfitting during the training process, a regularization term is introduced into the loss function to constrain the complexity of the model parameters; at the same time, the Adam optimization algorithm is used to dynamically adjust the learning rate and combine the accumulated information of the first-order and second-order momentum to optimize the model parameters in the non-convex optimization problem, and gradient clipping is introduced on the basis of Adam to prevent the gradient explosion from interfering with the training process. 7.根据权利要求6所述的一种基于多尺度特征融合的海洋数字孪生优化方法，其特征在于，所述利用优化后的数字孪生模型进行海洋环境多源数据预测，包括将优化后的数字孪生模型部署到实际场景中，实时进行监控与仿真，并通过模型预测的目标变量与观测值的对比计算误差，其中对于时间序列数据，基于预测值与观测值之间的均方误差 MSE，设定阈值为0.05，即预测值与真实值之间的误差平方平均值小于 5%；对于图像特征预测，基于像素点误差的平均绝对差 MAE，设定阈值为 3%，即像素值误差占图像灰度值范围的比例小于 3%，计算误差表示为：7. According to claim 6, a marine digital twin optimization method based on multi-scale feature fusion is characterized in that the use of the optimized digital twin model to predict multi-source data of the marine environment includes deploying the optimized digital twin model to the actual scene, monitoring and simulating in real time, and calculating the error by comparing the target variable predicted by the model with the observed value, wherein for time series data, based on the mean square error MSE between the predicted value and the observed value, the threshold is set to 0.05, that is, the square average of the error between the predicted value and the true value is less than 5%; for image feature prediction, based on the mean absolute difference MAE of pixel point error, the threshold is set to 3%, that is, the proportion of pixel value error to the image gray value range is less than 3%, and the calculated error is expressed as: 其中，目标变量 ，观测值。Among them, the target variable , observed value . 8.一种基于多尺度特征融合的海洋数字孪生优化系统，其特征在于，包括：8. An ocean digital twin optimization system based on multi-scale feature fusion, characterized by comprising: 数据获取模块，被配置为，获取海洋环境多源数据；The data acquisition module is configured to acquire multi-source data of the marine environment; 预处理模块，被配置为，对获取的海洋环境多源数据进行数据预处理；The preprocessing module is configured to perform data preprocessing on the acquired marine environment multi-source data; 特征模块，被配置为，基于数字孪生模型对多源数据进行特征提取和特征融合；The feature module is configured to perform feature extraction and feature fusion on multi-source data based on the digital twin model; 转化模块，被配置为，将融合后的特征通过维度对齐转化为统一的特征表示；The conversion module is configured to convert the fused features into a unified feature representation through dimension alignment; 优化模块，被配置为，基于融合的特征对数字孪生模型进行优化；The optimization module is configured to optimize the digital twin model based on the fused features; 预测模块，被配置为，利用优化后的数字孪生模型进行海洋环境多源数据预测。The prediction module is configured to use the optimized digital twin model to predict multi-source data of the marine environment. 9.一种计算机可读存储介质，其中存储有多条指令，其特征在于，所述指令适于由终端设备的处理器加载并执行如权利要求1所述的方法。9. A computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded by a processor of a terminal device and executing the method as claimed in claim 1. 10.一种终端设备，包括处理器和计算机可读存储介质，处理器用于实现各指令；计算机可读存储介质用于存储多条指令，其特征在于，所述指令适于由处理器加载并执行如权利要求1所述的方法。10. A terminal device, comprising a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, characterized in that the instructions are suitable for being loaded by the processor and executing the method according to claim 1.
CN202510031039.8A 2025-01-09 2025-01-09 A marine digital twin optimization method and system based on multi-scale feature fusion Pending CN119442921A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202510031039.8A CN119442921A (en) 2025-01-09 2025-01-09 A marine digital twin optimization method and system based on multi-scale feature fusion

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202510031039.8A CN119442921A (en) 2025-01-09 2025-01-09 A marine digital twin optimization method and system based on multi-scale feature fusion

Publications (1)

Publication Number Publication Date
CN119442921A true CN119442921A (en) 2025-02-14

Family

ID=94505983

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202510031039.8A Pending CN119442921A (en) 2025-01-09 2025-01-09 A marine digital twin optimization method and system based on multi-scale feature fusion

Country Status (1)

Country Link
CN (1) CN119442921A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119903758A (en) * 2025-03-31 2025-04-29 山东超华环保智能装备有限公司 A digital twin modeling method for environmental protection equipment based on multi-source data fusion
CN119940161A (en) * 2025-04-08 2025-05-06 中国科学院地理科学与资源研究所 Twin intelligent simulation method and device for coral reef beach sand evolution process
CN120070778A (en) * 2025-04-28 2025-05-30 自然资源部南海海域海岛中心(自然资源部南海标准计量与信息中心) Ocean environment real-time monitoring modeling system and method based on multi-mode fusion
CN120631191A (en) * 2025-08-18 2025-09-12 浙江弄潮儿智慧科技有限公司 A digital evolution method and system for an ecosystem
CN120748425A (en) * 2025-09-03 2025-10-03 中国海洋大学 UUV self-noise suppression method based on dual-channel collaborative noise reduction neural network

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118193982A (en) * 2024-03-30 2024-06-14 连云港鲸鸣信息科技有限公司 Ocean monitoring-oriented multitasking and perception decision-making large model
CN118195057A (en) * 2024-02-05 2024-06-14 通辽水文水资源分中心 Flood forecasting system and forecasting method based on digital twin technology
WO2024253782A1 (en) * 2023-06-03 2024-12-12 Istari Digital, Inc. Digital twin enhancement using external feedback within integrated digital model platform
CN119129421A (en) * 2024-09-11 2024-12-13 南方海洋科学与工程广东省实验室(珠海) A marine digital twin platform

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2024253782A1 (en) * 2023-06-03 2024-12-12 Istari Digital, Inc. Digital twin enhancement using external feedback within integrated digital model platform
CN118195057A (en) * 2024-02-05 2024-06-14 通辽水文水资源分中心 Flood forecasting system and forecasting method based on digital twin technology
CN118193982A (en) * 2024-03-30 2024-06-14 连云港鲸鸣信息科技有限公司 Ocean monitoring-oriented multitasking and perception decision-making large model
CN119129421A (en) * 2024-09-11 2024-12-13 南方海洋科学与工程广东省实验室(珠海) A marine digital twin platform

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119903758A (en) * 2025-03-31 2025-04-29 山东超华环保智能装备有限公司 A digital twin modeling method for environmental protection equipment based on multi-source data fusion
CN119940161A (en) * 2025-04-08 2025-05-06 中国科学院地理科学与资源研究所 Twin intelligent simulation method and device for coral reef beach sand evolution process
CN120070778A (en) * 2025-04-28 2025-05-30 自然资源部南海海域海岛中心(自然资源部南海标准计量与信息中心) Ocean environment real-time monitoring modeling system and method based on multi-mode fusion
CN120631191A (en) * 2025-08-18 2025-09-12 浙江弄潮儿智慧科技有限公司 A digital evolution method and system for an ecosystem
CN120631191B (en) * 2025-08-18 2025-11-07 浙江弄潮儿智慧科技有限公司 A digital evolution method and system for ecosystems
CN120748425A (en) * 2025-09-03 2025-10-03 中国海洋大学 UUV self-noise suppression method based on dual-channel collaborative noise reduction neural network

Similar Documents

Publication Publication Date Title
CN119442921A (en) A marine digital twin optimization method and system based on multi-scale feature fusion
CN112347859B (en) A method for salient object detection in optical remote sensing images
CN118298327B (en) Mamba-based North sea ice concentration prediction method
CN113469094A (en) Multi-mode remote sensing data depth fusion-based earth surface coverage classification method
CN115147731A (en) A SAR Image Target Detection Method Based on Full Spatial Coding Attention Module
CN112686207A (en) Urban street scene target detection method based on regional information enhancement
CN115223009A (en) Small target detection method and device based on improved YOLOv5
CN114550014A (en) Road segmentation method and computer device
CN113989612A (en) Remote sensing image target detection method based on attention and generation countermeasure network
CN117876349A (en) A method for predicting photovoltaic power generation
CN118410327A (en) A sea wave forecasting method, electronic device and medium for flat sea areas
Patil et al. Semantic segmentation of satellite images using modified U-Net
CN116912646B (en) Multimode remote sensing image fusion classification method based on double-branch dynamic modulation network
CN117220276A (en) Ultra-short-term photovoltaic power prediction method and system considering the interaction of spatiotemporal characteristics
CN120279437B (en) Remote sensing open-air mining area detection method based on double-branch structure and feature fusion mechanism
CN120070858B (en) Light-weight target detection system integrating attention mechanism, optimizing feature fusion method and self-supervision learning
KR102515608B1 (en) Photovoltaic forecasting method and system
CN118968317B (en) Blue algae distribution graph generation method and device based on diffusion model and server
CN120654567A (en) Method, device, equipment and medium for reconstructing high space-time resolution refined flow field
CN119598895A (en) Method and system for inverting ocean three-dimensional temperature structure based on sea surface data
CN117911701A (en) Light-weight double-prediction branch semantic segmentation water body extraction deep learning method and system
CN116977866A (en) A lightweight landslide detection method
Wang et al. Sensor data fusion optimization in UAV integrated navigation based on matrix factorization
CN113706650A (en) Image generation method based on attention mechanism and flow model
CN120047450B (en) Remote sensing image change detection method based on double-current time-phase characteristic adapter

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination