Machine learning-based rod channel cooling liquid drop behavior analysis method and equipment

Abstract

The invention provides a rod channel cooling liquid drop behavior analysis method based on machine learning, which comprises the following steps of: acquiring image data of a nuclear reactor fuel rod channel cooling liquid drop movement process, and temperature, pressure, liquid drop flow velocity distribution and liquid drop concentration in a channel; using a model trained based on machine learning to analyze coolant drop behaviors, including coolant drop behavior classification, short-term behavior prediction and trend analysis; the constructed model architecture is a hybrid model fusing a convolutional neural network, a recurrent neural network and a graph neural network. According to the method, the flowing state, the interaction and the phase change behavior of the cooling liquid drops can be classified in detail and accurately predicted, the probability value of each classification result is provided, and a more comprehensive and accurate decision basis is provided for operation and maintenance personnel.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear reactor thermal-hydraulics, and particularly to a method and device for analyzing the behavior of coolant droplets in a rod channel based on machine learning. Background Art

[0002] During the operation of a nuclear reactor, coolant flows through the rod channels around the fuel rods in the reactor core, thereby effectively removing the heat generated by the nuclear reactor. Therefore, analyzing the future dynamic behavior of coolant droplets is of great significance for the nuclear reactor cooling system.

[0003] Currently, the prediction research on the behavior of coolant droplets in rod channels mostly relies on conventional experimental data processing methods, such as using regression analysis in statistical methods to fit curves to a large amount of experimental data. The above methods cannot deeply mine the hidden information in droplet behavior data, are difficult to construct a high-precision behavior prediction model, and cannot meet the requirements of accurate prediction and safety determination of coolant droplet behavior in complex situations such as the nuclear reactor cooling process. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the present invention proposes a method for analyzing the behavior of coolant droplets in a rod channel based on machine learning to solve the problem that the existing prediction research methods for the behavior of coolant droplets in rod channels cannot deeply mine the hidden information in droplet behavior data and are difficult to obtain high-precision prediction results.

[0005] The technical solution adopted by the present invention is as follows:

[0006] In a first aspect, a method for analyzing the behavior of coolant droplets in a rod channel based on machine learning is provided, including the following steps:

[0007] Collect image data of the movement process of coolant droplets in the fuel rod channels of the nuclear reactor, as well as the temperature, pressure, droplet flow rate, and droplet concentration in the channels;

[0008] According to the coolant droplet image data and the temperature, pressure, droplet flow rate distribution, and droplet concentration in the channels, use a model trained based on machine learning to analyze the behavior of coolant droplets.

[0009] Further, the analysis of the behavior of coolant droplets includes: behavior classification, short-term behavior prediction, and trend analysis.

[0010] Further, the behavior classification includes:

[0011] Whether the phase flow pattern will change, whether it changes from a small number of discrete small droplets to an inverse annular flow, slug flow;

[0012] Whether the droplets will collide, aggregate, or break in the future;

[0013] Whether evaporation and condensation of the droplet will occur in the future.

[0014] Furthermore, the short-term behavior prediction and trend analysis include:

[0015] Predict the change trend of the droplet velocity within a certain period of time in the future, and calculate the numerical range of the droplet velocity change;

[0016] Calculate the change of the droplet position coordinates in the future time, and predict whether the droplet will deviate from the normal trajectory or collide with the channel wall or other droplets.

[0017] Furthermore, before analyzing the behavior of the coolant droplet using the model trained based on machine learning, feature extraction and preprocessing are performed on the coolant droplet image data, including:

[0018] First, apply the Canny edge detection algorithm and morphological processing technology to the image data to determine the droplet contour boundary, calculate the geometric features such as the diameter and perimeter of the droplet, and use Hu moments to obtain invariant moment features to describe the droplet shape;

[0019] Based on the optical flow algorithm, calculate the droplet displacement and velocity vector field and combine with PIV to optimize the velocity calculation. At the same time, perform gray histogram and LBP texture analysis to obtain gray and texture features.

[0020] Furthermore, for the model trained based on machine learning, the model architecture is a hybrid model architecture that integrates a convolutional neural network, a recurrent neural network, and a graph neural network.

[0021] Furthermore, the convolutional neural network contains 3 convolutional layers, 2 pooling layers, and 2 fully connected layers, and outputs the extracted image feature vectors for fusion with the recurrent neural network and the graph neural network;

[0022] The recurrent neural network selects a long short-term memory artificial neural network, which contains 2 LSTM layers. The first LSTM layer is used to capture the long-term dependencies in the time series data through the gating mechanism; the second LSTM layer is used to process the time series information, and the output of the second LSTM layer is used as the modeling result of the dynamic change of the coolant droplet behavior;

[0023] The graph neural network contains 3 graph convolutional layers. The first graph convolutional layer is used to construct edge weights according to the distance information between droplets and perform preliminary learning on the collective behavior patterns of droplets; the second graph convolutional layer is used to learn the interaction relationships between droplets; the third graph convolutional layer is used to output a vector expressing the collective behavior characteristics of droplets for feature fusion with the convolutional neural network and the recurrent neural network.

[0024] Furthermore, splicing and fusing the output features of the convolutional neural network, recurrent neural network, and graph neural network includes:

[0025] Concatenate the image feature vectors of the convolutional neural network, the time series feature vectors of the recurrent neural network, and the group behavior feature vectors of the graph neural network in terms of dimensions to obtain fused feature vectors;

[0026] Perform dimensionality reduction processing through a fully connected layer to compress and integrate the fused features.

[0027] Furthermore, present the prediction results obtained from analyzing the behavior of coolant droplets in a visual manner, including: plotting the droplet velocity-time curve and the position-time trajectory diagram.

[0028] In a second aspect, an electronic device is provided, including: one or more processors; a storage device for storing one or more programs; when the one or more programs are executed by the one or more processors, enabling the one or more processors to implement the machine learning-based analysis method for the behavior of coolant droplets in a rod channel according to the first aspect.

[0029] As can be seen from the above technical solutions, the beneficial technical effects of the present invention are as follows:

[0030] 1. The analysis method for the behavior of coolant droplets in a rod channel obtains comprehensive and detailed droplet behavior information through multi-modal data acquisition and advanced feature extraction techniques, and accurately identifies various behavior patterns through in-depth analysis of the hybrid model. During the model training process, a large amount of coolant droplet behavior data covering various working conditions is collected, and a reasonable stratified sampling method is used to divide the data set, ensuring the diversity and representativeness of the training data. The Adam optimization algorithm is selected and combined with strategies such as learning rate decay, mini-batch gradient descent method, and early stopping method, which can dynamically adjust the model parameters according to the loss of the validation set during the training process, effectively avoiding overfitting and improving the generalization ability of the model. At the same time, visualization tools are used for dynamic analysis of the training, which is convenient for timely adjusting the model structure and parameters to ensure that the model performance reaches the optimal.

[0031] 2. In terms of behavior analysis and prediction, it can classify and accurately predict the flow state, interaction, and phase change behavior of coolant droplets in detail, and provide the probability value of each classification result, providing more comprehensive and accurate decision-making basis for the operation and maintenance personnel. At the same time, by visually presenting the prediction results such as velocity and position, the operation and maintenance personnel can intuitively understand the future dynamics of the coolant droplets, facilitating the timely discovery of potential problems and taking measures, which helps to improve the safety and stability of the nuclear reactor cooling system. Description of the Drawings

[0032] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0033] Figure 1 Schematic flow diagram of the method for analyzing the behavior of coolant droplets in the rod channel according to an embodiment of the present invention. Specific embodiments

[0034] The following will describe in detail the embodiments of the technical solutions of the present invention in conjunction with the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, so they are only examples and cannot be used to limit the protection scope of the present invention.

[0035] It should be noted that unless otherwise specified, the technical terms or scientific terms used in this application should have the ordinary meaning understood by those skilled in the art to which the present invention belongs.

[0036] Embodiment

[0037] This embodiment provides a method for analyzing the behavior of coolant droplets in a rod channel based on machine learning, including the following steps:

[0038] Step S1: Collect image data of the movement process of coolant droplets in the fuel rod channel of the nuclear reactor, as well as the temperature, pressure, droplet flow rate and droplet concentration in the channel

[0039] In this embodiment, a data acquisition system is used to arrange a high-resolution, high-frame-rate high-speed camera array, high-precision sensors, and a laser scattering concentration meter around the rod channel of the nuclear reactor to collect the movement process of coolant droplets and physical parameters such as the temperature, pressure and flow rate distribution in the channel; at the same time, a high-speed and stable data acquisition and transmission network is built to ensure the data transmission rate and the temporal consistency of data from different data sources, and the data is transmitted to the processing center. In a specific implementation manner, the selection of sensors and the transmission network is not limited, and any implementable manner in the prior art can be used.

[0040] Step S2: Extract and preprocess the features of the coolant droplet image data

[0041] In some embodiments, optionally, before the subsequent model analyzes the behavior of coolant droplets, the image data collected in step S1 is first subjected to feature extraction and preprocessing, so as to improve the model training and prediction speed while retaining key information, and lay a foundation for the subsequent model analysis.

[0042] During feature extraction and preprocessing, first, for the image data, the Canny edge detection algorithm and morphological processing technology are used to determine the contour boundary of the droplet, calculate geometric features such as the diameter and perimeter of the droplet, and use Hu moments to obtain invariant moment features to describe the droplet shape. Then, based on the optical flow algorithm, the droplet displacement and velocity vector field are calculated, and combined with PIV (Particle Image Velocimetry) technology to optimize the velocity calculation. At the same time, gray histogram and LBP (Local Binary Pattern) texture analysis are performed to obtain gray and texture features

[0043] Step S3: According to the coolant droplet image data, as well as the temperature, pressure, droplet flow rate distribution, and droplet concentration in the channel, use the model trained based on machine learning to analyze the coolant droplet behavior

[0044] 1. Construction of the model

[0045] To achieve accurate prediction of the coolant droplet behavior, in this embodiment, a hybrid model architecture integrating Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Graph Neural Network (GNN) is constructed as the rod channel coolant behavior analysis model, hereinafter referred to as the model. The construction method of the model is as follows

[0046] Among them, the CNN contains 3 convolutional layers, 2 pooling layers, and 2 fully connected layers. The first convolutional layer uses 32 convolutional kernels of size 3x3, with a stride of 1 and a ReLU activation function, and is used to extract the preliminary local features of the droplet image. After the first convolutional layer, the data size becomes (original size - 2) x (original size - 2) x 32. Then there is a max pooling layer with a pooling kernel size of 2x2 and a stride of 2, which downsamples the feature map, making the data size become (original size / 2 - 1) x (original size / 2 - 1) x 32. The second convolutional layer uses 64 3x3 convolutional kernels, with the same stride and activation function as the first layer, to further extract more complex local features, and the output size is (original size / 2 - 3) x (original size / 2 - 3) x 64. After another max pooling layer with the same pooling kernel and stride, the data size becomes (original size / 4 - 2) x (original size / 4 - 2) x 64. The third convolutional layer uses 128 3x3 convolutional kernels, with operations similar to the previous two layers, and the output feature map size is (original size / 4 - 4) x (original size / 4 - 4) x 128. Then the feature map after convolution and pooling is flattened and passed through the first fully connected layer, which has 256 neurons and a ReLU activation function, and is used to integrate the feature information. The last fully connected layer has 128 neurons, and outputs the image feature vector extracted by the CNN, preparing for the subsequent fusion with the RNN and GNN

[0047] The RNN adopts the LSTM (Long Short-Term Memory artificial neural network) structure and contains 2 LSTM layers. The first LSTM layer has 64 hidden units. The input data is a combination of sensor data and the image feature vectors output by the CNN. It captures the long-term dependencies in the time series data through the gating mechanism, and the output size is (sequence length, 64). The second LSTM layer has 32 hidden units, receives the output of the first LSTM layer, further processes the time series information, and the output size is (sequence length, 32). The final output is used as the modeling result of the dynamic change of the coolant droplet behavior for subsequent analysis and prediction.

[0048] The GNN contains 3 graph convolutional layers. The first graph convolutional layer has 32 channels. It constructs edge weights based on information such as the distance between droplets and conducts preliminary learning on the group behavior patterns of droplets, with the output feature dimension being 32. The second graph convolutional layer has 64 channels and further deeply learns the interaction relationships between droplets, with the output feature dimension being 64. The third graph convolutional layer has 128 channels. After the graph convolutional operation of this layer, it outputs a vector that can fully express the group behavior characteristics of droplets for feature fusion with the CNN and RNN.

[0049] The output features of the CNN, RNN, and GNN are concatenated and fused. First, the 128-dimensional image feature vector of the CNN, the 32-dimensional time series feature vector of the RNN, and the 128-dimensional group behavior feature vector of the GNN are concatenated in dimension to obtain a 288-dimensional fused feature vector. Then, dimensionality reduction is performed through a fully connected layer. This fully connected layer has 64 neurons, and the activation function is ReLU. It further compresses and integrates the fused features, and the final output is used as the prediction result of the model.

[0050] Meanwhile, collect no less than [X] groups of coolant droplet behavior data covering various faults and extreme conditions in the nuclear reactor rod channels, and stratify and sample them according to the ratio of 70:20:10 into a training set, a validation set, and a test set. During the training process, the Adam optimization algorithm is selected. The initial learning rate is set to 0.001 and decays to 0.9 times the original every 10 training cycles according to the loss value of the validation set. The mini-batch gradient descent method is used to iterate and train with a batch size of 32. The gradients are calculated according to the corresponding loss function of the classification or regression task to update the parameters. Closely monitor the performance metrics of the validation set. If there is no improvement for 5 consecutive cycles, the early stopping method is used, and a visualization tool is used to analyze the training dynamics and effects. Continuously adjust the model structure and parameters until the model reaches satisfactory performance metrics on the test set to ensure that the model has good generalization and prediction capabilities.

[0051] The hybrid model that integrates the advantages of CNN, RNN, and GNN trained by the above method fully exploits the unique capabilities of each network in processing image features, time series data, and droplet interaction relationships. The integration of the three realizes multi-modal in-depth analysis of the behavior of coolant droplets. Compared with a single model or a simple combined model, this architecture can more accurately model the complex behavior of coolant droplets and improve the accuracy and reliability of predictions.

[0052] 2. Analysis of Coolant Droplet Behavior

[0053] In this embodiment, the analysis of the behavior of coolant droplets is mainly divided into two aspects. One is behavior classification, and the other is short-term behavior prediction and trend analysis, which are specifically as follows:

[0054] (1) Behavior Classification

[0055] The collected and preprocessed coolant droplet data is input into the trained model in a predetermined format and chronological order. The model first performs feature extraction and fusion processing on the input data, and then through forward propagation calculation, outputs the behavior classification result of the coolant droplets. For the flow state of the coolant droplets, based on the RNN's ability to model time series data and the image features extracted by the CNN in the model, the model analyzes information such as the current speed, position of the coolant droplets, and the temperature, pressure, and flow rate distribution in the channel. Including:

[0056] By learning the historical speed and temperature data through RNN, combined with the current pressure gradient and channel geometry, predict the movement trend and distribution changes of the coolant droplets in the short term in the future (such as in the next few seconds to minutes), and then judge whether the two-phase flow pattern will change, whether it changes from a small number of discrete small droplets to categories such as inverse annular flow and slug flow. The model takes into account the critical conditions for the transition between different flow patterns, which are obtained through learning a large amount of data during the training process, including the comprehensive influence of factors such as temperature, pressure, droplet flow rate, and droplet concentration.

[0057] For the interaction between coolant droplets, using the interaction rules between droplets learned by the GNN and the current distribution state of the droplets, the model predicts whether phenomena such as collision, aggregation, or fragmentation of droplets will occur in the future. If the model detects multiple droplets in close proximity and with a relatively large relative velocity, according to the learned collision probability model, it predicts that they may collide, and further analyzes the possible results after the collision, whether they will aggregate to form larger droplets or break into smaller droplets, and the impact of these changes on the overall flow and heat transfer is classified into categories such as normal collision, aggregation and fusion, fragmentation and dispersion, etc.; for the phase change behavior, according to the model's prediction of the phase change behavior of droplets, combined with the principles of heat transfer and thermodynamics, when it is detected that the temperature around the droplets increases and the vapor concentration increases, combined with the previously learned evaporation condition model, it is predicted that the droplets may evaporate, and the evaporation rate and the impact on the temperature and pressure of the surrounding environment are estimated. Conversely, when the temperature decreases and the vapor may reach a saturated state, it is predicted that the droplets may condense, and the effect of the condensation process on the flow and heat transfer is analyzed. It is classified into categories such as evaporation, condensation, etc. At the same time, the model outputs the probability value of each classification result, and determines the final behavior category according to a preset threshold (such as a probability greater than 0.5 is determined as the corresponding category), providing an intuitive and clear judgment of the coolant droplet behavior for the operation and maintenance personnel to timely detect signs of abnormal behavior.

[0058] (2) Short-term Behavior Prediction and Trend Analysis

[0059] Velocity Prediction: Based on the powerful processing ability of the RNN part in the model for time series data, especially the LSTM or GRU structure can effectively capture the long-term dependence relationships in historical data. During the training phase, the model learned the patterns of the coolant droplet velocity changing over time under different operating conditions, such as the velocity change rules under temperature gradient changes, pressure fluctuations, and the influence of channel geometry. When the current coolant droplet data is input, the model combines the learned historical velocity change trends and the current environmental parameters (such as temperature, pressure, channel position, etc.), and through internal neuron calculations and weight adjustments, predicts the change trend of the droplet velocity in the short term in the future (such as the next few seconds to minutes). For example, if the current droplet is in a region with increasing temperature and the flow velocity has an increasing trend, the model will refer to the learning results of past similar situations, predict that the velocity may continue to increase, and give a specific numerical range of the velocity change.

[0060] Position Prediction: Considering that the movement of coolant droplets in the channel is affected by various factors, the model comprehensively utilizes the velocity vector, current position information, and geometric feature information of the channel in the input data. During the training phase, the model learned the movement trajectory patterns of droplets in various channel shapes (such as straight segments, bends, pipe diameter change segments, etc.) under different initial positions and velocity conditions. By learning and memorizing these patterns, when the data of the coolant droplet at the current moment is input, the model calculates the change in the position coordinates of the droplet at future time steps based on the current position and velocity of the droplet, combined with the specific geometric shape and physical parameters of the channel, using the principles of physical kinematics and the prediction ability of the machine learning model, so as to draw the position-time trajectory diagram of the droplet and predict whether it will deviate from the normal trajectory or collide with the channel wall or other droplets.

[0061] In some embodiments, the prediction results are presented to the operation and maintenance personnel in a visual manner, by plotting the droplet velocity-time curve and position-time trajectory diagram, enabling them to intuitively understand the future dynamics of the coolant droplets, make preparations in advance, and ensure the safe and stable operation of the nuclear reactor cooling system.

[0062] The method for analyzing the behavior of coolant droplets in rod channels provided in this embodiment obtains comprehensive and detailed droplet behavior information through multi-modal data acquisition and advanced feature extraction techniques, and accurately identifies various behavior patterns through in-depth analysis of the hybrid model. During the model training process, a large amount of coolant droplet behavior data covering various working conditions is collected, and a reasonable stratified sampling method is used to divide the data set, ensuring the diversity and representativeness of the training data. The Adam optimization algorithm is selected and combined with strategies such as learning rate decay, mini-batch gradient descent method, and early stopping method, which can dynamically adjust the model parameters according to the loss of the validation set during the training process, effectively avoid overfitting, and improve the generalization ability of the model. At the same time, a visualization tool is used for dynamic analysis of the training, which is convenient for timely adjusting the model structure and parameters to ensure that the model performance reaches the optimal.

[0063] In terms of behavior analysis and prediction, it can classify and accurately predict the flow state, interaction, and phase change behavior of coolant droplets in detail, and provide the probability value of each classification result, providing more comprehensive and accurate decision-making basis for the operation and maintenance personnel. At the same time, by visually presenting the prediction results such as velocity and position, the operation and maintenance personnel can intuitively understand the future dynamics of the coolant droplets, facilitating the timely discovery of potential problems and taking measures, which helps to improve the safety and stability of the nuclear reactor cooling system.

[0064] In some embodiments, an electronic device is further provided, including: one or more processors; a storage device for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the method for analyzing the behavior of coolant droplets in rod channels based on machine learning described above.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present invention, and they should all be covered within the scope of the claims and the description of the present invention.

Claims

1. A rod channel cooling droplet behavior analysis method based on machine learning, characterized in that: The following steps are involved: Collect image data of the motion process of cooling droplets in the fuel rod channel of a nuclear reactor, as well as the temperature, pressure, droplet velocity distribution and droplet concentration in the channel; Based on the cooling droplet image data as well as the temperature, pressure, droplet velocity distribution and droplet concentration in the channel, the cooling droplet behavior is analyzed using a model trained based on machine learning.

2. The rod channel cooling droplet behavior analysis method based on machine learning according to claim 1 is characterized in that: The analysis of cooling droplet behavior includes: behavior classification, short-term behavior prediction and trend analysis.

3. The rod channel cooling droplet behavior analysis method based on machine learning according to claim 2 is characterized in that: The behavior classification includes: Will the phase flow pattern change, from a small number of discrete droplets to an anti-annular flow or slug flow? whether the droplets will collide, aggregate, or break up in the future; Whether the droplets will evaporate or condense in the future.

4. The rod channel cooling droplet behavior analysis method based on machine learning according to claim 2 is characterized in that: The short-term behavior prediction and trend analysis include: Predict the change trend of droplet velocity within a certain period of time in the future and calculate the value range of droplet velocity change; Calculate the change in the position coordinates of the droplet at future times and predict whether the droplet will deviate from the normal trajectory or collide with the channel wall or other droplets.

5. The rod channel cooling droplet behavior analysis method based on machine learning according to claim 1 is characterized in that: Before using the machine learning-based trained model to analyze the behavior of cooling droplets, the cooling droplet image data is subjected to feature extraction and preprocessing, including: The Canny edge detection algorithm and morphological processing technology are first used to determine the droplet contour boundary of the image data, calculate the diameter and circumference geometric characteristics of the droplet, and use the Hu moment to obtain the invariant moment feature to describe the droplet shape; The droplet displacement and velocity vector field are calculated based on the optical flow algorithm and the velocity calculation is optimized in combination with PIV. At the same time, grayscale histogram and LBP texture analysis are performed to obtain grayscale and texture features.

6. The rod channel cooling droplet behavior analysis method based on machine learning according to claim 1 is characterized in that: The model is trained based on machine learning, and the model architecture is a hybrid model architecture that integrates convolutional neural networks, recurrent neural networks, and graph neural networks.

7. The rod channel cooling droplet behavior analysis method based on machine learning according to claim 6 is characterized in that: The convolutional neural network includes 3 convolutional layers, 2 pooling layers and 2 fully connected layers, and outputs the extracted image feature vectors for fusion with the recurrent neural network and the graph neural network; The recurrent neural network uses a long short-term memory artificial neural network, which includes two LSTM layers. The first LSTM layer is used to capture the long-term dependency in the time series data through a gating mechanism; the second LSTM layer is used to process the time series information, and the output of the second LSTM layer is used as the modeling result of the dynamic change of the cooling droplet behavior; The graph neural network includes three graph convolutional layers. The first graph convolutional layer is used to construct edge weights based on the distance information between droplets and perform preliminary learning on the behavior pattern of droplet groups. The second graph convolutional layer is used to learn the interaction relationship between droplets. The third graph convolutional layer is used to output a vector expressing the characteristics of droplet group behavior, which is integrated with the features of the convolutional neural network and the recurrent neural network.

8. The rod channel cooling droplet behavior analysis method based on machine learning according to claim 7 is characterized in that: The output features of the fused convolutional neural network, recurrent neural network and graph neural network are spliced ​​and fused, including: The image feature vector of the convolutional neural network, the time series feature vector of the recurrent neural network, and the group behavior feature vector of the graph neural network are concatenated in dimension to obtain a fused feature vector; The dimensionality reduction is performed through a fully connected layer to compress and integrate the fusion features.

9. The rod channel cooling droplet behavior analysis method based on machine learning according to claim 1, characterized in that: The predicted results obtained from analyzing the cooling droplet behavior are presented in a visual way, including: plotting the droplet velocity-time curve and position-time trajectory.

10. An electronic device, characterized in that: include: one or more processors; A storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the rod channel cooling droplet behavior analysis method based on machine learning as described in any one of claims 1-9.

Images