A multi-dimensional information parallel targeting extraction method based on spatial characteristics of droplets / bubbles of multiphase flow

Abstract

A kind of multi-dimensional information parallel targeting extraction method based on multiphase flow droplet / bubble spatial characteristics, relate to a kind of multiphase flow liquid multi-dimensional information extraction method, the method includes establishing the convolutional neural network model based on self-improvement YOLOv11, the high deformation degree, large overlap, large angle of deviation multi-dimensional complex bubble / droplet image under turbulent state is as data set;Further, the bubble / droplet image is automatically oriented, size scaling, gray scale conversion, cropping and isolation object and other data preprocessing procedures;According to the different information contained in image, different backbone quantity is set, and the features are accurately extracted using attention mechanism;In addition, the extracted information is combined and transmitted to the next layer for subsequent convolution reorganization;The application embeds the C2DAT (Cross Stage Partial Network with Deformable Attention Transformer) module in each backbone channel, achieves the goal of improving information extraction performance and reducing computational complexity, and makes the model more lightweight to adapt to low-power environment.

Description

TECHNICAL FIELD

[0001] The present application relates to a multi-phase flow liquid multi-dimensional information extraction method, in particular to a multi-dimensional information parallel targeted extraction method based on the spatial characteristics of multi-phase flow droplets / bubbles. BACKGROUND

[0002] Gas-liquid / liquid-liquid two-phase flow is widely used in petroleum and chemical industry, nuclear power generation, wastewater treatment, mass transfer equipment and pipeline industrial applications. Gas-liquid / liquid-liquid two-phase flow has important practical significance in petroleum and chemical industry, especially in the fields of petroleum cracking, natural gas treatment and chemical synthesis. Petroleum cracking process is a key step in petroleum and chemical industry. Through the mass transfer and heat transfer characteristics of gas-liquid / liquid-liquid two-phase flow, the cracking efficiency of hydrocarbon molecules can be effectively improved. The high-efficiency contact of catalyst and gas-liquid mixture can strengthen the reaction rate of hydrocarbon compounds in high-temperature and high-pressure environment, thereby producing more high-value-added light oil products. The optimized application of gas-liquid / liquid-liquid two-phase flow in these processes can not only improve the utilization rate of crude oil, but also reduce energy consumption and production cost, which meets the demand of the current petroleum and chemical industry towards high efficiency, green and low carbon.

[0003] In the gas-liquid / liquid-liquid two-phase flow system, the parameters of bubbles / droplets are crucial for the characterization of the system characteristics. For example, the shape, surface area and size of bubbles / droplets are affected by the flow pattern. Therefore, high-precision bubble / droplet recognition can more accurately evaluate a gas-liquid / liquid-liquid two-phase flow system. Convolutional neural networks with high universality and robustness can accurately identify bubbles / droplets, obtain bubble / droplet parameters and track the process of bubble / droplet coalescence and breakup in complex gas-liquid / liquid-liquid systems. Therefore, the parameters obtained by the neural network play an important role in optimizing the design of the mixer and obtaining the stability of the gas-liquid / liquid-liquid system.

[0004] The features of complex images may need a deep network for processing, and the deeper network has large computation, and the training and inference are more time-consuming. Training a large-scale CNN requires high-performance hardware support, especially a GPU or TPU with sufficient memory. The problem of gradient vanishing or explosion in deep network will affect the training stability and convergence speed, especially when dealing with complex tasks. In complex images, details may be hidden in the background, and the model needs to have stronger resolution capability. Bubble / droplet images have multiple poses, perspectives and shape deformations, increasing the difficulty of recognition. And bubbles / droplets appear at different scales, and traditional CNN may not be able to adapt.

[0005] Therefore, a neural network design method is needed to extract complex multi-dimensional data and make the model more lightweight. SUMMARY

[0006] The purpose of this invention is to provide a parallel targeted extraction method for multidimensional information based on the spatial features of multiphase flow droplets / bubbles. This method effectively solves the problems of high computational cost, long training time, vanishing or exploding gradients, difficulty in extracting feature details, and various deformations and scale inconsistencies in the recognition of complex bubble / droplet images. This method employs a multidimensional backbone lightweight neural network for optimization. The multidimensional backbone network extracts image features hierarchically, capturing local texture information at shallow layers and aggregating global semantic features at deep layers, effectively handling targets of different scales and deformations. Simultaneously, the lightweight design uses depthwise separable convolution, pruning, and quantization techniques to significantly reduce computational costs, decrease hardware dependence, and improve inference efficiency. Furthermore, combining the residual structure ResNet and a dynamic feature fusion mechanism can alleviate the vanishing gradient problem, improving model stability and convergence speed. This method improves model resolution while balancing fine-grained feature extraction and computational efficiency in complex scenes, providing an efficient solution for complex image recognition.

[0007] The technical solution adopted in this invention is as follows:

[0008] A parallel target extraction method based on multidimensional information of multiphase flow droplets / bubbles, comprising multidimensional information-based target dynamic detection and prediction, and overall model lightweighting: the specific steps are as follows:

[0009] Step 1: Use bubble / droplet images under different working conditions as the dataset, with 70%~80% used for the training set and 30%~20% used for the validation set;

[0010] Step 2: Calibrate the bubbles / droplets in the image, and then perform automatic orientation, resizing, grayscale conversion, cropping, and object isolation operations on the image;

[0011] Step 3: When extracting bubble / droplet information in three-dimensional space, it is first necessary to clarify the types of bubble / droplet information, such as size, shape, distribution location, number density, and trajectory. Different types of bubble / droplet information usually correspond to different research needs. For example, research on dynamic characteristics may focus on trajectory and velocity information, while research on structural characteristics may focus more on size and shape. Once the types of bubble / droplet information to be extracted are determined, the number of backbones needs to be designed according to the complexity of the information and the data dimensions. Different backbones can be used to collect one or more different types of information, thereby achieving comprehensive coverage and efficient analysis of bubble / droplet characteristics. When designing backbones, it is necessary to balance acquisition accuracy, processing speed, and system complexity. For some simple research needs, the system structure can be optimized by reducing the number of backbones and merging functions. For complex research objectives, the number of backbones should be increased to ensure independent acquisition and efficient processing of various types of information.

[0012] Step 4: Embed C2DAT modules at the head and tail of each backbone level to optimize information extraction and computational resource allocation; improve the backbone's information processing efficiency and ensure the relevance and accuracy of data processing. The C2DAT module embedded at the backbone head focuses on and strengthens the specific information extracted by the backbone. By extracting and filtering specific features of the input data, this module can accurately capture key information related to the target task while suppressing interference factors unrelated to the backbone's task. The introduction of this module effectively reduces the redundancy of input information in the initial stage, thereby optimizing data flow efficiency. Similarly, a C2DAT module is embedded at the tail of the backbone to further focus on data output directly related to the target information. At this stage, the main function of this module is to process data passing through the backbone's intermediate layers. The intermediate data generated after processing is filtered and simplified. By eliminating redundant or irrelevant calculation results, the tail module can reduce the occupation of computing resources by redundant information, thereby freeing up computing power to be invested in critical tasks. In addition, the tail module can also enhance the expressiveness and interpretability of target information, providing highly focused data output for subsequent tasks. This dual-module design optimizes the input and output data streams at the head and tail of the backbone, respectively. At the head, the C2DAT module can capture specific features in a targeted manner when data flows in. At the tail, the module further focuses and eliminates redundancy to enhance the expression of target information. Through this structure, the backbone can not only achieve more efficient information processing, but also improve the efficiency of computing power utilization, ensuring a more rational and scientific allocation of computing resources.

[0013] Step 5: Training employs a complex iterative training method, including: feature extraction stage, optimization stage, and iterative adjustment.

[0014] Step 6: Use a neural network model to identify bubbles / droplets, obtain bubble / droplet information, track bubble / droplet motion, and reconstruct bubble / droplet morphology in three-dimensional space; accurately separate bubbles / droplets from the background, extract geometric features and dynamic information, and combine multi-view data to achieve high-precision three-dimensional reconstruction of bubble / droplet morphology.

[0015] The method described above is a parallel targeted extraction method for multidimensional information based on the spatial characteristics of multiphase flow droplets / bubbles. This method utilizes optical images containing multiple types of information, such as... Figure 2 The multi-backbone neural network shown extracts various feature information from an image, and the process is as follows:

[0016] Step 1: Preprocess the image, set the resolution, and then input it into the neural network's Input layer;

[0017] Step 2: Feed the dataset after simple convolution into a multidimensional backbone network composed of multiple backbones to deeply mine the required information. Obtain various information about the target, including its location, size, shape, and motion state.

[0018] Step 3: Embed the C2DAT dynamic attention module at the head and tail of each backbone channel to focus on the target information, remove redundant calculations and free up computing power;

[0019] Step 4: The information obtained from the multidimensional backbone is spliced ​​and combined, and then used in subsequent convolutional pooling and NMS, Softmax and other functions to complete the target recognition, classification, tracking and prediction operations.

[0020] The method for parallel targeted extraction of multidimensional information based on the spatial characteristics of multiphase flow droplets / bubbles involves constructing a dataset in step one by collecting multidimensional complex bubble / droplet images with high deformation, large overlap, and large offset angle caused by different operating conditions, strong disturbances, and high turbulence. The dataset is divided into a training set and a validation set, with 70%~80% used for the training set and 30%~20% used for the validation set, to ensure the generalization ability and robustness of the model in complex environments.

[0021] The method for parallel targeted extraction of multidimensional information based on the spatial characteristics of multiphase flow droplets / bubbles includes a step two method where the acquired images are labeled to extract key information and the images are uniformly processed, including automatic orientation, size scaling, grayscale conversion, cropping, and object isolation data preprocessing steps to standardize the data format.

[0022] The method for parallel targeted extraction of multidimensional information based on the spatial features of multiphase flow droplets / bubbles, wherein the complex optical image obtained in step three has a variety of information including: semantic information, structural information, temporal information, contextual information, spatial information, feature interaction information, and scale invariance.

[0023] Semantic information: The multidimensional backbone model uses semantic information to identify bubbles / droplets in images and the differences between bubbles / droplets and fluids, helping the model understand the key elements in the image and their meanings;

[0024] Structural information: Structural information covers the external contour features and internal details of bubbles / droplets; by reading structural information, the multidimensional backbone model can accurately capture the morphological features of bubbles / droplets and efficiently represent and visualize them.

[0025] Temporal information: When analyzing video data, the multidimensional backbone model can track the changes of bubbles / droplets over time through temporal information, such as their speed, direction, merging or splitting behavior;

[0026] Contextual information: The multidimensional backbone model analyzes the contextual relationships of bubbles / droplets throughout the flow field by identifying their relative positions and interactions;

[0027] Spatial information: Spatial information is crucial for analyzing the distribution, density, and overall flow patterns of bubbles / droplets in fluids. Multidimensional backbone models can accurately extract spatial features from images, including the precise location of bubbles / droplets and the spatial relationships between them.

[0028] Feature interaction information: Feature interaction information helps multidimensional backbone models learn and understand the complex interactions between bubble / droplet features (such as size and shape), which helps to analyze the characteristics of gas-liquid flow at a higher level;

[0029] Scale invariance: The multidimensional backbone model identifies similar features at different scales through various convolution and pooling layers; this means that the model can effectively identify and classify bubbles / droplets regardless of their size.

[0030] Step 3 consists of a multi-level backbone composed of a main branch, a reversible auxiliary branch, ResNet, GELAN, and MobileNet. The main branch only acquires some surface information and obtains global feature information from the original image data through preliminary convolution, then directly enters the tail of the backbone for feature stitching. The reversible auxiliary branch provides backpropagation support to the main branch by generating reliable gradient information. Multiple backbones extract different information from various unique information. ResNet introduces a residual module to solve the gradient vanishing problem in deep networks and capture deeper features. GELAN is used to identify the size changes and shape diversity of bubbles / droplets, while segmenting the image to obtain complete spatial information. MobileNet is used to assist other branches in real-time computation, ensuring sufficient computing power for each level of the backbone.

[0031] The method for parallel targeted extraction of multidimensional information based on the spatial features of multiphase flow droplets / bubbles includes a C2DAT module embedded in step four to enhance computational performance and contribute to the overall lightweighting of the model. The C2DAT module is introduced into the head and tail of the backbone network to efficiently extract bubble / droplet features. The input to the head C2DAT module in step four is a feature map from the previous layer, which has already been processed by the preceding convolutional layers and contains preliminary image information. The C2DAT in step four includes standard convolution, depthwise convolution, pointwise convolution, and the merging formulas shown below:

[0032]

[0033] in Y It outputs the feature map. X inIt is the input feature map. W It represents the convolution kernel, and * represents the convolution calculation. z It is a bias;

[0034] In step four, C2DAT addresses the high memory and computational costs inherent in traditional methods due to the input layer receiving a wide range of features. It also effectively reduces the influence of irrelevant parts of the image. The C2DAT network structure is as follows: Figure 3 As shown, C2DAT dynamically selects attention keypoints and value pairs based on data requirements, allowing the model to focus on relevant regions and extract more informative features. C2DAT's deformable attention mechanism focuses only on a small portion of the image. In this mechanism, C2DAT dynamically selects sampling points instead of processing the entire image. This dynamic selection allows the model to concentrate on the regions most important to the current task.

[0035] The method for parallel targeted extraction of multidimensional information based on the spatial characteristics of multiphase flow droplets / bubbles, wherein step five of the training adopts a complex iterative training method, including: feature extraction stage, optimization stage, and iterative adjustment;

[0036] In the feature extraction stage, the model extracts features from the input data through forward propagation; these feature representations form the basis for future classification or prediction tasks.

[0037] The optimization phase uses a loss function to optimize the model's parameters; this step typically involves calculating the loss and then adjusting the network weights using the backpropagation algorithm to reduce prediction error.

[0038] In the iterative adjustment phase, after the first iteration, the feature extractor is adjusted and the network structure or parameters are modified based on the model's performance. Then, feature extraction and optimization are performed again. Multiple iterations are conducted to gradually improve the model's understanding and prediction capabilities of the data.

[0039] The method for parallel targeted extraction of multidimensional information based on the spatial features of multiphase flow droplets / bubbles, in step six, estimates the three-dimensional dimensions of feature bubbles / droplets using the planar projection of the bubbles / droplets; and obtains the major axis of the bubbles / droplets from the recognized image. a short axis c and tilt angle β Therefore, it is also necessary to obtain the major axis of the elliptical bubble / droplet perpendicular to the image plane. b To estimate the volume and cross-sectional area of ​​the bubble / droplet; to establish a spatial coordinate system using the equatorial radius and polar radius of the ellipsoid; in order to obtain the ellipsoid at arbitrary heights h The horizontal cross section at that location needs to be determined. β Inclination angle in a plane coordinate system θ As shown below:

[0040] (5)

[0041] ellipsoid at any height h The equation of the elliptical section at point is as follows:

[0042] (6)

[0043] In the formula, a 1 is the major axis of the elliptical cross section.

[0044] (7)

[0045] b 1 is the minor axis of the elliptical cross section.

[0046] (8)

[0047] The formula for the cross-sectional area of ​​an ellipse is as follows:

[0048] (9)

[0049] The formula for bubble / droplet volume is:

[0050] (10).

[0051] The method described above is a parallel targeted extraction method for multidimensional information based on the spatial characteristics of multiphase flow droplets / bubbles. This method improves the performance of neural network models and provides a reference for multidimensional information acquisition and overall model lightweighting.

[0052] The beneficial effects of this invention are:

[0053] Based on a multi-backbone neural network design, the system comprehensively extracts multi-dimensional information from complex images. The introduction of the C2DAT module not only significantly improves computational performance but also contributes to the overall lightweight design of the model. Through parallel design, the multi-backbone neural network fully utilizes the characteristics of main branches, auxiliary branches, and the lightweight network to comprehensively mine local and global features in images from different levels and perspectives. This structure ensures that the model captures global semantic information while also taking into account local details and dynamic changes.

[0054] In a multidimensional backbone network, information extracted from each branch is fused through feature concatenation and combination to achieve comprehensive acquisition of global information. This fusion strategy organically integrates features from different network branches, enabling the model to more accurately model the feature distribution and interrelationships in complex scenes. The embedding of the C2DAT module further enhances the efficiency and effectiveness of feature processing. Through a dynamic attention mechanism, C2DAT can focus on the most relevant target regions in an image while ignoring unnecessary background information and noise, thereby optimizing the allocation of computational resources. Its deformable structure makes the attention mechanism more flexible, adaptable to various scenes and feature distributions, and improves the model's ability to capture target features.

[0055] The introduction of the C2DAT module significantly contributes to the lightweight design of the model. By reducing the processing and computation of irrelevant regions, this module effectively reduces the overall computational load of the model, making the inference process in complex scenes more efficient. Furthermore, the multi-backbone neural network works in conjunction with the C2DAT module to achieve a balance between performance and efficiency, ensuring high-precision feature extraction while meeting the requirements of real-time computation and lightweight design. This design provides an efficient, flexible, and high-performance solution for complex image recognition tasks. Attached Figure Description

[0056] Figure 1 This is a diagram of the overall neural network architecture of the present invention, which includes information about each layer of the neural network;

[0057] Figure 2 The overall flowchart of image recognition in this invention reveals the entire process of image recognition and target detection;

[0058] Figure 3 This is a structural diagram of the C2DAT module of the present invention;

[0059] Figure 4 These are graphs showing various performance indicators after the training of this invention is completed. Detailed Implementation

[0060] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention based on the content disclosed in this specification. Furthermore, the present invention can be implemented or applied through various other specific embodiments, and the details in this specification can also be modified or adjusted based on different perspectives and uses without departing from the spirit of the present invention. It should be particularly noted that, without conflict, the following embodiments and their features can be flexibly combined to adapt to different needs and application scenarios.

[0061] A parallel target extraction method based on multidimensional information of multiphase flow droplets / bubbles, comprising multidimensional information-based target dynamic detection and prediction, and overall model lightweighting: the specific steps are as follows:

[0062] Step 1: Use bubble / droplet images under different working conditions as the dataset, with 70%~80% used for the training set and 30%~20% used for the validation set;

[0063] Step 2: Calibrate the bubbles / droplets in the image, and then perform automatic orientation, resizing, grayscale conversion, cropping, and object isolation operations on the image;

[0064] Step 3: When extracting bubble / droplet information in three-dimensional space, clearly define the types of bubbles / droplets to be extracted, including size, shape, distribution location, number density, and movement trajectory; design the number of backbones based on the complexity of the information and the data dimensions; different backbones are used to collect one or more different types of information respectively;

[0065] Step 4: In the design of the head and tail of each level of the backbone, C2DAT modules are embedded to optimize information extraction and computing resource allocation. The C2DAT module embedded at the head of the backbone focuses on and enhances the specific information extracted by the backbone. By extracting and filtering specific features of the input data, this module can accurately capture key information related to the target task while suppressing interference factors unrelated to the backbone task. At the tail of the backbone, a C2DAT module is also embedded to further focus on data output directly related to the target information. At this stage, the main function of this module is to filter and simplify the intermediate data generated after processing by the backbone's intermediate layer. This dual-module design optimizes the input and output data flows at the head and tail of the backbone, respectively. At the head, the C2DAT module can capture specific features in a targeted manner when data flows in; at the tail, the module further focuses and removes redundancy to enhance the expression of the target information.

[0066] In step four, embedding the C2DAT module enhances computational performance while contributing to the overall lightweighting of the model. The C2DAT module is introduced into the head and tail of the backbone network to efficiently extract bubble / droplet features. The feature maps input to the head C2DAT module in step four are feature maps from the previous layer, which have already been processed by the preceding convolutional layers and contain preliminary image information. The C2DAT in step four includes standard convolution, depthwise convolution, pointwise convolution, and the merging formula is shown below:

[0067] Where Y is the output feature map, X inis the input feature map, W is the convolution kernel, * is the convolution calculation, and z is the bias;

[0068] In step four, C2DAT addresses the high memory and computational costs inherent in traditional methods due to the input layer receiving a wide range of features. It also effectively reduces the influence of irrelevant parts of the image. C2DAT dynamically selects attention keypoints and value pairs based on data requirements, allowing the model to focus on relevant regions and extract more informative features. The deformable attention mechanism introduced by C2DAT focuses only on a small portion of the image. In this mechanism, C2DAT dynamically selects sampling points instead of processing the entire image. This dynamic selection allows the model to concentrate on the regions most important to the current task.

[0069] Step 5: Training employs a complex iterative training method, including: feature extraction stage, optimization stage, and iterative adjustment.

[0070] Step 6: Use a neural network model to identify bubbles / droplets, obtain bubble / droplet information, track bubble / droplet motion, and reconstruct bubble / droplet morphology in three-dimensional space; accurately separate bubbles / droplets from the background, extract geometric features and dynamic information, and combine multi-view data to achieve high-precision three-dimensional reconstruction of bubble / droplet morphology.

[0071] The method is based on optical images containing multiple types of information, and uses a multi-branch neural network to extract various feature information from the image. The process is as follows:

[0072] (1) Preprocess the image to set the resolution and then input it into the neural network Input layer;

[0073] (2) Feed the dataset after simple convolution into a multidimensional backbone network composed of multiple backbones to deeply mine the required information; obtain the target's various positions, sizes, shapes, and motion states;

[0074] (3) Embed the C2DAT dynamic attention module at the head and tail of each backbone channel to focus on the target information, remove redundant calculations and release computing power;

[0075] (4) The information obtained from the multidimensional backbone is spliced ​​and combined and then used for subsequent convolutional pooling and target recognition, classification, tracking and prediction operations are completed through NMS and Softmax functions.

[0076] In step one, a dataset is constructed by collecting multidimensional complex bubble / droplet images with high deformation, large overlap, and large offset angle caused by different operating conditions, strong disturbances, and high turbulence. The dataset is divided into a training set and a validation set, with 70% to 80% used for the training set and 30% to 20% used for the validation set, to ensure the model's generalization ability and robustness in complex environments.

[0077] In step two, the acquired images are labeled to extract key information, and the images are processed uniformly, including automatic orientation, size scaling, grayscale conversion, cropping, and object isolation data preprocessing steps to standardize the data format.

[0078] Step three involves extracting bubble / droplet information in three-dimensional space, including semantic information, structural information, temporal information, contextual information, spatial information, feature interaction information, and scale invariance.

[0079] Semantic information: The multidimensional backbone model uses semantic information to identify bubbles / droplets in images and the differences between bubbles / droplets and fluids, helping the model understand the key elements in the image and their meanings;

[0080] Structural information: Structural information covers the external contour features and internal details of bubbles / droplets; by reading structural information, the multidimensional backbone model can accurately capture the morphological features of bubbles / droplets and efficiently represent and visualize them.

[0081] Temporal information: When analyzing video data, the multidimensional backbone model can track the changes of bubbles / droplets over time through temporal information, including their movement speed, direction, merging or splitting behavior;

[0082] Contextual information: The multidimensional backbone model analyzes the contextual relationships of bubbles / droplets throughout the flow field by identifying their relative positions and interactions;

[0083] Spatial information: Spatial information is crucial for analyzing the distribution, density, and overall flow patterns of bubbles / droplets in fluids; multidimensional backbone models can accurately extract spatial features from images, including the precise location of bubbles / droplets in the image and the spatial relationships between bubbles / droplets;

[0084] Feature interaction information: Feature interaction information helps the multidimensional backbone model learn and understand the complex interactions between bubble / droplet features, which helps to analyze the characteristics of gas-liquid flow at a higher level;

[0085] Scale invariance: The multidimensional backbone model identifies similar features at different scales through various convolution and pooling layers; this means that the model can effectively identify and classify bubbles / droplets regardless of their size.

[0086] Step 3 consists of a multi-level backbone composed of a main branch, a reversible auxiliary branch, ResNet, GELAN, and MobileNet. The main branch only acquires some surface information and obtains global feature information from the original image data through preliminary convolution, then directly enters the tail of the backbone for feature stitching. The reversible auxiliary branch provides backpropagation support to the main branch by generating reliable gradient information. Multiple backbones extract different information from various unique information. ResNet introduces a residual module to solve the gradient vanishing problem in deep networks and capture deeper features. GELAN is used to identify the size changes and shape diversity of bubbles / droplets, while segmenting the image to obtain complete spatial information. MobileNet is used to assist other branches in real-time computation, ensuring sufficient computing power for each level of the backbone.

[0087] The training in step five employs a complex iterative training method, which includes: a feature extraction stage, an optimization stage, and iterative adjustment.

[0088] In the feature extraction stage, the model extracts features from the input data through forward propagation; these feature representations form the basis for future classification or prediction tasks.

[0089] The optimization phase uses a loss function to optimize the model's parameters; this step includes calculating the loss and then adjusting the network weights through the backpropagation algorithm to reduce prediction error.

[0090] In the iterative adjustment phase, after the first iteration, the feature extractor is adjusted and the network structure or parameters are modified based on the model's performance. Then, feature extraction and optimization are performed again. Multiple iterations are conducted to gradually improve the model's understanding and prediction capabilities of the data.

[0091] In step six, the three-dimensional dimensions of the feature bubbles / droplets are estimated using the planar projection of the bubbles / droplets; the major axis of the bubbles / droplets is obtained from the recognized image. a short axis c and tilt angle β Therefore, it is also necessary to obtain the major axis of the elliptical bubble / droplet perpendicular to the image plane. b To estimate the volume and cross-sectional area of ​​the bubble / droplet; to establish a spatial coordinate system using the equatorial radius and polar radius of the ellipsoid; in order to obtain the ellipsoid at arbitrary heights h The horizontal cross section at that location needs to be determined. β Inclination angle in a plane coordinate system θ As shown below:

[0092] (5)

[0093] ellipsoid at any height h The equation of the elliptical section at point is as follows:

[0094] (6)

[0095] In the formula, a 1 is the major axis of the elliptical cross section.

[0096] (7)

[0097] b 1 is the minor axis of the elliptical cross section.

[0098] (8)

[0099] The formula for the cross-sectional area of ​​an ellipse is as follows:

[0100] (9)

[0101] The formula for bubble / droplet volume is:

[0102] (10).

[0103] The method described above improves the metrics of neural network models and provides a reference for multi-dimensional information acquisition and overall model lightweighting. The specific embodiments of this invention are described in further detail below.

[0104] Data processing includes steps such as normalization, data augmentation, and resizing. The core of normalization is adjusting the pixel values ​​of an image to a specific range, such as [0, 1] or [-1, 1]. This eliminates dimensional differences in the data, making the data distribution more uniform, thereby improving the stability of neural network training and accelerating the model's convergence. Regarding data augmentation, operations such as rotation, cropping, scaling, flipping, or adjusting brightness can artificially increase the diversity of training data. This not only helps improve the model's generalization ability but also effectively reduces overfitting. For resizing, images are typically standardized to a fixed size to fit the network's input requirements. This can be achieved through scaling. If maintaining the original image proportions is necessary, zero-padding can be added to prevent information loss due to distortion. These processing steps complement each other, building a solid data foundation for deep learning tasks and improving model performance and training efficiency. Regarding normalization:

[0105] (11)

[0106] in μ The average pixel value. σ Standard deviation 。

[0107] The backbone network adopts a multi-dimensional backbone network architecture consisting of a main branch, a reversible auxiliary branch, ResNet, GELAN, and MobileNet. The backbone network is responsible for extracting multi-scale features. This mainly includes convolutional operations, batch normalization, activation functions, and feature fusion, as shown in Equations 12-15:

[0108] (12)

[0109] in X ( i, j ) is the input feature map, W ( m, n ) is the convolution kernel, b For bias, Y ( i, j ) is the output feature map.

[0110] (13)

[0111] in μ and σ 2 For the mean and variance. γ,β These are learnable parameters.

[0112] (14)

[0113] Feature fusion is as follows: (15)

[0114] (16)

[0115] The Neck layer consists of a Feature Pyramid Network (FPN) and a Path Aggregation Network (PANet). The FPN fuses high-level features (rich in semantic information) and low-level features (rich in detail information) from top to bottom, while PANet aggregates low-level features (containing spatial information) from low to high levels, further enhancing the detection capability of small targets. A multi-scale feature fusion method is used to merge features from different layers. Upsampling involves interpolating or deconvolving low-resolution features.

[0116] (17)

[0117] Given a k×k In a window, the max pooling operation selects the maximum value within the window for output:

[0118] (18)

[0119] Average pooling is: (19)

[0120] in x It is the input feature map.y This is the feature map after pooling. i,j Indicates the starting position of the pooling operation. k It is the size of the pooling window.

[0121] The average value is calculated over the entire input feature map, and a single value is output. Global average pooling is as follows:

[0122] (20)

[0123] in H and W These are the height and width of the feature map, respectively.

[0124] Besides pooling, convolution operations can also achieve downsampling. By setting the stride to be greater than 1, the convolution operation can be directly converted into downsampling.

[0125] (twenty one)

[0126] Among them stride S When the value is greater than 1, adjacent input signals will not be fully sampled, and the size of the output signal will be reduced.

[0127] The detector head predicts the target classification and bounding box parameters for each grid cell. The Softmax function predicts the target category:

[0128] (twenty two)

[0129] Bounding box regression outputs bounding box parameters, which are then converted into actual coordinates:

[0130] (twenty three)

[0131] in( c x ,c y ) represents the center coordinates of the grid cell. p w and p h The prior frame width and height

[0132] A composite loss function is used to comprehensively optimize the errors of classification, regression, and object detection. IoU optimization is used to improve bounding box matching, and Focal Loss is used to reduce the weight of easily classified samples.

[0133] (twenty four)

[0134] Focal Loss is:

[0135] (25)

[0136] in It is a "modulation term", when p t Larger (i.e., when the model's predictions are relatively accurate). This will become smaller, thus reducing the loss on easily classified samples. The existence of this term makes Focal Loss focus more on samples that are more difficult to classify. α It is a balancing factor, typically used when classes are imbalanced, to adjust the weights of positive and negative classes. For cases where positive and negative classes are imbalanced, a suitable... α To balance their effects. γ Parameters control the degree of attention given to difficult-to-classify samples. 。 when γ When = 0, Focal Loss degenerates into the standard cross-entropy loss; as = 0, Focal Loss degenerates into the standard cross-entropy loss; γ As the value increases, the model will suffer less loss on easily classified samples and pay more attention to difficult-to-classify samples.

[0137] The total loss formula is: (26)

[0138] Non-maximum suppression (NMS) is used to remove overlapping bounding boxes, retaining only the boxes with the highest confidence, thus ensuring clear and accurate detection results. Given two boxes... B i and B j Their intersection-union ratio (IoU) is calculated by the following formula:

[0139] (27)

[0140] (28)

[0141] in This represents the area of ​​the intersection region of the two boxes. This represents the area of ​​the union of the two bounding boxes. Quantization and knowledge distillation are used in the model to reduce computational complexity and memory requirements while maintaining good performance. The following are their basic principles and formulas.

[0142] The goal of quantization is to convert parameters, activation values, etc., in a high-precision model (e.g., 32-bit floating-point numbers) into a lower-precision representation (e.g., 8-bit integers), thereby reducing the model's storage and computational costs. Assume the original parameters... w It is a floating-point number, quantized parameter It is converted into an integer. The quantization process can be represented by the following formula:

[0143] (29)

[0144] round() represents the rounding operation. min ( w )yes w The smallest value in the range is typically used to shift data to a non-negative range. Δ is the quantization step size, calculated using the following formula:

[0145] (30)

[0146] in Q It refers to the number of bits used in quantization. Quantization is a process that achieves compression in terms of storage and computation by mapping continuous floating-point values ​​to discrete integer values.

[0147] Knowledge distillation is a method for transferring knowledge from a complex, large model (the teacher model) to a smaller model (the student model). The student model is trained by mimicking the output probability distribution of the teacher model, thus achieving similar performance but with fewer parameters and faster inference speed. Assume the probability distribution output by the teacher model is... p teacher The probability distribution output by the student model is as follows: p student The loss function for knowledge distillation typically consists of two parts: a standard classification loss and a distillation loss, which guides the student model's learning by minimizing the difference in the distributions of the student and teacher model outputs.

[0148] (31)

[0149] L CE ( p teacher , y ) is the cross-entropy loss of the student model. y It's a real label. L KL ( p student , p teacher The σ is the Kullback-Leibler divergence between the student model and the teacher model, which measures the difference between the two probability distributions. T It is a temperature parameter used to smooth the output probability distribution of the teacher model, enabling students to learn more detailed distribution information. α These are the weighting coefficients for the two loss components. α It controls the balance between the labeled true loss and the distillation loss of the student model.

[0150] (32)

[0151] Finally, the detection results are visualized. The bounding boxes are mapped back to the original image: (33)

[0152] Convert the fixed-size network output to the original image coordinates.

[0153] Figure 1 This is a diagram of the overall neural network architecture of the present invention, which includes information about each layer of the neural network. Figure 2 The overall flowchart of the image recognition process of this invention reveals the entire process of image recognition and target detection. Figure 3 This is a structural diagram of the C2DAT module of the present invention. Figure 4 These are graphs showing various performance indicators after the training of this invention is completed.

[0154] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0155] The above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Those skilled in the art can make various modifications and improvements to the present invention without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the present invention, and the patent protection scope of the present invention should be defined by the claims.

Claims

1. A parallel targeted extraction method for multidimensional information based on the spatial characteristics of multiphase flow droplets / bubbles, characterized in that, The method includes dynamic target detection and prediction based on multi-dimensional information, as well as overall model lightweighting: the specific steps are as follows: Step 1: Use bubble / droplet images under different working conditions as the dataset, with 70%~80% used for the training set and 30%~20% used for the validation set; Step 2: Calibrate the bubbles / droplets in the image, and then perform automatic orientation, resizing, grayscale conversion, cropping, and object isolation operations on the image; Step 3: When extracting bubble / droplet information in 3D space, clearly define the types of bubble / droplet information to be extracted, including size, shape, distribution location, number density, and movement trajectory; design the number of backbones based on the complexity of the information and the data dimensions; different backbones are used to collect one or more different types of information respectively; Step 4: In the design of the head and tail of each level of the backbone, C2DAT modules are embedded to optimize information extraction and computational resource allocation. The C2DAT module embedded at the head of the backbone focuses on and strengthens the information extracted by the backbone. By extracting and filtering the features of the input data, this module can accurately capture key information related to the target task while suppressing interference factors unrelated to the backbone task. At the tail of the backbone, a C2DAT module is also embedded to further focus on the data output directly related to the target information. At this stage, the module's function is to filter and simplify the intermediate data generated after processing by the backbone's intermediate layer. This dual-module design optimizes the input and output data flows at the head and tail of the backbone, respectively. At the head, the C2DAT module can capture features in a targeted manner as data flows in; at the tail, the module further focuses and removes redundancy to enhance the expression of the target information. In step four, embedding the C2DAT module enhances computational performance while contributing to the overall lightweighting of the model. The C2DAT module is introduced into the head and tail of the backbone network to efficiently extract bubble / droplet features. The feature maps input to the head C2DAT module in step four are feature maps from the previous layer, which have already been processed by the preceding convolutional layers and contain preliminary image information. The C2DAT in step four includes standard convolution, depthwise convolution, pointwise convolution, and the merging formula is shown below: ； Where Y is the output feature map, X in is the input feature map, W is the convolution kernel, * is the convolution calculation, and z is the bias; In step four, C2DAT solves the problem of high memory and computational cost caused by the input layer receiving a wide range of features in traditional methods. At the same time, it can effectively reduce the influence of irrelevant parts in the image. C2DAT can dynamically select attention key points and value pairs according to the needs of the data, so that the model focuses on relevant regions and extracts more informative features. The deformable attention mechanism introduced by C2DAT only focuses on a small part of the image; in the deformable attention mechanism, C2DAT dynamically selects sampling points instead of processing the entire image in a fixed way. This dynamic selection mechanism allows the model to focus on processing the areas most important to the current task. Step 5: Training employs a complex iterative training method, including: feature extraction stage, optimization stage, and iterative adjustment. Step 6: Use a neural network model to identify bubbles / droplets, obtain bubble / droplet information, track bubble / droplet motion, and reconstruct bubble / droplet morphology in three-dimensional space; accurately separate bubbles / droplets from the background, extract geometric features and dynamic information, and combine multi-view data to achieve high-precision three-dimensional reconstruction of bubble / droplet morphology.

2. The method for parallel targeted extraction of multidimensional information based on the spatial characteristics of multiphase flow droplets / bubbles according to claim 1, characterized in that, The method is based on optical images containing multiple types of information, and uses a multi-branch neural network to extract various feature information from the image. The process is as follows: (1) Preprocess the image to set the resolution and then input it into the neural network Input layer; (2) Feed the dataset after simple convolution into a multidimensional backbone network composed of multiple backbones to deeply mine the required information; obtain the target's various positions, sizes, shapes, and motion states; (3) Embed the C2DAT dynamic attention module at the head and tail of each backbone channel to focus on the target information, remove redundant calculations and release computing power; (4) The information obtained from the multidimensional backbone is spliced ​​and combined and then used for subsequent convolutional pooling and target recognition, classification, tracking and prediction operations are completed through NMS and Softmax functions.

3. The method for parallel targeted extraction of multidimensional information based on the spatial characteristics of multiphase flow droplets / bubbles according to claim 1, characterized in that, In step one, a dataset is constructed by collecting multidimensional complex bubble / droplet images with high deformation, large overlap, and large offset angle caused by different operating conditions, strong disturbances, and high turbulence. The dataset is divided into a training set and a validation set, with 70% to 80% used for the training set and 30% to 20% used for the validation set, to ensure the model's generalization ability and robustness in complex environments.

4. The method for parallel targeted extraction of multidimensional information based on the spatial characteristics of multiphase flow droplets / bubbles according to claim 1, characterized in that, In step two, the acquired images are labeled to extract key information, and the images are processed uniformly, including automatic orientation, size scaling, grayscale conversion, cropping, and object isolation data preprocessing steps to standardize the data format.

5. The method for parallel targeted extraction of multidimensional information based on the spatial characteristics of multiphase flow droplets / bubbles according to claim 1, characterized in that, Step three involves extracting bubble / droplet information in three-dimensional space, including semantic information, structural information, temporal information, contextual information, spatial information, feature interaction information, and scale invariance. Semantic information: The multidimensional backbone model uses semantic information to identify bubbles / droplets in images and the differences between bubbles / droplets and fluids, helping the model understand the key elements in the image and their meanings; Structural information: Structural information covers the external contour features and internal details of bubbles / droplets; by reading structural information, the multidimensional backbone model can accurately capture the morphological features of bubbles / droplets and efficiently represent and visualize them. Temporal information: When analyzing video data, the multidimensional backbone model can track the changes of bubbles / droplets over time through temporal information, including their movement speed, direction, merging or splitting behavior; Contextual information: The multidimensional backbone model analyzes the contextual relationships of bubbles / droplets throughout the flow field by identifying their relative positions and interactions; Spatial information: Spatial information is crucial for analyzing the distribution, density, and overall flow patterns of bubbles / droplets in fluids; multidimensional backbone models can accurately extract spatial features from images, including the precise location of bubbles / droplets in the image and the spatial relationships between bubbles / droplets; Feature interaction information: Feature interaction information helps the multidimensional backbone model learn and understand the complex interactions between bubble / droplet features, which helps to analyze the characteristics of gas-liquid flow at a higher level; Scale invariance: The multidimensional backbone model identifies similar features at different scales through various convolution and pooling layers; this means that the model can effectively identify and classify bubbles / droplets regardless of their size. Step 3 consists of a multi-level backbone composed of a main branch, a reversible auxiliary branch, ResNet, GELAN, and MobileNet. The main branch only acquires some surface information and obtains global feature information from the original image data through preliminary convolution, then directly enters the tail of the backbone for feature stitching. The reversible auxiliary branch provides backpropagation support to the main branch by generating reliable gradient information. Multiple backbones extract different information from various unique information. ResNet introduces a residual module to solve the gradient vanishing problem in deep networks and capture deeper features. GELAN is used to identify the size changes and shape diversity of bubbles / droplets, while segmenting the image to obtain complete spatial information. MobileNet is used to assist other branches in real-time computation, ensuring sufficient computing power for each level of the backbone.

6. The method for parallel targeted extraction of multidimensional information based on the spatial characteristics of multiphase flow droplets / bubbles according to claim 1, characterized in that, The training in step five employs a complex iterative training method, which includes: a feature extraction stage, an optimization stage, and iterative adjustment. In the feature extraction stage, the model extracts features from the input data through forward propagation; these feature representations form the basis for future classification or prediction tasks. The optimization phase uses a loss function to optimize the model's parameters; this step includes calculating the loss and then adjusting the network weights through the backpropagation algorithm to reduce prediction error. In the iterative adjustment phase, after the first iteration, the feature extractor is adjusted and the network structure or parameters are modified based on the model's performance. Then, feature extraction and optimization are performed again. Multiple iterations are conducted to gradually improve the model's understanding and prediction capabilities of the data.

7. The method for parallel targeted extraction of multidimensional information based on the spatial characteristics of multiphase flow droplets / bubbles according to claim 1, characterized in that, In step six, the three-dimensional dimensions of the feature bubbles / droplets are estimated using the planar projection of the bubbles / droplets; the major axis of the bubbles / droplets is obtained from the recognized image. a short axis c and tilt angle β ; Therefore, it is also necessary to obtain the major axis of the elliptical bubble / droplet perpendicular to the image plane. b To estimate the volume and cross-sectional area of ​​the bubble / droplet; to establish a spatial coordinate system using the equatorial radius and polar radius of the ellipsoid; in order to obtain the ellipsoid at arbitrary heights h The horizontal cross section at that location needs to be determined. β Inclination angle in a plane coordinate system θ As shown below: （5） ellipsoid at any height h The equation of the elliptical section at point is as follows: （6） In the formula, a 1 is the major axis of the elliptical cross section; （7） b 1 is the minor axis of the elliptical cross section; （8） The formula for the cross-sectional area of ​​an ellipse is as follows: （9） The formula for bubble / droplet volume is: （10）。 8. The method for parallel targeted extraction of multidimensional information based on the spatial characteristics of multiphase flow droplets / bubbles according to claim 1, characterized in that, The method improves the metrics of neural network models and provides a reference for multi-dimensional information acquisition and overall model lightweighting.

Images