High-time-resolution flame three-dimensional reconstruction system and method

Abstract

The invention discloses a high-time-resolution flame three-dimensional reconstruction system and method, and relates to the technical field of combustion measurement, and the system mainly comprises a high-speed camera, a plurality of low-speed cameras, a lens, an optical filter, a trigger signal controller and a data acquisition and control system. The high-speed cameras and the low-speed cameras are evenly arranged in the circumferential direction with the flame as the center. The trigger signal controller controls the high-speed cameras to continuously shoot flames, and the low-speed cameras at all the visual angles sequentially shoot the flames in the circumferential direction at a certain time interval; a multi-view flame image prediction model associated with flame images of the high-speed camera and the low-speed camera is constructed through machine learning, and high-time-resolution flame images of view angles of the low-speed cameras are obtained through calculation; and realizing three-dimensional reconstruction of the high-time-resolution flame structure through a three-dimensional reconstruction algorithm. The high-time-resolution flame three-dimensional reconstruction is realized through one high-speed camera and a plurality of low-speed cameras, the hardware configuration requirement of the high-time-resolution flame three-dimensional reconstruction can be obviously reduced, and the flame three-dimensional structure measurement cost is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of combustion measurement, and in particular relates to a high time-resolution flame three-dimensional reconstruction system and method. Background Art

[0002] Combustion phenomena are widely present in energy and power equipment such as boilers, kilns, gas turbines, and aircraft engines, and are used in the fields of electricity, aviation, and ships. Combustion processes, such as ignition, flameout, flame pulsation, and flame propagation, are typical non-steady-state processes, and their transient flame structures have significant three-dimensional spatial characteristics. Traditional two-dimensional flame images cannot accurately depict the three-dimensional information of the flame space, reveal the complex non-steady-state combustion development process, and thus provide support for the optimization design of the combustion process and burner. Therefore, the development of a high-time-resolution three-dimensional flame structure reconstruction system and method is of great significance for revealing the flame space development process and optimizing the non-steady-state combustion process.

[0003] Flame 3D measurement technologies primarily include chemiluminescence tomography (CTC), volume laser-induced fluorescence (VLIF), volume laser-induced infrared (VLII), and passive infrared radiation. These methods capture flame projections from different perspectives using multiple image sensors surrounding the flame, and then reconstruct the flame's 3D structure using 3D reconstruction algorithms. These high-time-resolution 3D flame reconstruction methods typically rely on multiple high-speed cameras, resulting in complex experimental systems and high measurement costs. Furthermore, methods such as volume laser-induced fluorescence (VLIF) and volume laser-induced infrared (VLII) require equipment such as lasers, placing high demands on both the site and the environment, thus limiting their application in actual combustion equipment.

[0004] The present invention proposes a high-time-resolution flame 3D reconstruction system and method, which uses a high-speed camera combined with multiple low-speed cameras to achieve high-time-resolution flame 3D reconstruction, which can significantly reduce the hardware requirements for high-time-resolution flame 3D reconstruction and reduce the cost of flame 3D structure measurement. Summary of the Invention

[0005] The present invention discloses a high-time-resolution three-dimensional flame reconstruction system and method. The system primarily comprises a high-speed camera, multiple low-speed cameras, lenses, filters, a trigger signal controller, and a data acquisition and control system. The high-speed and low-speed cameras are evenly arranged circumferentially around the flame. The high-speed camera continuously captures the flame at a high frame rate, and the trigger signal controller generates trigger signals to control the low-speed cameras at each viewing angle to sequentially capture images circumferentially at regular time intervals. A multi-view flame image prediction model is constructed through machine learning to correlate the flame images from the high-speed and low-speed cameras. This model is used to obtain high-time-resolution flame images from each low-speed camera's viewing angle. High-time-resolution three-dimensional reconstruction of the flame structure is achieved using a three-dimensional reconstruction algorithm.

[0006] The technical solution adopted by the present invention comprises the following steps:

[0007] Step 1: Use a high-speed camera and multiple low-speed cameras to collect two-dimensional flame images from different perspectives;

[0008] Step 2: preprocess the flame image by background noise reduction;

[0009] Step 3: Use machine learning algorithms to build a multi-view flame image prediction model that correlates high-speed camera images and low-speed camera flame images;

[0010] Step 4: Using a multi-view flame image prediction model, calculate the high time resolution flame image of each low-speed camera perspective;

[0011] Step 5: Based on the high time-resolution flame images at each viewing angle, reconstruct the high time-resolution flame three-dimensional structure using a three-dimensional reconstruction algorithm.

[0012] In step 1, a high-speed camera and multiple low-speed cameras capture two-dimensional flame images from different perspectives, centered around the combustion chamber. A trigger signal controller generates trigger signals to control the high-speed camera to continuously capture the flame at a high frame rate, acquiring high-temporal-resolution flame images. The low-speed cameras at each perspective sequentially capture low-temporal-resolution flame images along the circumference at regular intervals, ensuring that at every moment the high-speed camera is capturing, a low-speed camera from a different perspective is also capturing.

[0013] In step 2, the flame images collected by cameras at various angles are subjected to denoising processing to eliminate noise in the image, thereby removing redundant information; during the denoising process, the corresponding background noise and image noise are removed according to the viewing angle distribution, brightness distribution and flame characteristics of the flame image to smooth the image.

[0014] In step three, a multi-view flame image prediction model is constructed using a machine learning algorithm based on multiple sets of simultaneous high-speed and low-speed camera images. This model then calculates a high-temporal-resolution image from each low-speed camera's perspective. The calculated image from that specific perspective is then compared with a real-world image captured at that specific perspective, and this data is used as the validation set input into the model. Finally, the model undergoes a cycle of modification, training, and verification, ensuring that the image output by the multi-view flame image prediction model gradually approximates the real-world flame image recorded at that moment. The frame rate of the high-temporal-resolution image derived by the model for each low-speed camera perspective corresponds to the frame rate of the high-speed camera.

[0015] In step 4, the image captured by a single high-speed camera is input. Using the multi-view flame image prediction model, high-temporal-resolution flame images from each low-speed camera's perspective are obtained. Because the flame images used for training have been preprocessed, the derived high-temporal-resolution flame images can be used for 3D flame reconstruction without requiring denoising or other operations.

[0016] In step 5, the image is used as a pixel matrix and then projected back to the flame 3D structure matrix by transferring the model matrix. The 3D reconstruction algorithm is then used to iteratively calculate and eventually gradually approximate the true flame 3D structure matrix. Specifically, the method for reconstructing the 3D flame structure mentioned in step 5 is as follows:

[0017] First, the low-speed camera perspectives derived in step 4 and the images of the high-speed camera at the same time, that is, the two-dimensional projections of the flame three-dimensional structure at each perspective at the same time, are used as pixel matrices for iterative verification of the three-dimensional structure reconstruction.

[0018] Next, the light propagation method is set and a transfer model matrix is established. The transfer model matrix simulates the propagation, absorption, and scattering of light in the flame. It serves as a weight coefficient matrix to calculate the pixel matrix of the flame's three-dimensional structure projected to each viewing angle.

[0019] Finally, an algebraic iteration method is used as a 3D reconstruction algorithm to establish matrix formulas for the camera image and the radiation transfer model. This method is a commonly used numerical calculation method for flame 3D reconstruction. Its principle is to approximate the exact solution through continuous iteration. In this process, the camera image is matrixed and back-projected using the transfer model to gradually approximate the true 3D structure of the flame. It should be noted that other algorithms besides the algebraic iteration method are also feasible for flame 3D reconstruction in this invention.

[0020] Through the above steps, the present invention realizes high time resolution three-dimensional flame structure reconstruction by using a high-speed camera in combination with multiple low-speed cameras.

[0021] The present invention has the following advantages:

[0022] (1) Reduce the hardware requirements for high time resolution flame 3D reconstruction

[0023] Compared with the traditional method of reconstructing the high-time-resolution three-dimensional structure of flames using multiple high-speed cameras, the present invention uses a high-speed camera combined with multiple low-speed cameras to achieve high-time-resolution three-dimensional reconstruction of flames, reducing the hardware configuration requirements for high-time-resolution three-dimensional reconstruction of flames and saving the cost of flame three-dimensional structure measurement.

[0024] (2) A single camera can achieve high time resolution 3D dynamic reconstruction of flames

[0025] The present invention obtains a multi-view flame image prediction model through machine learning training. Through the model, a single high-speed camera is used to realize three-dimensional reconstruction of flames. High-time-resolution three-dimensional dynamic reconstruction of flames with a single camera can be achieved, providing a feasible approach for three-dimensional reconstruction of flames.

[0026] (3) Wide range of application environments

[0027] This method does not rely on a large number of high-speed camera equipment and laser equipment. It uses a combination of a high-speed camera and multiple low-speed cameras to perform high-time-resolution flame three-dimensional structure measurement, which greatly reduces the site and environmental requirements for this type of flame three-dimensional reconstruction measurement, thereby broadening the application environment range of the flame three-dimensional reconstruction method. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of the flame three-dimensional reconstruction system of the present invention;

[0029] Figure 2 This is a timing diagram of the camera trigger signal of the present invention;

[0030] Figure 3 This is a schematic diagram of the flame shooting timing experimental platform of the present invention;

[0031] Figure 4 Schematic diagram of the circumferential position of the flame shooting sequence of the present invention

[0032] Figure 5 This is a schematic diagram of multi-view flame image prediction according to the present invention;

[0033] Figure 6 The flame image (left) and the predicted flame image (right) taken by a low-speed camera at a certain moment of the present invention;

[0034] Figure 7 This is a flow chart of three-dimensional flame reconstruction of the present invention;

[0035] Figure 8 The three-dimensional flame structure reconstructed by the present invention;

[0036] Figure 9 This is a flowchart of flame three-dimensional reconstruction based on the multi-view flame image prediction model of the present invention. DETAILED DESCRIPTION

[0037] The present invention will be further described below in conjunction with the accompanying drawings and specific examples. It should be understood that these examples are only intended to illustrate the present invention and are not intended to limit the scope of the present invention. After reading this invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the appended claims.

[0038] A high time-resolution flame three-dimensional reconstruction system and method, comprising the following steps:

[0039] Step 1: Use a high-speed camera and multiple low-speed cameras to capture two-dimensional flame images from different perspectives:

[0040] First, with the flame as the center, a high-speed camera and multiple low-speed cameras are used to collect flame images. The flame 3D reconstruction system is shown in the figure below. Figure 1 As shown. The high-speed camera is used to capture high-temporal-resolution images of the flame, ensuring that the rapidly changing details of the flame are captured. The low-speed camera is used to sequentially acquire flame images from various perspectives to provide multi-perspective information support. The trigger signal generator generates a trigger signal to control the high-speed camera to continuously shoot the flame at a high frame rate. The low-speed cameras at each perspective shoot sequentially along the circumference at a low frame rate at a certain interval. The specific time sequence of the high-speed and low-speed cameras is controlled by the trigger signal controller. The signal diagram of the trigger signal generator is shown in FIG. Figure 2 The shooting sequence of the flame low-speed camera and high-speed camera is shown as follows. Figure 3 As shown in the figure, a gray rectangle indicates a capture at that moment, and a white rectangle indicates no capture at that moment. The high-speed camera captures continuously, with the time it takes for the high-speed camera to capture one frame as the unit time. The low-speed cameras at each viewing angle capture images at intervals of unit time, sequentially along the circumferential viewing angle. A trigger signal is issued at each unit time interval, prompting the low-speed camera at the next viewing angle to capture the image at that moment. In this way, each high-speed camera can obtain the image captured by the low-speed camera at a specific viewing angle at the corresponding capture moment.

[0041] Step 2: Preprocess the flame image by background noise reduction:

[0042] The captured flame images were preprocessed with noise reduction and other techniques to facilitate subsequent machine learning and feature extraction. First, background image subtraction was performed on the flame images captured by the high-speed and low-speed cameras to remove background noise. Image filtering techniques were then used to reduce noise on the flame images, smoothing the images and reducing noise in the local flame structure.

[0043] Step 3: Use machine learning algorithms to build a multi-view flame image prediction model that correlates high-speed camera images and low-speed camera flame images:

[0044] A multi-view flame image prediction model is trained by combining multiple sets of images collected by high-speed cameras and low-speed cameras at the same time using a machine learning method. The images of multiple sets of high-speed cameras and low-speed cameras at the same time are used as training sets and validation sets. For example, the present invention constructs a multi-view flame image prediction model that combines an autoencoder neural network and a spatial mapping neural network (CNN-STN). The autoencoder neural network is used to process low-temporal-resolution low-speed camera data and map the low-temporal-resolution image to a high-dimensional feature space; the spatial mapping neural network is used to improve the spatial reconstruction capability of the image and enhance the detail recovery of the image by mapping the spatial information. The corresponding machine learning process is as follows: Figure 4 As shown, it should be noted that the machine learning method in the present invention is merely an example, and different machine learning methods can be used to construct a multi-view flame image prediction model. First, an autoencoder neural network and a spatial mapping neural network are used to preliminarily construct a multi-view flame image prediction model. This model is then used to calculate a high-time-resolution image from each low-speed camera's perspective. The calculated image from that specific perspective at that moment is then compared and verified with the actual image taken from that specific perspective, and this is used as the data input into the model for validation set verification. Finally, the model is subjected to a cycle of modification, training, and verification, so that the image output by the multi-view flame image prediction model gradually approaches the actual flame image recorded at that moment. The frame rate of the high-time-resolution image derived by the model for each low-speed camera from each perspective corresponds to the frame rate of the high-speed camera.

[0045] Step 4: Use the multi-view flame image prediction model to calculate the high-time-resolution flame image from each low-speed camera perspective:

[0046] The multi-view flame image prediction model is combined with the low-time resolution images of the low-speed cameras at each view to obtain the high-time resolution images of the low-speed cameras at each view. The real image taken by the low-speed camera at a certain moment (left) is compared with the image built by the multi-view flame image prediction model (right). Figure 5 The multi-view flame image prediction model can be used to obtain high-time-resolution images from low-speed cameras at each viewpoint, indirectly increasing the shooting frequency of the low-speed camera to the same frame rate as the high-speed camera.

[0047] Step 5: Based on the high-time-resolution flame images from each viewing angle, reconstruct the high-time-resolution flame 3D structure using a 3D reconstruction algorithm:

[0048] Based on the high time resolution flame images from each perspective, the flame 3D structure is reconstructed. The 3D flame reconstruction process is as follows: Figure 6 The main principle of this step is to use the optical intensity data of multiple angles to form the flame projection data matrix P m×1 , and the radiation transfer matrix W is calculated by the Radiative Transfer Equation (RTE) and the linear imaging model.m×n Finally, the principle of formula (1) is used to perform back projection through the three-dimensional reconstruction algorithm to restore the three-dimensional structure of the flame, where m and n refer to the total number of pixels in the projected image and the total number of pixels in the flame radiation matrix, respectively.

[0049] W m×n ×F n×1 =P m×1 (1)

[0050] First, in data processing, two-dimensional projection images of the flame are captured from different angles by multiple cameras and converted into a series of optical intensity matrices. These matrices represent the projection intensity distribution of the flame in the corresponding view at different angles, but do not directly reveal the flame distribution in three-dimensional space. Further processing is required to restore its three-dimensional structure. To achieve this goal, a projection reconstruction method based on the CTC (Computed tomography of chemiluminescence) principle is adopted, as shown in formula (2). The three-dimensional space of the flame is discretized into voxels to use a linear imaging model. The flame intensity of each voxel will be represented by its weight coefficient contribution to the projection intensity matrix, thereby establishing a relationship between the two-dimensional projection and the three-dimensional flame.

[0051]

[0052] After establishing the linear imaging model, the optical intensity matrix is mathematically modeled using a projection weight coefficient matrix. In this process, the intensity matrix at each projection angle is the weighted sum of the flame intensities of each voxel in the three-dimensional flame space. The contribution of each voxel to the projected image is represented by a weight coefficient matrix, which maps the three-dimensional flame to the two-dimensional projection image. By properly modeling the weight coefficient matrix, the projection matrix is represented as a linear combination of the three-dimensional flame intensity matrix and the projection weight coefficient matrix, thus establishing a mathematical model of the flame intensity distribution.

[0053] Using this mathematical model, the projection data is gradually solved with the help of the algebraic iterative reconstruction algorithm - SART (Simultaneous Algebraic Reconstruction Technique), namely formula (3). During the iteration process, the SART algorithm gradually approaches the real flame three-dimensional structure by continuously updating the three-dimensional model. By repeatedly adjusting the model parameters, the SART algorithm can extract accurate three-dimensional structural information from the projection intensity matrix of multiple angles. In each iteration, the algorithm adjusts according to the error of each projection image, and finally outputs the flame three-dimensional structure intensity distribution matrix after multiple iterations. The flame three-dimensional structure intensity distribution matrix result is as follows Figure 7 As shown, the three-dimensional matrix can accurately reflect the three-dimensional structure distribution of the flame in space.

[0054]

[0055] After the above steps, using the multi-view flame image prediction model, it is finally possible to capture flame images from all angles and reconstruct the flame 3D structure with high time resolution using only one high-speed camera. The flame 3D reconstruction process based on the multi-view flame image prediction model is as follows: Figure 8 shown.

Claims

1. A high time-resolution flame 3D reconstruction system, characterized by: The system consists of a high-speed camera, multiple low-speed cameras, lenses, filters, trigger signal controllers, and data acquisition and control systems. The high-speed cameras and low-speed cameras are evenly arranged along the circumference with the flame as the center.

2. The high time-resolution flame 3D reconstruction system according to claim 1, characterized in that: Cameras include CCD cameras, CMOS cameras, infrared thermal imaging cameras, and ultraviolet imaging cameras, which can be used for flame imaging and analysis in different bands when paired with filters and lenses of different bands.

3. The high time-resolution flame 3D reconstruction system according to claim 1, characterized in that: Optical filters include bandpass filters, longpass filters, shortpass filters, neutral density filters, and multi-channel spectral filters. Flame self-luminous groups used to characterize flame structure include CH groups, OH groups, C2 groups, CO groups, H2 groups, NH groups, CN groups, NO groups, and C3 groups.

4. The high time-resolution flame 3D reconstruction system according to claim 1, characterized in that: The trigger signal controller in the system sends a control signal to control the high-speed camera and the low-speed camera to shoot. The shooting frame rate f1 of the low-speed camera should satisfy the relationship f2 = N × f1 with the shooting frame rate f2 of the high-speed camera. N is an integer. In this system, N is the number of low-speed cameras.

5. A high time-resolution flame 3D reconstruction method, characterized by: The high-time-resolution flame 3D reconstruction method uses the high-time-resolution flame 3D reconstruction system as described in any one of claims 1 to 4, wherein a high-speed camera continuously shoots the flame at a high frame rate, and a trigger signal controller generates a trigger signal to control the low-speed cameras of each perspective to shoot sequentially along the circumference at a certain time interval, and ensures that at every moment the high-speed camera shoots, there is a low-speed camera of the same perspective shooting; a multi-perspective flame image prediction model associating the flame images of the high-speed camera and the low-speed camera is constructed through machine learning, and the high-time-resolution images of the perspectives of the low-speed cameras are obtained using the model; and a high-time-resolution 3D reconstruction of the flame structure is achieved through a 3D reconstruction algorithm.

6. The high time-resolution flame 3D reconstruction method according to claim 5, characterized in that: The algorithms for flame 3D reconstruction include ART (Algebraic Reconstruction Technique), SART (Simultaneous Algebraic Reconstruction Technique), MART (Multiplicative Algebraic Reconstruction Technique) algebraic iterative algorithms and various flame 3D reconstruction algorithms. The models used for iterative calculation include geometric optics model and wave optics model.

7. The high time-resolution flame 3D reconstruction method according to claim 6, characterized in that: A multi-view flame image prediction model is constructed by associating flame images from high-speed cameras and low-speed cameras through training using machine learning methods. The machine learning methods include deep learning (DL), deep reinforcement learning (DRL), Q-learning, domain adaptation, U-Net, support vector machine (SVM), convolutional neural network (CNN), deep neural network (DNN), generative adversarial network (GAN), long short-term memory network (LSTM), multi-task learning (MTL), fully convolutional network (FCN), K-means clustering, autoencoder, transfer learning, deep convolutional generative adversarial network (DCGAN), reinforcement learning, random forest, gradient boosting machine (GBDT) algorithm and their combination.

Images