A droplet velocity prediction method based on a CNN-RF model
By using a droplet velocity prediction method based on a CNN-RF model, the measurement difficulties of PIV technology in extreme environments are solved, and high-precision, high-speed droplet velocity measurement is achieved, which is suitable for flow field diagnosis in extreme environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENYANG AEROSPACE UNIVERSITY
- Filing Date
- 2026-02-25
- Publication Date
- 2026-05-29
AI Technical Summary
Existing particle image velocimetry (PIV) technology suffers from poor tracking of tracer particles in extreme environments, expensive and complex equipment, and long debugging time. Meanwhile, machine learning methods lack end-to-end solutions and do not fully utilize features, resulting in low measurement accuracy and efficiency in extreme spray environments.
A droplet velocity prediction method based on a CNN-RF model is adopted. By simultaneously acquiring nozzle atomization images and standard velocity field data, deep features are extracted using a convolutional neural network and combined with a random forest regression model to achieve high-precision prediction of droplet velocity, thus eliminating the dependence on tracer particles and complex optical systems.
It achieves high-precision, high-speed droplet velocity measurement in extreme environments, eliminating the dependence on tracer particles and complex optical systems, improving measurement efficiency, and is suitable for extreme environments such as those with limited optical access, high temperature and pressure, and dense sprays.
Smart Images
Figure CN122116070A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of flow field measurement technology, and in particular to a method for predicting droplet velocity based on a CNN-RF model. Background Technology
[0002] Particle image velocimetry (PIV) technology, developed since the 1980s, has become a standard method for measuring atomized flow fields. It involves seeding the flow field with tracer particles, illuminating it with pulsed laser sheet light, and recording particle displacements using a synchronized dual-frame camera. The overall velocity is then calculated using a cross-correlation algorithm. Over decades of development, various techniques have emerged, including microscopic PIV, high-speed PIV, and stereoscopic PIV, with continuously improving spatial and temporal resolution.
[0003] Despite the maturity of PIV technology, its inherent physical and technical limitations are fully exposed when applied to extreme environments such as engine fuel spray, gas turbine combustion chambers, and industrial nozzles under extreme pressure.
[0004] The shortcomings and deficiencies of traditional PIV systems: (1) In high-speed, high-acceleration (such as the strong shear region at the nozzle outlet) or two-phase flow, tracer particles cannot accurately follow the movement of fluid micro-clusters, resulting in systematic deviations in velocity measurement. For liquid-gas two-phase flow, it is extremely difficult to find tracer particles that can follow both the gas and liquid phases.
[0005] (2) A complete PIV system includes a high-energy dual-pulse laser, a precise synchronization controller, a high-sensitivity scientific-grade camera, a special lens and complex optical components, with a total price ranging from hundreds of thousands to millions of RMB.
[0006] (3) System debugging, calibration and maintenance require extremely high professional knowledge and experience. PIV experiments are conducted on extreme test benches (such as high temperature and high pressure), and the preparation and debugging time for each experiment is very long and inefficient.
[0007] Current shortcomings in the field of machine learning: (1) Lack of end-to-end solutions for extreme spraying environments: There is no mature and complete technical solution for directly and robustly predicting droplet-scale velocity fields from a single static atomization image obtained from extreme environments.
[0008] (2) The model structure is simple and the features are not fully utilized: Most studies use a single CNN (such as U-Net) for regression, which fails to effectively integrate the macroscopic morphological features of the spray (captured by the CNN) with the precise spatial location information of the droplets (key physical coordinates). This limits its prediction accuracy and physical consistency in complex and non-uniform spray fields.
[0009] (3) The design is not specifically designed for the working conditions where PIV fails: The training and validation data are mostly from the normal environment where PIV can work well. The robustness of the model has not been fully verified in the extremely dense and highly dynamic spray where PIV itself has failed.
[0010] In summary, there is a clear gap in the existing technology: on the one hand, mature PIV technology is ineffective or unusable in extreme environments due to physical limitations; on the other hand, emerging deep learning-based methods have not yet provided a complete, reliable, and alternative measurement scheme specifically designed to overcome the shortcomings of PIV. Summary of the Invention
[0011] The purpose of this invention is to propose a droplet velocity prediction method based on a CNN-RF model to solve the problems mentioned in the background technology, get rid of the dependence on tracer particles and complex optical systems, achieve high-precision and high-efficiency prediction of droplet velocity in atomized fields under extreme environments, and replace the traditional PIV technology to complete flow field diagnosis in extreme environments.
[0012] To achieve the above objectives, this invention provides a droplet velocity prediction method based on a CNN-RF model, comprising the following steps: Step S1: Data Acquisition and Preprocessing: Simultaneously acquire atomization images of the nozzle and corresponding standard velocity field data, perform preprocessing, and construct a structured labeled dataset; Step S2: Feature Extraction and Fusion: Input the structured label dataset into the pre-configured convolutional neural network (CNN) to extract deep features of the fogged image and convert them into one-dimensional feature vectors; obtain the spatial coordinates of each droplet in the fogged image, and fuse the one-dimensional feature vectors with the spatial coordinates to form a comprehensive feature vector for each droplet; Step S3: Model Training: Using the comprehensive feature vector as input, train the model using the Random Forest (RF) regression model to output the corresponding droplet velocity prediction value, thus obtaining the CNN-RF prediction model. Step S4: Velocity prediction and cloud map generation: Obtain the static fogged image to be predicted, repeat steps S1 and S2, input the generated comprehensive feature vector into the CNN-RF prediction model, and finally output the velocity component of each droplet. After smoothing the velocity components, generate a full-field velocity cloud map or velocity vector map.
[0013] Preferably, in step S1, synchronous acquisition is achieved through a composite measurement system, which includes: Imaging unit: a high-resolution, high-frame-rate CMOS or CCD camera, equipped with a macro lens or telephoto lens, with the camera lens optical axis perpendicular to or at a preset angle to the spray main axis; Standard velocity field measurement unit: This is a dual-pulse laser PIV system, LDV system, or simulation equipment used to acquire high-confidence standard velocity field data; Synchronization and Control Unit: Synchronizer, used to ensure spatiotemporal synchronization between the imaging unit and the standard velocity field measurement unit, and to complete the spatial coordinate system calibration.
[0014] Preferably, in step S1, the preprocessing of the fogged image includes denoising, contrast enhancement, and scale normalization. Denoising can be achieved using median filtering, Gaussian filtering, or CNN-based deep learning denoising algorithms. Contrast enhancement employs histogram equalization or the contrast-limited adaptive histogram equalization CLAHE algorithm. Scale normalization scales or crops an image to a uniform pixel size, and normalizes the pixel grayscale values to a uniform value. Interval.
[0015] Preferably, in step S1, the standard velocity field data preprocessing includes outlier removal, data interpolation, and spatiotemporal alignment; Outlier removal employs a combination of global thresholding and local median testing. The global thresholding method sets the effective range based on the mean and standard deviation of the standard velocity field data. Data interpolation employs nearest neighbor interpolation, bilinear interpolation, or interpolation algorithms based on fluid physics constraints; Spatiotemporal alignment is the process of transforming the physical coordinates of standard velocity field data to the same pixel coordinate system as the preprocessed image.
[0016] Preferably, step S2 includes the following steps: Step S21: Extract deep features of the fogged image and convert them into one-dimensional feature vectors: First, a convolutional neural network (CNN) with a multi-level structure is selected as the feature extractor. Features are extracted through multiple alternating convolutional and pooling layers. Then, the output three-dimensional feature vector is converted into a one-dimensional feature vector through a flattening operation. Step S22: Obtain the spatial coordinates of each droplet in the atomized image: Obtain the spatial coordinates of each droplet in the atomized image through connected component analysis, watershed segmentation algorithm or pre-trained lightweight segmentation network; Step S23: Merge the one-dimensional feature vector with the spatial coordinates to form a comprehensive feature vector for each droplet: The one-dimensional feature vector extracted by the convolutional neural network (CNN) is concatenated and fused with the spatial coordinates of each droplet to form a comprehensive feature vector for each droplet.
[0017] Preferably, in step S21, the calculation formula for the convolutional layer is as follows: ; in, Indicates the first The first layer Each feature map Represents a non-linear activation function. Represents the convolution kernel weight matrix. Indicates the first Number of layer feature maps Represents the bias vector; The pooling layer uses max pooling, and the formula is as follows: ; in, This is the feature map after pooling. This is the starting position of the pooling window. This represents the pixel offset within the window.
[0018] Preferably, in step S3, the random forest regression model is trained in parallel by integrating multiple decision trees, and the final velocity prediction value is the average of the prediction values of all decision trees, as shown in the following formula: ; ; in, , These are the velocity components of the predicted velocity value. For the number of decision trees, , For the first Decision trees are based on comprehensive feature vectors The predicted value.
[0019] Preferably, in step S4, the smoothing of the velocity components uses Gaussian filtering or median filtering; the velocity cloud map converts the velocity components into a pseudo-color map using color mapping technology, and the velocity vector map visually displays the flow field direction in the form of arrows.
[0020] Therefore, the droplet velocity prediction method based on the CNN-RF model described above has the following advantages: (1) Strong environmental adaptability: It gets rid of the dependence on tracer particles and complex optical systems. Only a conventional or high-speed camera is needed to acquire a single static fogging image to complete the measurement. It avoids the failure problems caused by poor tracking of tracer particles and laser attenuation and shielding in PIV technology. It can be stably applied to extreme environments such as optical access restriction, high temperature and high pressure, and dense spray. (2) High prediction accuracy: Multi-level visual features of fogged images are extracted by CNN, droplet spatial coordinate information is fused, and the powerful nonlinear regression capability of random forest is used to achieve accurate mapping of "shape + position" features to velocity values; (3) High efficiency and real-time capability: It provides an end-to-end automated solution. The entire process does not require manual intervention. After the model training is completed, the velocity prediction of a single image can be completed in seconds or even milliseconds. Compared with the complex debugging of PIV and the point-by-point scanning of LDV, it greatly improves the measurement efficiency and provides a fast tool for nozzle performance evaluation, combustion process control and fault diagnosis.
[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0022] Figure 1 This is a flowchart illustrating the steps of a droplet velocity prediction method based on a CNN-RF model mentioned in an embodiment of the present invention. Figure 2 This is the initial image of the atomization process in a droplet velocity prediction method based on a CNN-RF model mentioned in an embodiment of the present invention. Figure 3 The image shown is a preprocessed image of the atomized image in a droplet velocity prediction method based on a CNN-RF model mentioned in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the preprocessing of atomized images and corresponding standard velocity field data in a droplet velocity prediction method based on a CNN-RF model mentioned in an embodiment of the present invention. Figure 5 This is a schematic diagram of the feature extraction process in a droplet velocity prediction method based on a CNN-RF model mentioned in an embodiment of the present invention; Figure 6 This is a schematic diagram of the random forest pre-trained model in a droplet velocity prediction method based on a CNN-RF model mentioned in an embodiment of the present invention. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0024] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0025] Example like Figure 1-6 As shown, this embodiment provides a droplet velocity prediction method based on a CNN-RF model, including the following steps: Step S1: Data Acquisition and Preprocessing: Under controllable experimental conditions, the atomization images of the nozzle and the corresponding standard velocity field data are acquired simultaneously and preprocessed to construct a structured label dataset. The structured label dataset is then randomly divided into a training set, a validation set, and a test set in a ratio of 7:2:1.
[0026] Synchronous data acquisition is achieved through a composite measurement system, which includes: Imaging unit: a high-resolution, high-frame-rate CMOS or CCD camera equipped with a macro lens or telephoto lens, with the camera lens optical axis perpendicular to or at a preset angle to the spray main axis; Standard velocity field measurement unit: a dual-pulse laser PIV system, LDV system, or simulation equipment, used to acquire high-confidence standard velocity field data; Synchronization and control unit: a synchronizer used to ensure spatiotemporal synchronization between the imaging unit and the standard velocity field measurement unit, and to complete the spatial coordinate system calibration.
[0027] Preprocessing of fogged images includes denoising, contrast enhancement, and scale normalization: Denoising can be achieved using median filtering, Gaussian filtering, or CNN-based deep learning denoising algorithms. The formula for calculating Gaussian filtering is as follows: ; ; in, This represents the pixel coordinates of the hazy image after Gaussian filtering. grayscale value at that location Represents the x and y coordinates of pixels in a fogged image. This indicates the two-dimensional Gaussian kernel function at the offset. The value at that location, This represents the standard deviation of the Gaussian filter. This represents the half-width of the Gaussian filter window.
[0028] Contrast enhancement employs histogram equalization or the contrast-limited adaptive histogram equalization (CLAHE) algorithm. The histogram equalization formula is as follows: ; in, This represents the grayscale value of the hazy image after histogram equalization. The transformation function representing histogram equalization, This represents the grayscale level of the original fogged image. This represents the total number of gray levels in the fogged image. This indicates that the gray levels in the original fogged image are... The number of pixels, This represents the total number of pixels in the fogged image.
[0029] Scale normalization scales or crops an image to a uniform pixel size of 224×224 pixels or 512×512 pixels, and normalizes the pixel grayscale values to a uniform size. Interval.
[0030] Standard velocity field data preprocessing includes outlier removal, data interpolation, and spatiotemporal alignment. Outlier removal employs a combination of global thresholding and local median testing. The global thresholding method sets an effective range based on the mean and standard deviation of the standard velocity field data. Data interpolation uses nearest neighbor interpolation, bilinear interpolation, or interpolation algorithms based on fluid physics constraints. Spatiotemporal alignment transforms the physical coordinates of the standard velocity field data to the same pixel coordinate system as the preprocessed image.
[0031] The formula for removing outliers using the global threshold method is as follows: Wherein, the effective range of the standard velocity field data is: ; ; ; in, This represents the mean of the velocity field data. This represents the threshold adjustment coefficient. The standard deviation of the velocity field data is represented by... This represents the total number of samples in the velocity field data. Indicates the first The velocity values of each velocity field sample.
[0032] The spatiotemporal alignment formula is as follows: ; ; ; in, Represents the temporary pixel coordinates after matrix transformation. The scale factor representing the homography matrix. Represents the elements of the homography matrix. Physical coordinates representing standard velocity field data, This represents the pixel coordinates that match the fogged image after conversion.
[0033] Step S2: Feature Extraction and Fusion: Input the structured label dataset into the pre-configured VGG16-bn convolutional neural network (CNN) to extract deep features from the fogged image and convert them into one-dimensional feature vectors; obtain the spatial coordinates of each droplet in the fogged image, and fuse the one-dimensional feature vectors with the spatial coordinates to form a comprehensive feature vector for each droplet. This includes the following steps: Step S21: Extract deep features from the fogged image and convert them into a one-dimensional feature vector: Using a 224×224×3 image as input, a VGG16-bn convolutional neural network (CNN) with a multi-level structure is selected as the feature extractor. Features are extracted through multiple alternating convolutional layers and 2×2 max pooling layers. The conv5 layer is selected as the optimal feature extraction layer. Then, a flattening operation is used to convert the output three-dimensional feature vector into a one-dimensional feature vector. The non-linear activation function of the convolutional layer is ReLU. The calculation formula for the convolutional layer is as follows: ; in, Indicates the first The first layer Each feature map Represents a non-linear activation function. Indicates the first The total number of feature maps in each convolutional layer Represents the convolution kernel weight matrix. Indicates the first Number of layer feature maps This represents the bias vector; the pooling layer uses max pooling, as shown in the following formula: ; in, This is the feature map after pooling. The starting pixel coordinates of the pooling window. This is the pixel offset within the window.
[0034] Step S22: Obtain the spatial coordinates of each droplet in the atomized image: Obtain the spatial coordinates of each droplet in the atomized image through connected component analysis, watershed segmentation algorithm or pre-trained lightweight segmentation network.
[0035] Step S23: Merge the one-dimensional feature vector with the spatial coordinates to form a comprehensive feature vector for each droplet: The one-dimensional feature vector extracted by the convolutional neural network (CNN) is concatenated and fused with the spatial coordinates of each droplet to form a comprehensive feature vector for each droplet.
[0036] Step S3: Model Training: Using the comprehensive feature vector as input, train the Random Forest (RF) regression model to output the corresponding droplet velocity prediction value, thus obtaining the CNN-RF prediction model. The Random Forest regression model is trained in parallel by integrating multiple decision trees, and the final velocity prediction value is the average of the prediction values of all decision trees, as shown in the following formula: ; ; in, Based on comprehensive feature vectors Predicted droplets velocity components in the direction and The velocity component in the direction, This represents the comprehensive feature vector of the droplet. For the number of decision trees, In a random forest, the i-th The sequence number of each decision tree. , For the first Decision trees are based on comprehensive feature vectors Predicted Directional velocity value and Directional velocity value.
[0037] Step S4: Velocity prediction and cloud map generation: Obtain the static fogged image to be predicted, repeat steps S1 and S2, input the generated comprehensive feature vector into the CNN-RF prediction model, and finally output the velocity component of each droplet. After smoothing the velocity components, generate a full-field velocity cloud map or velocity vector map.
[0038] The smoothing of velocity components uses Gaussian filtering or median filtering; the velocity contour map converts the velocity components into a pseudo-color map using color mapping technology, and the velocity vector map visually displays the flow field direction in the form of arrows.
[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A method for predicting droplet velocity based on a CNN-RF model, characterized in that, Includes the following steps: Step S1: Data Acquisition and Preprocessing: Simultaneously acquire atomization images of the nozzle and corresponding standard velocity field data, perform preprocessing, and construct a structured labeled dataset; Step S2: Feature Extraction and Fusion: Input the structured label dataset into the pre-configured convolutional neural network (CNN) to extract deep features of the fogged image and convert them into one-dimensional feature vectors; obtain the spatial coordinates of each droplet in the fogged image, and fuse the one-dimensional feature vectors with the spatial coordinates to form a comprehensive feature vector for each droplet; Step S3: Model Training: Using the comprehensive feature vector as input, train the model using the Random Forest (RF) regression model to output the corresponding droplet velocity prediction value, thus obtaining the CNN-RF prediction model. Step S4: Velocity prediction and cloud map generation: Obtain the static fogged image to be predicted, repeat steps S1 and S2, input the generated comprehensive feature vector into the CNN-RF prediction model, and finally output the velocity component of each droplet. After smoothing the velocity components, generate a full-field velocity cloud map or velocity vector map.
2. The droplet velocity prediction method based on a CNN-RF model according to claim 1, characterized in that: In step S1, synchronous data acquisition is achieved through a composite measurement system, which includes: Imaging unit: a high-resolution, high-frame-rate CMOS or CCD camera, equipped with a macro lens or telephoto lens, with the camera lens optical axis perpendicular to or at a preset angle to the spray main axis; Standard velocity field measurement unit: This is a dual-pulse laser PIV system, LDV system, or simulation equipment used to acquire high-confidence standard velocity field data; Synchronization and Control Unit: Synchronizer, used to ensure spatiotemporal synchronization between the imaging unit and the standard velocity field measurement unit, and to complete the spatial coordinate system calibration.
3. The droplet velocity prediction method based on a CNN-RF model according to claim 1, characterized in that: In step S1, the preprocessing of the fogged image includes denoising, contrast enhancement, and scale normalization. Denoising can be achieved using median filtering, Gaussian filtering, or CNN-based deep learning denoising algorithms. Contrast enhancement employs histogram equalization or the contrast-limited adaptive histogram equalization CLAHE algorithm. Scale normalization scales or crops an image to a uniform pixel size, and normalizes the pixel grayscale values to a uniform value. Interval.
4. The droplet velocity prediction method based on a CNN-RF model according to claim 1, characterized in that: In step S1, the standard velocity field data preprocessing includes outlier removal, data interpolation, and spatiotemporal alignment; Outlier removal employs a combination of global thresholding and local median testing. The global thresholding method sets the effective range based on the mean and standard deviation of the standard velocity field data. Data interpolation employs nearest neighbor interpolation, bilinear interpolation, or interpolation algorithms based on fluid physics constraints; Spatiotemporal alignment is the process of transforming the physical coordinates of standard velocity field data to the same pixel coordinate system as the preprocessed image.
5. The droplet velocity prediction method based on a CNN-RF model according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: Extract deep features of the fogged image and convert them into one-dimensional feature vectors: First, a convolutional neural network (CNN) with a multi-level structure is selected as the feature extractor. Features are extracted through multiple alternating convolutional and pooling layers. Then, the output three-dimensional feature vector is converted into a one-dimensional feature vector through a flattening operation. Step S22: Obtain the spatial coordinates of each droplet in the atomized image: Obtain the spatial coordinates of each droplet in the atomized image through connected component analysis, watershed segmentation algorithm or pre-trained lightweight segmentation network; Step S23: Merge the one-dimensional feature vector with the spatial coordinates to form a comprehensive feature vector for each droplet: The one-dimensional feature vector extracted by the convolutional neural network (CNN) is concatenated and fused with the spatial coordinates of each droplet to form a comprehensive feature vector for each droplet.
6. The droplet velocity prediction method based on a CNN-RF model according to claim 5, characterized in that: In step S21, the calculation formula for the convolutional layer is as follows: ； in, Indicates the first The first layer Each feature map Represents a non-linear activation function. Represents the convolution kernel weight matrix. Indicates the first Number of layer feature maps Represents the bias vector; The pooling layer uses max pooling, and the formula is as follows: ； in, This is the feature map after pooling. This is the starting position of the pooling window. This represents the pixel offset within the window.
7. The droplet velocity prediction method based on a CNN-RF model according to claim 1, characterized in that: In step S3, the random forest regression model is trained in parallel by integrating multiple decision trees. The final velocity prediction value is the average of the prediction values of all decision trees, as shown in the following formula: ； ； in, , These are the velocity components of the predicted velocity value. For the number of decision trees, , For the first Decision trees are based on comprehensive feature vectors The predicted value.
8. The droplet velocity prediction method based on a CNN-RF model according to claim 1, characterized in that: Step S4: The smoothing of the velocity components uses Gaussian filtering or median filtering; the velocity cloud map converts the velocity components into a pseudo-color map using color mapping technology, and the velocity vector map visually displays the flow field direction in the form of arrows.