CN-122023674-A - Real-time reconstruction method for temperature field of multi-source constraint hearth based on micro-rendering optimization

CN122023674ACN 122023674 ACN122023674 ACN 122023674ACN-122023674-A

Abstract

The invention discloses a real-time reconstruction method of a multisource constraint hearth temperature field based on micro-rendering optimization. The method comprises the steps of constructing a parameterized representation and differentiable forward radiation projection model of a three-dimensional temperature field corresponding to a combustion space of a hearth, collecting flame radiation images, thermocouple temperatures and water-cooled wall temperatures of the combustion space of the hearth, constructing a joint loss function comprising a reprojection error constraint term, a thermocouple deviation constraint term, a water-cooled wall deviation constraint term and a smoothness constraint term, calculating gradients of the joint loss function on three-dimensional voxel temperature parameters, and updating the three-dimensional voxel temperature parameters through an iterative optimization algorithm to obtain the three-dimensional temperature field of the hearth. The invention converts the traditional inverse radiation problem into the parameter optimization problem based on gradient, takes the thermocouple temperature and the water-cooled wall temperature as strong physical constraint, solves the uncertainty problem of three-dimensional reconstruction under a sparse view angle while guaranteeing the millisecond computing speed, and realizes high-precision and high-robustness online monitoring.

Inventors

WANG DI
HUANG QUNXING
GAO YULONG
WANG SHOUKANG
YIN XIAOYONG
FANG LI
WANG YAFEI
SHI ZHENG
WANG BIN
WU WEI

Assignees

上海华电电力发展有限公司望亭发电分公司
浙江大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. The real-time reconstruction method of the temperature field of the multi-source constraint hearth based on micro-rendering optimization is characterized by comprising the following steps of: S1, discretizing a hearth combustion space into a three-dimensional voxel grid according to a preset spatial resolution, taking a temperature parameter corresponding to each three-dimensional voxel as a continuously adjustable variable to be optimized, and constructing parameterized representation of a three-dimensional temperature field; S2, based on the Stefan-Boltzmann law, establishing a forward radiation projection model from a three-dimensional voxel grid to at least one two-dimensional imaging plane; S3, acquiring flame radiation images, thermocouple temperature and water-cooled wall temperature of a combustion space of a hearth in the same time window; S4, forward radiation projection is carried out based on the forward radiation projection model and the current three-dimensional voxel temperature parameters, a virtual image corresponding to the flame radiation image is generated, and a joint loss function is constructed, wherein the joint loss function comprises: the reprojection error constraint term is the mean square error of the radiation intensity between corresponding points in all the virtual images and the flame radiation image; A thermocouple deviation constraint term is the mean square error between all thermocouple temperatures and three-dimensional voxel temperature parameters corresponding to the thermocouple temperatures in the three-dimensional voxel grid; the water-cooled wall deviation constraint term is the mean square error between all water-cooled wall temperatures and three-dimensional voxel temperature parameters corresponding to the water-cooled wall temperatures in the three-dimensional voxel grid; The smooth constraint term is a total variation regularization term of the three-dimensional voxel grid; And S5, calculating the gradient of the joint loss function on the three-dimensional voxel temperature parameter, and updating the three-dimensional voxel temperature parameter based on the obtained gradient by adopting an iterative optimization mode until convergence or the preset iteration times are reached, so as to obtain the three-dimensional temperature field of the hearth at the current moment.
2. The method for reconstructing the temperature field of the multi-source constrained furnace based on micro-rendering optimization in real time according to claim 1, wherein in the step S2, the forward radiation projection model is constructed based on a discrete coordinate method or a ray stepping method, and specifically comprises the following steps: Wherein, the For the received radiation intensity of the corresponding pixel on the two-dimensional imaging plane, Is a geometrical projection matrix, S (T) is a source term, according to the Stefan-Boltzmann law, , Indicating the effective emissivity of the combustion medium in the furnace, Representing the stefin-boltzmann constant, Representing the absolute temperature of the corresponding three-dimensional voxel.
3. The method for reconstructing the temperature field of the multi-source constrained hearth based on micro-rendering optimization in real time according to claim 2 is characterized in that in the step S2, the calculation method of the geometric projection matrix A is as follows: Constructing a geometric mapping relation from three-dimensional voxels to two-dimensional pixels, determining a corresponding sight line direction for any pixel i on a two-dimensional imaging plane according to calibration parameters of an imaging device, traversing paths in a hearth three-dimensional voxel grid along the sight line direction, determining an intersection relation between the sight line direction and any three-dimensional voxel j and an effective propagation length of the intersection relation in the three-dimensional voxel j, calculating the sum of the effective propagation lengths of all three-dimensional voxels in the sight line direction, dividing the effective propagation length in the three-dimensional voxel j by the sum of the effective propagation lengths corresponding to the sight line direction to obtain a radiation transfer coefficient between the pixel i and the three-dimensional voxel j And collecting radiation transfer coefficients between all pixels and three-dimensional voxels to obtain a geometric projection matrix A, wherein the geometric projection matrix A is stored in a sparse tensor format.
4. The method for reconstructing the temperature field of the multi-source constraint furnace based on micro-rendering optimization according to claim 1, wherein in the step S4, the mathematical expression of the joint loss function is: Wherein, the For the joint loss function value, A radiation intensity matrix of the virtual image generated for the forward radiation projection, The radiation intensity matrix is the radiation image of the flame which is actually collected, N is the number of temperature measuring points, For the three-dimensional voxel temperature of the current iteration at the kth measurement point coordinate, The temperature is measured for the thermocouple of the kth measuring point; Representing a set of three-dimensional voxels attached to the physical location of the furnace water wall; Is a collection The temperature of the three-dimensional voxel of the current iteration of the p-th boundary voxel; Is a collection The water-cooled wall temperature corresponding to the p-th boundary voxel; the method is characterized in that a total variation regularization term of a three-dimensional voxel grid is adopted as a smoothness constraint term and is used for constraining the spatial continuity of a temperature field; 、、、 and the weight coefficients of the reprojection error constraint term, the thermocouple deviation constraint term, the water wall deviation constraint term and the smoothing constraint term are respectively calculated.
5. The method for reconstructing the temperature field of the multi-source constraint furnace based on micro-rendering optimization in real time is characterized in that in the joint loss function, a dynamic self-adaptive adjustment strategy is adopted for weight coefficients of a re-projection error constraint item, a thermocouple deviation constraint item, a water-cooled wall deviation constraint item and a smooth constraint item, wherein the weight of the re-projection error constraint item is given to be larger in the initial stage of iterative optimization, and the weights of the thermocouple deviation constraint item and the water-cooled wall deviation constraint item are dynamically increased along with the fact that the re-projection error tends to be stable in the middle and later stages of iteration.
6. The method for reconstructing the temperature field of the multi-source constrained furnace based on the micro-rendering optimization according to claim 1, wherein in the step S5, an Adam optimizer or a random gradient descent optimizer is adopted for the iterative optimization.
7. The method for reconstructing the temperature field of the multi-source constraint furnace chamber based on micro-rendering optimization in real time is characterized by further comprising the steps of S6, transmitting the reconstructed three-dimensional temperature field of the furnace chamber to a front-end visual terminal through a network interface, and S7, dynamically drawing the three-dimensional temperature field of the furnace chamber by using a volume rendering technology at the front-end visual terminal, and displaying temperature distribution of any section according to user interaction instructions.
8. The method for reconstructing the temperature field of the multi-source constrained furnace based on the micro-rendering optimization in real time according to claim 1, wherein the iterative optimization in the step S5 adopts a coarse-to-fine multi-scale grid reconstruction strategy, namely firstly forward radiation projection and gradient iteration are carried out on a three-dimensional voxel grid with lower spatial resolution, after the joint loss function is primarily converged, the temperature distribution of the grid with low spatial resolution is used as an initial temperature parameter of the grid with high spatial resolution by utilizing a three-dimensional interpolation algorithm, and gradient iteration under high spatial resolution is continued.
9. The method for reconstructing the temperature field of the multi-source constraint furnace based on micro-rendering optimization in real time according to claim 1, wherein when the parameterized representation of the three-dimensional temperature field is constructed in the step S1, three-dimensional voxels of a peripheral boundary layer attached to the physical position of a water wall of the furnace are set as limited variables, and the temperature of the water wall in the actual operation of the boiler is obtained as a boundary condition.
10. A multi-source constrained furnace temperature field real-time reconstruction system based on micro-rendering optimization, for implementing the method of any one of claims 1 to 9, comprising: the data acquisition module is used for acquiring a hearth flame radiation image, thermocouple temperature data and water-cooled wall temperature data; the edge computing module is internally provided with a graphic processor and is used for running the forward radiation projection model and the gradient optimization algorithm and resolving a three-dimensional temperature field in real time by combining the data transmitted by the data acquisition module; And the visual interaction module is used for receiving the three-dimensional temperature field data calculated by the edge calculation module, and performing three-dimensional visual rendering and man-machine interaction display.

Description

Real-time reconstruction method for temperature field of multi-source constraint hearth based on micro-rendering optimization Technical Field The invention relates to the technical field of industrial combustion monitoring and diagnosis, in particular to a multisource constraint hearth temperature field real-time reconstruction method based on micro-rendering optimization by combining a computer vision, a radiation heat transfer mechanism and a numerical optimization technology. Background The distribution of the combustion temperature field in the furnace of the coal-fired boiler is an important index reflecting the combustion stability, the combustion efficiency and the pollutant emission level. The traditional hearth temperature monitoring means mainly comprise contact measurement (such as a thermocouple), acoustic wave temperature measurement and the like. However, the thermocouple can only acquire single-point temperature, cannot reflect three-dimensional temperature distribution of the whole hearth, has short service life in a high-temperature environment, and has low spatial resolution and is easily interfered by fly ash and soot blowing noise although the acoustic temperature measurement can acquire section temperature. In recent years, three-dimensional temperature field reconstruction techniques (i.e., inverse radiation problem solving) based on flame radiation images have received widespread attention. The existing reconstruction methods are mainly divided into two categories: The traditional algebraic iteration method is to establish a radiation equation set through a Discrete Transfer Method (DTM) or a Monte Carlo method, and solve an inverse matrix through algorithms such as a Least Squares Method (LSMR), singular Value Decomposition (SVD) and the like. The method has extremely high computational complexity, relates to large-scale matrix inversion, is difficult to meet the requirement of millisecond-level real-time control, and has extremely unstable reconstruction results under the condition of a small number of cameras (sparse view angles) and serious pathological conditions of an equation set. The deep learning method driven by pure data is to directly establish the mapping from the image to the temperature field by using a Convolutional Neural Network (CNN). Such methods rely heavily on massive amounts of high-precision tag data (typically from CFD simulations), and models lack physical interpretability, and generalization ability is poor when actual conditions differ significantly from training data. Therefore, a new reconstruction method is needed that can ensure accuracy by using a physical mechanism, can realize high-speed calculation by using modern hardware, and can integrate existing sensor data to improve robustness. Disclosure of Invention The invention provides a method and a system for reconstructing a multi-source constraint hearth temperature field in real time based on micro-rendering optimization in order to solve the problems in the prior art. The invention innovatively introduces a micro-rendering (Differentiable Rendering) technology, converts the ill-condition inverse radiation problem into a good-condition gradient-based parameter optimization problem, and utilizes thermocouple temperature and water-cooled wall temperature data as physical anchor points, thereby realizing high-precision and real-time three-dimensional temperature field reconstruction. The technical scheme of the invention is as follows: a real-time reconstruction method of a multi-source constraint hearth temperature field based on micro-rendering optimization comprises the following steps: S1, discretizing a hearth combustion space into a three-dimensional voxel grid according to a preset spatial resolution, taking a temperature parameter corresponding to each three-dimensional voxel as a continuously adjustable variable to be optimized, and constructing parameterized representation of a three-dimensional temperature field; Step S2, based on the Stefan-Boltzmann law, a forward radiation projection model from the three-dimensional voxel grid to at least one two-dimensional imaging plane is established, and the forward radiation projection model is continuously conductive to the voxel temperature parameters, so that the pixel radiation intensity on the two-dimensional imaging plane can perform gradient calculation relative to the temperature parameters of the three-dimensional voxels; step S3, synchronously acquiring flame radiation images acquired by at least one imaging device arranged on the periphery of the hearth, actual measurement temperatures acquired by at least one thermocouple temperature sensor arranged in the hearth or on the wall surface of the hearth and water-cooling wall temperatures acquired by at least one thermocouple sensor arranged on the water-cooling wall in the same time window; S4, forward radiation projection is carried out based on the forward radiation projection model and th