A multi-source infrared-based copper-smelting anode furnace brick thickness online detection and life prediction method
By combining multi-source infrared thermal field fusion and deep learning with physical constraints, online detection and life prediction of furnace brick thickness in copper smelting anode furnaces were achieved. This solved the monitoring difficulties of traditional methods under complex operating conditions, provided high-precision and stable prediction results, and supported long-term operation and preventive maintenance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YUNNAN COPPER CO LTD
- Filing Date
- 2026-01-06
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies struggle to achieve non-contact, online, continuous monitoring of furnace brick thickness in copper smelting anode furnaces, and they also struggle to obtain a stable temperature field and accurate thickness prediction under complex operating conditions. In particular, traditional methods are susceptible to interference and lack physical constraints in high-temperature, high-radiation, and dusty environments.
By employing multi-source infrared thermal field fusion, physical consistency deep inversion, and furnace body zoning modeling, and combining multi-view infrared imaging data fusion, dynamic emissivity compensation, flue gas obstruction identification, and temperature field purification with deep regression networks and time series prediction, a thickness inversion model is constructed to achieve online detection of furnace brick thickness and life prediction.
It achieves high-precision and stable monitoring and prediction of furnace brick thickness under complex working conditions, reduces reliance on parameters that are difficult to measure in real time, has adaptive capabilities, supports long-term stable operation, and provides a basis for preventive maintenance decisions.
Smart Images
Figure CN122108028A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial equipment health monitoring and intelligent diagnosis technology, specifically a method for online detection of furnace brick thickness and life prediction of copper smelting anode furnace based on multi-source infrared radiation. Background Technology
[0002] Copper smelting anode furnaces need to operate in a high-temperature environment of 1150–1250 °C for extended periods during actual production, undergoing multiple stages of operation including oxidation, reaction, and refining. The furnace temperature, slag composition, oxygen levels, and material agitation methods often change continuously with the operating rhythm, resulting in significant time-varying and complex stress states and thermal environments in the furnace lining.
[0003] Furnace brick materials (such as magnesia-carbon bricks, magnesia-chrome bricks, and alumina-silicon bricks) are subjected to the combined effects of multiple factors in such an environment. First, molten copper and high-temperature slag directly wet the brick body, causing chemical reactions and dissolving the surface structure. Second, the liquid flow and strong gas flow near the copper tapping area and the redox tapping area continuously scour the brick body, easily forming areas of accelerated local wear. In addition, temperature fluctuations caused by furnace start-up and shutdown, charging, etc., cause the brick body to frequently experience thermal expansion and contraction, resulting in thermal shock cracks. Long-term contact between the surface and oxidizing flue gas will cause oxidation and spalling, reducing the mechanical strength of the brick body. In some furnace runs with uneven operating conditions, fluctuations in the charge level, molten impact, or changes in furnace pressure can also create stress concentrations in local areas of the brick body, inducing the propagation of microcracks. The superposition of multiple mechanisms results in the brick thickness exhibiting typical characteristics of "non-linearity, significant regional differences, and gradual decay over time". If the remaining thickness of the bricks cannot be determined in time and its future trend cannot be predicted, serious accidents such as perforation and leakage may occur in the furnace body during peak production periods. Therefore, thickness monitoring and life prediction are crucial to smelting safety.
[0004] Based on the aforementioned current situation and problems, it is evident that the smelting industry's demand for furnace brick thickness monitoring technology far exceeds what traditional methods can meet. Enterprises desire an online monitoring method capable of continuous operation without interrupting the furnace, enabling real-time capture of thickness changes without relying on furnace shutdown or manual measurement. Simultaneously, the smelting environment is extremely complex, with factors such as smoke and dust obstruction, flame flickering, and high-gloss reflection from metal being prevalent. Therefore, the monitoring method must be able to cope with highly disruptive operating conditions and extract reliable and clean temperature information from unstable infrared images. Since single-view infrared thermal images often fail to cover the entire furnace wall and are easily affected by local anomalies, multi-view data fusion becomes a crucial means of reconstructing a stable temperature field. Furthermore, the thickness inversion process cannot rely solely on statistical relationships between data; it should incorporate heat conduction laws to ensure the prediction results are physically reliable and consistent. Moreover, the heat load and loss mechanisms differ significantly across different regions of the furnace body; the erosion rates of the copper taphole, redox taphole, sidewalls, and furnace bottom are fundamentally different. Therefore, monitoring and modeling methods should also reflect these structural differences.
[0005] In practical applications, companies are not only concerned with the current thickness status but also with the decay trend over the next few days or even weeks, hoping to predict in advance whether the furnace bricks will be at risk before the next maintenance cycle. Therefore, the monitoring system should have the ability to identify trends in thickness changes and predict lifespan. Meanwhile, changes in furnace age, fluctuations in operating conditions, and evolution of surface conditions can all cause drift in the statistical characteristics of the temperature field, requiring the system to be able to self-update and automatically correct to adapt to data changes during long-term operation. Finally, a system truly suitable for industrial deployment must minimize its reliance on parameters that are difficult to measure in real time, such as internal temperature and convection coefficients, and maintain stable operation in long-term, high-intensity, and high-dust environments.
[0006] Therefore, the smelting industry urgently needs a novel solution truly applicable to industrial environments. This approach must meet multiple core requirements: First, it must achieve non-contact, online continuous monitoring without furnace shutdown; second, it must effectively integrate multi-view infrared information and suppress complex interference to obtain a stable and reliable temperature field; more importantly, thickness inversion must deeply integrate the physical mechanisms of heat conduction, not just data fitting, to ensure the physical consistency and adaptability of the model; simultaneously, it should be able to model differentiated erosion patterns in different regions of the furnace body and possess the ability to predict thickness evolution and assess remaining life based on time-series data; furthermore, the system must have online self-correction and adaptive capabilities to cope with long-term operational condition and state drift. Ultimately, the solution should minimize reliance on internal parameters that are difficult to measure in real time, forming a highly reliable, predictable, and long-term stable intelligent monitoring system. Summary of the Invention
[0007] To address the aforementioned issues, this invention provides a method for online detection and life prediction of furnace brick thickness in copper smelting anode furnaces based on multi-source infrared radiation. This invention combines a comprehensive methodology of multi-source infrared thermal field fusion, physical consistency inversion, furnace body zoning modeling, and thickness time-series prediction. It aims to provide reliable, continuous, and predictable furnace brick health monitoring capabilities under complex operating conditions, thereby significantly improving the safety and operational efficiency of smelting equipment.
[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for online detection of furnace brick thickness and life prediction of copper smelting anode furnace based on multi-source infrared radiation, specifically including the following steps: S1. Multi-source infrared thermal field acquisition and dynamic emissivity compensation: Collect multi-view infrared imaging data and furnace brick material parameters. After feature point matching and registration based on the brightness consistency of multi-view images, construct a dynamic emissivity estimation model. Use background radiation temperature to correct the original infrared temperature measurement to obtain the multi-view compensated surface temperature field. Perform weighted fusion on the multi-view compensated surface temperature field to output the multi-view compensated temperature field. The multi-view infrared imaging data consists of furnace outer wall temperature images collected simultaneously by at least four different-view infrared thermal imagers. The material parameters of the furnace bricks include: thermal conductivity, density, and specific heat capacity; The key components are the copper port area and the redox port area; This invention constructs a dynamic emissivity estimation model based on brightness uniformity, surface texture features, and brick material properties. The expression is: in, Based on the material parameters of the furnace bricks The reference emissivity is determined by factors such as thermal conductivity, density, and specific heat. These are the material parameters for the furnace bricks. This indicates the brightness difference of the corresponding pixels between this viewpoint and adjacent viewpoints after registration. The texture roughness index is calculated from local gray-level variance, gradient, etc. The first empirical coefficient, This is the second empirical coefficient, which is obtained through calibration experiments. It represents three-dimensional space and time coordinates; After obtaining the dynamic emissivity estimation model, the background radiation temperature is used. The original infrared thermometry is corrected to obtain a multi-view compensated surface temperature field. The expression is: The multi-view compensated surface temperature field is obtained by weighted fusion. The expression is: in, The field of view number is assigned to the infrared camera. These are the weighting coefficients.
[0009] S2. Smoke Occlusion Recognition and Temperature Field Purification: Based on the multi-view compensated surface temperature field, the motion estimation of two adjacent frames is performed by using a dense optical flow algorithm, and the optical flow amplitude, root mean square contrast and luminance variance are calculated. After inputting into the infrared occlusion detection network to obtain the output feature map, the median interpolation of the adjacent frames within the time window is used by setting interference judgment to output the purified real temperature field. First, a dense optical flow algorithm is used to estimate motion between two adjacent frames. Then, a constant brightness constraint and a multi-scale pyramid model are used to solve for the velocity vector field at each pixel. The expression is as follows: in, For the optical flow velocity vector field, Define the horizontal component of the optical flow and calculate its amplitude; Secondly, to depict "local brightness variations" and texture features, adjustments are made to each frame. Calculate the value centered at this pixel, with a window size of [value missing]. Local mean of a pixel Contrast with root mean square The expression is as follows: in, For local windows, These are local coordinates within the window. The number of pixels within the window. This is the local mean. Then, the brightness variance over time is calculated for the same window. , used to characterize time-domain stability, is expressed as follows: in, This is the time average of the local mean within this time window.
[0010] After obtaining the above features, the input features are constructed as follows: An infrared occlusion detection network. Output feature map. , representing the pixel category probabilities of background, smoke, flame, and highly reflective areas, respectively, with the output being a probability map. During training, manually labeled pixel-level interference masks are used as labels, and cross-entropy loss is employed for supervised learning.
[0011] The interference determination logic of this invention is set as follows: when the probability of smoke pixel category is... When the pixel is obscured by smoke, it is classified as a smoke-obstructed pixel; when the probability of the flame pixel category is... When the pixel type probability of a high-reflectivity area is high, it is determined to be a flame area; when the pixel type probability of a high-reflectivity area is high, it is determined to be a flame area. When a pixel is identified as a high-reflectivity point, the rest are classified as background, thus obtaining a multi-class segmentation mask. For areas marked as smoke or flame, median interpolation within the time window is used to output the purified true temperature field, as shown in the following expression: in, This represents the number of neighboring frames. For highly reflective areas, utilize the surrounding area. Region inpainting is performed using the weighted average of unlabeled pixels in the neighborhood. Simultaneously, the estimated radiative background component is subtracted from all frames to complete radiative background subtraction. The result obtained after the above processing... This is the purified temperature field after the interference of smoke, flame, and reflection has been suppressed, and it is used as the input for the subsequent furnace brick thickness inversion model.
[0012] S3. Constructing a deep regression network for the physical consistency constraint module: Based on the purified real temperature field, a deep regression network is constructed to extract multi-scale thermal features. After embedding the physical consistency constraint module, the theoretical heat flux density is calculated based on the steady-state heat conduction relationship. The total loss function is then constructed for training, and the predicted furnace brick thickness distribution map is output, thus completing the construction of the furnace brick thickness inversion model based on the physical consistency constraint module. The total loss function includes the physical residual loss function, the data fitting loss function, and the spatial smoothing loss function; The inputs to the physical consistency constraint module include: firstly, the preliminary thickness prediction feature map output from the intermediate layer of the thickness regression backbone network. Secondly, the surface temperature field after purification at the corresponding location. and the temperature of the gas inside the furnace This module is based on steady-state heat conduction relationships and theoretical heat flux density. The expression is as follows: in, The thermal conductivity of the furnace brick material. The surface temperature field after purification. The temperature of the gas inside the furnace. This is the initial thickness prediction feature map. A pixel-level physical consistency residual is constructed, and further combined with the thickness spatial gradient. Constrain the smoothness of the thickness field.
[0013] During the training phase, the output of the physical consistency constraint module is represented as a physical residual loss. Its relationship with data fitting loss and spatial smoothing loss Together they constitute the total loss function, expressed as follows: in, The weighting coefficients for spatial smoothing loss are... This is the weighting coefficient for physical residual loss; The total loss function is used to optimize the network parameters through backpropagation; during the prediction phase, the physical consistency constraint module does not directly output the scalar loss value, but instead optimizes it by... The implicit constraint effect outputs a predicted furnace brick thickness distribution map that satisfies the physical laws of heat conduction. The thickness prediction results obtained in this way are consistent in both statistical significance and thermal mechanism, thus significantly improving the stability and reliability of the model under complex working conditions.
[0014] Based on infrared imaging data, the expression is as follows: This embodiment constructs a brick thickness inversion model based on a convolutional neural network (CNN), achieving pixel-level mapping from the surface temperature field to the thickness field. The network input is the surface temperature distribution after emissivity compensation and interference removal. The output is the predicted value of the remaining thickness of the furnace brick at the corresponding location. .
[0015] To simultaneously constrain data fitting accuracy, spatial smoothness, and consistency with the physical laws of heat conduction, the model's total loss function consists of three parts: a data loss term, a smoothing constraint term, and a physical constraint term. The data loss term is defined as follows: in, The total number of pixels in the training samples. For the first Predicted thickness per pixel This represents the measured thickness at the corresponding location.
[0016] The smoothing constraint term is used to suppress abrupt changes in thickness caused by local noise, and is defined as follows: in, Indicates the predicted thickness field in the th... Spatial gradient at each pixel.
[0017] The physical constraints are constructed based on steady-state heat conduction relationships. The expression for the theoretical heat flux density of a pixel is as follows: in, The thermal conductivity of the furnace bricks, The temperature of the gas inside the furnace (measured by the on-site DCS system and taken as the average value of each furnace). For the first The outer wall surface temperature at each pixel. Comparing the theoretical heat flux density with the reference heat flux density. By comparison, the physical residual loss is constructed, as expressed below: In this invention, the weight of the smoothing constraint term is... The range of values is Physical constraint weights The range of values is To achieve a balance between fitting accuracy and physical consistency. After training, the model obtains the following mapping relationship: in, For including parameters A deep learning thickness inversion model.
[0018] S4. Furnace body zoning modeling and fusion based on structural differences: Based on the predicted furnace brick thickness distribution map and furnace body region division map, thickness inversion sub-models are trained according to the furnace body region division map. By cropping the temperature field image according to the preset region and inputting it into the corresponding sub-model, the optimized thickness map of each region is obtained. Then, the outputs of multiple models are weighted and fused by defining the region sensitivity and prediction confidence to output the thickness distribution map of the whole furnace. The furnace body area division diagram includes: copper tapping area, redox tapping area, sidewall area, and furnace bottom area; This invention is used to divide pixel-level region label maps into different areas. The following methods were used to obtain the image: First, spatial calibration was performed based on the furnace body CAD structural drawing and the infrared camera imaging coordinate system. The projection boundaries of the copper tapping area, redox tapping area, sidewall area, and furnace bottom area in the infrared image were determined by manual annotation. Second, the high-temperature zone of the copper tapping area and the periodically high-fluctuation area of the redox tapping area were supplemented by combining the long-term temperature mean field, thereby obtaining a pixel-level region label map. The expression of the pixel-level region label map is as follows: These correspond to the copper tapping area, the redox tapping area, the sidewall area, and the furnace bottom area, respectively.
[0019] To address the differences in erosion scale across different regions, this invention jointly regulates the receptive field size by adjusting the input image preprocessing scale and the network structure depth, thereby constructing sub-models with different feature scales, as detailed below: The input scale of the copper mouth area sub-model is The network employs a 6-layer convolutional structure to characterize local erosion features dominated by small-scale, strong scouring. The input scale of the redox sub-model is The network uses a 7-layer convolutional structure; The input scale for both the sidewall region and the furnace bottom region sub-models is [scale value missing]. The network employs an 8-layer convolutional structure to characterize the thickness evolution features dominated by large-scale, slow erosion. Therefore, "having different receptive fields and feature scales" is reflected both in the multi-scale pyramid cropping method of the input image and in the differentiated design of the network structure depth.
[0020] The training data for each regional sub-model comes from: purification temperature sub-maps cropped according to regional labels. ; Corresponding thickness label ; The thickness of the furnace bricks was obtained by aligning the measured data from borehole drilling during the furnace shutdown and maintenance with the temperature field over time. In this embodiment, each regional sub-model was trained independently and its parameters were optimized separately. During the overall furnace thickness prediction stage, each sub-model only participated in joint inference and weighted fusion, and no further reverse updates were performed.
[0021] Define the region sensitivity coefficient This is used to represent the relative importance of different functional areas in furnace safety, and its specific value is: .
[0022] The thickness prediction results output by each sub-model are denoted as: And simultaneously output the prediction confidence map: The weighted fusion function expression is constructed as follows: in, For regional indexes, For summation index; The final overall furnace thickness distribution diagram is as follows: S5. Spatiotemporal evolution modeling and remaining life prediction: Based on the thickness inversion model, the thickness distribution map of the whole furnace and the minimum safe thickness, the thickness change sequence is constructed and input into the time series prediction model to predict the thickness at the next moment. Then, the thickness loss rate of the furnace bricks is calculated to predict the remaining life of the furnace body. This invention will monitor each point in time. The thickness is expressed as: in, For a moment Infrared temperature field, This is a thickness inversion model.
[0023] Constructing a thickness variation sequence: The thickness evolution is modeled using a Long Short-Term Memory (LSTM) network to predict the thickness at the next time step, as shown in the following expression: in, This is a time series prediction model.
[0024] Calculate the brick thickness loss rate based on the prediction results: in, It is a differential operator; Assume the minimum safe thickness of the furnace bricks is The formula for predicting the remaining life of the furnace body is: in, This represents the average loss rate due to recent thickness variations.
[0025] The system calculates the remaining operating time of the furnace body based on the furnace body remaining life prediction formula. A maintenance warning is triggered when the value is below a preset threshold.
[0026] S6. Based on the thickness inversion model and time series prediction model, a method and system for online thickness detection and lifetime prediction are completed through online drift monitoring and self-correction mechanism and online deployment and dynamic updating. Compared with existing technologies, this invention provides a method for online detection of furnace brick thickness and life prediction of copper smelting anode furnace based on multi-source infrared radiation, which has the following beneficial effects: (1) Through multi-view infrared fusion, dynamic emissivity compensation and active purification technology of smoke and flame interference, the system can obtain a stable furnace wall temperature field without stopping production, which solves the defects of traditional methods that require furnace shutdown, rely on manual labor and have poor anti-interference ability. (2) This invention achieves high-precision thickness spatial distribution inversion by embedding the heat conduction mechanism as a constraint into a deep learning model and constructing differentiated sub-models for different regions of the furnace body for fusion. Furthermore, based on the time-series evolution model, the thickness decay trend is predicted, realizing the process from current state perception to remaining life assessment, providing key decision-making basis for preventive maintenance, and proposing a paradigm of thickness inversion and life prediction that deeply integrates physical constraints and data-driven approaches; (3) The system of the present invention can adapt to changes in operating conditions and equipment aging through an online drift monitoring mechanism, and automatically update the model to maintain accuracy. At the same time, the method of the present invention significantly reduces the dependence on internal boundary conditions that are difficult to obtain in real time, and supports edge computing deployment through modular design, making it easy to integrate with existing monitoring systems, forming a complete solution that can be implemented and operated sustainably. Attached Figure Description
[0027] Figure 1 This is a schematic diagram of the furnace brick thickness detection and life prediction method and system structure of the present invention; Figure 2 This is a temperature distribution diagram of the outer wall of the furnace body of the present invention; Figure 3 This is a flowchart illustrating the furnace brick thickness inversion process of the present invention. Figure 4 This is a schematic diagram of the thickness time series and loss trend of the present invention. Detailed Implementation
[0028] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0029] This invention stems from the long-standing difficulties and operational safety hazards encountered in monitoring the thickness of furnace bricks in actual copper smelting anode furnace production. As the most critical refractory material in the furnace body, the thickness variation of furnace bricks directly affects the structural stability, operational safety, and production continuity. However, under conditions of high temperature, strong radiation, abundant smoke and dust, and complex fluctuations in operating conditions, traditional thickness measurement methods struggle to achieve accurate, continuous, and spatially complete monitoring of furnace brick thickness without interrupting production. Furthermore, due to significant differences in heating conditions and erosion mechanisms across different parts of the furnace body, the wear rate exhibits obvious regional non-uniformity, making single-point measurements or simple calculations insufficient to accurately reflect the overall health status of the furnace lining.
[0030] While existing infrared detection technology can achieve non-contact temperature measurement, it is susceptible to factors such as changes in emissivity, strong reflection, and dust obstruction, often resulting in significant errors when directly inferring internal thickness from surface temperature. Furthermore, data-driven deep learning models lack stability in smelting environments and are prone to prediction inaccuracies under varying operating conditions when physical constraints are lacking, making them unsuitable for long-term online operation. Therefore, there is an urgent need to organically combine the advantages of infrared detection with heat conduction mechanisms, deep learning methods, and furnace structural characteristics to achieve high-precision and highly adaptable furnace brick thickness inversion.
[0031] This invention takes the physical mapping relationship between temperature field and thickness as its theoretical basis, as follows: On the outer wall of the furnace, the heat transfer process of the furnace bricks can be described by the unsteady-state heat conduction equation: in, ρ For the density of furnace bricks, c p For specific heat capacity, The thermal conductivity of the furnace brick material. T(x,y,t) Indicates the temperature field of the furnace bricks. This is the term for heat conduction.
[0032] When the furnace body is in a relatively stable operating range, local heat transfer can be approximated as a steady-state process, at which point the heat flux density satisfies Fourier's law: In the formula, Indicates the temperature of the melt inside the furnace. The surface temperature of the outer wall of the furnace. This represents the remaining thickness of the furnace bricks.
[0033] From this steady-state heat flow relationship, the inverse relationship between the thickness of the furnace bricks and the temperature of the outer wall can be derived, and the remaining furnace brick thickness... The expression is: in, As the first empirical parameter, The second empirical parameter is obtained from on-site calibration experiments. In this embodiment, taking a 350-ton copper smelting anode furnace as an example, 52 sets of sample data of the furnace brick thickness measured by drilling during the furnace shutdown and maintenance period and the corresponding outer wall temperature were selected. The least squares method was used to fit the above functional relationship, and the following results were obtained: This qualitative relationship provides a good physical constraint for deep learning models, enabling them to still have a reasonable predictive trend under data noise or operating condition fluctuations.
[0034] Example 1: A method for online detection of furnace brick thickness and life prediction in copper smelting anode furnace based on multi-source infrared radiation; Please see Figures 1-4 A method for online detection of furnace brick thickness and life prediction in copper smelting anode furnace based on multi-source infrared radiation, specifically including the following steps: S1. Multi-source infrared thermal field acquisition and dynamic emissivity compensation: Collect multi-view infrared imaging data and furnace brick material parameters. After feature point matching and registration based on the brightness consistency of multi-view images, construct a dynamic emissivity estimation model. Use background radiation temperature to correct the original infrared temperature measurement to obtain the multi-view compensated surface temperature field. Perform weighted fusion on the multi-view compensated surface temperature field to output the multi-view compensated temperature field. This invention deploys four infrared thermal imagers with different viewing angles at key locations on the outer wall of the anode furnace, simultaneously acquiring the temperature field of the furnace surface from different perspectives to obtain the original brightness and temperature images from each perspective. ;like Figure 2 As shown, multi-view imaging not only expands the monitoring coverage but also corrects temperature anomalies through redundant information between viewpoints, improving the stability of temperature field data. Considering that the surface emissivity of furnace bricks drifts with oxidation, ash accumulation, slag buildup, and furnace age, this invention constructs a dynamic emissivity estimation model based on brightness consistency, surface texture features, and brick material properties. The expression is: in, Based on the material parameters of the furnace bricks The reference emissivity is determined by factors such as thermal conductivity, density, and specific heat. These are the material parameters for the furnace bricks. This indicates the brightness difference of the corresponding pixels between this viewpoint and adjacent viewpoints after registration. The texture roughness index is calculated from local gray-level variance, gradient, etc. The first empirical coefficient, This is the second empirical coefficient, which is obtained through calibration experiments. It represents three-dimensional space and time coordinates.
[0035] After obtaining the dynamic emissivity estimation model, the background radiation temperature is used. The original infrared thermometry is corrected to obtain a multi-view compensated surface temperature field. The expression is: This invention assumes that infrared radiation approximately satisfies the Stefan-Boltzmann relation and uses a fourth power form for inversion.
[0036] To further eliminate local occlusion and measurement noise, this invention constructs weighting coefficients based on the signal-to-noise ratio of imaging from each viewpoint, the observation angle, and the occlusion determination results. The compensated temperatures from each field of view are fused to obtain the final multi-view compensated temperature field. The expression is: in, The field of view of the infrared camera is numbered in this embodiment. =1, 2, 3, 4; thus forming a steady-state surface temperature field dataset after dynamic emissivity correction and multi-view fusion. This serves as the input for subsequent flue gas interference identification and thickness inversion models.
[0037] S2. Smoke Occlusion Recognition and Temperature Field Purification: Based on the multi-view compensated surface temperature field, the motion estimation of two adjacent frames is performed by using a dense optical flow algorithm, and the optical flow amplitude, root mean square contrast and luminance variance are calculated. After inputting into the infrared occlusion detection network to obtain the output feature map, the median interpolation of the adjacent frames within the time window is used by setting interference judgment to output the purified real temperature field. Because smelting sites commonly experience interference factors such as smoke and dust dispersion, flame flickering, and metal reflection, these factors can generate significant noise and false temperature hotspots in infrared images. Therefore, this invention addresses this issue by using a length of... (This embodiment takes) (time window) Internal selection Surface temperature field after frame compensation , For the frame index within the time window, The current frame index of the time series is used, and an infrared occlusion detection network based on optical flow and local contrast features is constructed to perform pixel-level identification of smoke, flames and high-radiation reflectivity points.
[0038] First, a dense optical flow algorithm is used to estimate motion between two adjacent frames. This embodiment uses the Farnebäck dense optical flow method, employing constant brightness constraints and a multi-scale pyramid model to solve for the velocity vector field at each pixel. The expression for calculating the optical flow amplitude is shown below: in, For the optical flow velocity vector field, This refers to the horizontal component of optical flow. Secondly, to depict "local brightness variations" and texture features, adjustments are made to each frame. Calculate the value centered at this pixel, with a window size of [value missing]. Local mean of a pixel Contrast with root mean square The expression is as follows: in, For local windows, These are local coordinates within the window. The number of pixels within the window. This is the local mean. Then, the brightness variance over time is calculated for the same window. , used to characterize time-domain stability, is expressed as follows: in, This is the time average of the local mean within this time window.
[0039] After obtaining the above features, the input features are constructed as follows: An infrared occlusion detection network is described. This network employs an encoder-decoder structure: the encoder contains three convolutional modules (each module consists of two layers). Convolution + ReLU activation + The max pooling algorithm consists of 32, 64, and 128 channels respectively. The decoder uses three upsampling modules (bilinear upsampling + convolution), and finally passes through one layer. Convolution yields a 4-channel output feature map. , representing the pixel category probabilities of background, smoke, flame, and highly reflective areas, respectively, with the output being a probability map. During training, manually labeled pixel-level interference masks are used as labels, and cross-entropy loss is employed for supervised learning.
[0040] The interference determination logic of this invention is set as follows: when the probability of smoke pixel category is... Time (in this embodiment, the flue gas determination threshold is used) The probability of classifying the smoke pixel as smoke obstruction is determined; when the probability of classifying the flame pixel is... Time (Flame Detection Threshold) The region is identified as a flame area; the pixel category probability of a high-reflectivity area is... Time (reflection determination threshold) The pixel 0 is identified as a high-reflectivity point; the remaining pixels are classified as background based on their probability categories, thus obtaining a multi-class segmentation mask. For areas marked as smoke or flame, median interpolation within the time window is used to output the purified true temperature field, as shown in the following expression: in, This represents the number of neighboring frames. (This embodiment takes) To restore the temperature of the blocked area; for high-reflectivity areas, utilize the surrounding area. Region inpainting is performed using the weighted average of unlabeled pixels in the neighborhood. Simultaneously, the estimated radiative background component is subtracted from all frames to complete radiative background subtraction. The result obtained after the above processing... This is the purified temperature field after the interference of smoke, flame, and reflection has been suppressed, and it is used as the input for the subsequent furnace brick thickness inversion model.
[0041] S3. Constructing a deep regression network for the physical consistency constraint module: Based on the purified real temperature field, a deep regression network is constructed to extract multi-scale thermal features. After embedding the physical consistency constraint module, the theoretical heat flux density is calculated based on the steady-state heat conduction relationship. The total loss function is then constructed for training, and the predicted furnace brick thickness distribution map is output, thus completing the construction of the furnace brick thickness inversion model based on the physical consistency constraint module. The total loss function includes the physical residual loss function, the data fitting loss function, and the spatial smoothing loss function; In the thickness prediction stage, this invention constructs a deep learning inversion model that includes a physical consistency constraint module, achieving pixel-level mapping from the purified surface temperature field to the brick thickness field. The model as a whole adopts a cascaded structure of "temperature feature extraction network + thickness regression backbone network + physical consistency constraint module". Specifically, the thickness regression backbone network uses the purified surface temperature field... As input, multi-scale thermal features are extracted and a preliminary thickness prediction feature map is output. .
[0042] The inputs to the physical consistency constraint module include: firstly, the preliminary thickness prediction feature map output from the intermediate layer of the thickness regression backbone network. Secondly, the surface temperature field after purification at the corresponding location. and the temperature of the gas inside the furnace This module is based on steady-state heat conduction relationships and theoretical heat flux density. The expression is as follows: in, The thermal conductivity of the furnace brick material. The surface temperature field after purification. The temperature of the gas inside the furnace. This is the initial thickness prediction feature map. A pixel-level physical consistency residual is constructed, and further combined with the thickness spatial gradient. Constrain the smoothness of the thickness field.
[0043] During the training phase, the output of the physical consistency constraint module is represented as a physical residual loss. Its relationship with data fitting loss and spatial smoothing loss Together they constitute the total loss function, expressed as follows: in, The weighting coefficients for spatial smoothing loss are... This is the weighting coefficient for physical residual loss; The total loss function is used to optimize the network parameters through backpropagation; during the prediction phase, the physical consistency constraint module does not directly output the scalar loss value, but instead optimizes it by... The implicit constraint effect outputs a predicted furnace brick thickness distribution map that satisfies the physical laws of heat conduction. The thickness prediction results obtained in this way are consistent in both statistical significance and thermal mechanism, thus significantly improving the stability and reliability of the model under complex working conditions.
[0044] This embodiment focuses on the areas with the most concentrated heat load in copper smelting anode furnaces, namely the copper tapping area and the redox tapping area. In these locations, furnace brick erosion is typically more pronounced and is a key focus for monitoring and prediction. This embodiment uses an infrared thermal imager to acquire the temperature distribution of the furnace outer wall in real time and combines this with on-site manual thickness measurement data to establish a mapping relationship between temperature and thickness. Figure 3 As shown, based on this, a deep learning model is used to extract thermal features from the temperature field to achieve non-contact inversion of the thickness of the furnace bricks, providing basic data for subsequent thickness evaluation and trend analysis.
[0045] Based on infrared imaging data, the expression is as follows: This embodiment constructs a brick thickness inversion model based on a convolutional neural network (CNN), achieving pixel-level mapping from the surface temperature field to the thickness field. The network input is the surface temperature distribution after emissivity compensation and interference removal. The output is the predicted value of the remaining thickness of the furnace brick at the corresponding location. .
[0046] To simultaneously constrain data fitting accuracy, spatial smoothness, and consistency with the physical laws of heat conduction, the model's total loss function consists of three parts: a data loss term, a smoothing constraint term, and a physical constraint term. The data loss term is defined as follows: in, The total number of pixels in the training samples. For the first Predicted thickness per pixel This represents the measured thickness at the corresponding location.
[0047] The smoothing constraint term is used to suppress abrupt changes in thickness caused by local noise, and is defined as follows: in, Indicates the predicted thickness field in the th... Spatial gradient at each pixel.
[0048] The physical constraints are constructed based on steady-state heat conduction relationships. The expression for the theoretical heat flux density of a pixel is as follows: in, The thermal conductivity of the furnace bricks, The temperature of the gas inside the furnace (measured by the on-site DCS system and taken as the average value of each furnace). For the first The outer wall surface temperature at each pixel. Comparing the theoretical heat flux density with the reference heat flux density. By comparing the results, a physical residual loss function is constructed, expressed as follows: In this embodiment, the weight of the smoothing constraint term The range of values is Finally selected Weights of physical constraint terms The range of values is Finally selected To achieve a balance between fitting accuracy and physical consistency.
[0049] In this embodiment, the network is trained using the Adam optimization algorithm, with an initial learning rate set to... The batch size is 16, and the maximum number of training epochs is 200. Training is terminated early when the validation set loss no longer decreases within 20 consecutive epochs to prevent overfitting. Training and validation data are randomly divided in an 8:2 ratio.
[0050] After training, the model obtains the following mapping relationship: in, For including parameters A deep learning-based thickness inversion model is proposed. Based on the loss function design described above, the model ensures prediction accuracy while also taking into account the spatial smoothness of the thickness field and the physical consistency of thermal conduction. As a result, it can still output stable and reliable thickness prediction results under noise interference and operating condition fluctuations.
[0051] S4. Furnace body zoning modeling and fusion based on structural differences: Based on the predicted furnace brick thickness distribution map and furnace body region division map, thickness inversion sub-models are trained according to the furnace body region division map. By cropping the temperature field image according to the preset region and inputting it into the corresponding sub-model, the optimized thickness map of each region is obtained. Then, the outputs of multiple models are weighted and fused by defining the region sensitivity and prediction confidence to output the thickness distribution map of the whole furnace. The furnace body area division diagram includes: copper tapping area, redox tapping area, sidewall area, and furnace bottom area; Because the erosion mechanisms differ significantly across different regions of the furnace body—for example, the copper tapping area is subjected to high-temperature melt and intense liquid flow, the oxidation-reduction tapping area is affected by severe chemical reactions and periodic thermal shocks, while the heat loads of the sidewall and furnace bottom areas are relatively stable—this invention employs a furnace body zoning modeling approach, training separate thickness inversion sub-models for each region. In this embodiment, the furnace exterior is divided into four functional regions: the copper tapping area, the oxidation-reduction tapping area, the sidewall area, and the furnace bottom area.
[0052] Pixel-level region label map used to divide each region The following methods were used to obtain the image: First, spatial calibration was performed based on the furnace body CAD structural drawing and the infrared camera imaging coordinate system. The projection boundaries of the copper tapping area, redox tapping area, sidewall area, and furnace bottom area in the infrared image were determined by manual annotation. Second, the high-temperature zone of the copper tapping area and the periodically high-fluctuation area of the redox tapping area were supplemented by combining the long-term temperature mean field, thereby obtaining a pixel-level region label map. The expression of the pixel-level region label map is as follows: These correspond to the copper tapping area, the redox tapping area, the sidewall area, and the furnace bottom area, respectively.
[0053] To address the differences in erosion scale across different regions, this invention jointly regulates the receptive field size by adjusting the input image preprocessing scale and the network structure depth, thereby constructing sub-models with different feature scales, as detailed below: The input scale of the copper mouth area sub-model is The network employs a 6-layer convolutional structure to characterize local erosion features dominated by small-scale, strong scouring. The input scale of the redox sub-model is The network uses a 7-layer convolutional structure; The input scale for both the sidewall sub-model and the furnace bottom sub-model is [scale value missing]. The network employs an 8-layer convolutional structure to characterize the thickness evolution features dominated by large-scale, slow erosion. Therefore, "having different receptive fields and feature scales" is reflected both in the multi-scale pyramid cropping method of the input image and in the differentiated design of the network structure depth.
[0054] The training data for each regional sub-model comes from: purification temperature sub-maps cropped according to region labels. Corresponding thickness label: The thickness of the furnace bricks was obtained by aligning the measured data from borehole drilling during the furnace shutdown and maintenance with the temperature field over time. In this embodiment, each regional sub-model was trained independently and its parameters were optimized separately. During the overall furnace thickness prediction stage, each sub-model only participated in joint inference and weighted fusion, and no further reverse updates were performed.
[0055] Define the region sensitivity coefficient This is used to represent the relative importance of different functional areas in furnace safety, and its specific value is: .
[0056] The thickness prediction results output by each sub-model are denoted as: And simultaneously output the prediction confidence map: The weighted fusion function expression is constructed as follows: The final overall furnace thickness distribution diagram is as follows: The aforementioned weighted fusion method takes into account both regional failure sensitivity and sub-model prediction confidence, thereby ensuring that the weight of key parts such as the copper gate region and redox gate region is automatically increased in the fusion result, so that the prediction result maintains both local accuracy and global consistency.
[0057] In this invention, the model's predicted output of the real-time temperature field is converted into a cloud map of the remaining thickness distribution of the furnace bricks, and overlaid in pseudo-color onto the schematic diagram of the furnace structure to form an intuitive thickness visualization interface. The colors, from red to blue, represent the thickness of the furnace bricks from thin to thick. The system continuously processes infrared temperature image data in the background, and the update frequency can be set according to the production rhythm. In this embodiment, a full furnace thickness distribution map can be automatically generated weekly or per batch. This visualization result can be used to identify local weak points in the copper tapping area and the redox tapping area in real time, and provide a reliable basis for subsequent analysis of the remaining life of the furnace bricks and maintenance decisions.
[0058] S5. Spatiotemporal evolution modeling and remaining life prediction: Based on the thickness inversion model, the thickness distribution map of the whole furnace and the minimum safe thickness, the thickness change sequence is constructed and input into the time series prediction model to predict the thickness at the next moment. Then, the thickness loss rate of the furnace bricks is calculated to predict the remaining life of the furnace body. like Figure 4As shown, based on the furnace brick thickness inversion model, this embodiment further analyzes the thickness data of continuous time series and uses a time series deep learning model to predict the furnace brick wear trend and remaining life.
[0059] This invention will monitor each point in time. The thickness is expressed as: in, For a moment Infrared temperature field, This is a thickness inversion model.
[0060] Constructing a thickness variation sequence: The thickness evolution is modeled using a Long Short-Term Memory (LSTM) network to predict the thickness at the next time step, as shown in the following expression: in, This is a time series prediction model.
[0061] Calculate the brick thickness loss rate based on the prediction results: in, It is a differential operator; Assume the minimum safe thickness of the furnace bricks is The formula for predicting the remaining life of the furnace body is: in, This represents the average loss rate due to recent thickness variations.
[0062] The system calculates the remaining operating time of the furnace body based on the furnace body remaining life prediction formula. A maintenance warning is triggered when the value is below a preset threshold.
[0063] S6. Based on the thickness inversion model and time series prediction model, a method and system for online thickness detection and lifetime prediction are completed through online drift monitoring and self-correction mechanism and online deployment and dynamic updating. This invention takes into account the changes in temperature distribution, brick surface condition, and sensor characteristics over time in actual production. It incorporates an online drift monitoring module into the system architecture to implement an online drift monitoring and self-correction mechanism. The system continuously monitors indicators such as temperature field statistics, gradient distribution, and eigenvector changes to determine if the data deviates from historical patterns. When significant drift is detected, the system adaptively corrects the inversion model through parameter fine-tuning, incremental training, or model rollback, enabling the system to maintain accuracy and stability in long-term, high-intensity industrial operating environments.
[0064] Online Deployment and Dynamic Updates: Through the synergy of the aforementioned multiple stages, this invention constructs a complete closed-loop system encompassing temperature acquisition and image purification, thickness inversion and regional fusion, and spatiotemporal prediction and self-correction. This system can achieve high-precision continuous thickness monitoring under harsh operating conditions and predict future lining changes, providing truly online and intelligent decision support for smelting operations, enabling online deployment and dynamic updates.
[0065] The overall system adopts a dual-model architecture of "spatial thickness inversion + time series prediction", as shown in the following expression: in, Infrared temperature field; That is, through the infrared temperature field and spatial thickness inversion model Invert the thickness, and then use the time series model. Predict lifespan.
[0066] Example 2: An online detection and life prediction system for furnace brick thickness in copper smelting anode furnaces based on multi-source infrared radiation; like Figure 3 As shown, the present invention also provides an online detection and life prediction system for furnace brick thickness of copper smelting anode furnace based on multi-source infrared radiation, and a method for implementing this system, including: The multi-view infrared acquisition module is used by S1 to simultaneously acquire multi-source temperature images of the outer wall of the furnace. The emissivity estimation and radiation compensation module is used by S1 to dynamically correct the emissivity of the acquired images; The occlusion detection and temperature purification module is used by S2 to eliminate interference from smoke, flames, reflected light and other factors, and to perform smoke occlusion identification and thermal radiation interference suppression. The thickness inversion module is used by S3 to construct a furnace brick thickness inversion model based on physical consistency constraints, including a deep mapping network with physical consistency constraints. The partition fusion module is used by S4 to uniformly fuse the inversion results of multiple regions; The spatiotemporal evolution prediction module is used to calculate the S5 thickness variation trend and furnace brick life. The model self-calibration module is used to automatically update model parameters during long-term operation of S6; The visualization and alarm module is used to output and display thickness cloud maps throughout the entire process and to perform lifespan warnings.
[0067] Compared with traditional finite element thermal inversion and manual inspection methods, this invention can realize non-contact, online, and intelligent furnace health monitoring, significantly improving equipment safety and operational reliability.
[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for online detection of furnace brick thickness and life prediction in copper smelting anode furnaces based on multi-source infrared radiation, characterized in that, Includes the following steps: S1. Multi-source infrared thermal field acquisition and dynamic emissivity compensation: Collect multi-view infrared imaging data and furnace brick material parameters. After feature point matching and registration based on the brightness consistency of multi-view images, construct a dynamic emissivity estimation model. Use background radiation temperature to correct the original infrared temperature measurement to obtain the multi-view compensated surface temperature field. Perform weighted fusion on the multi-view compensated surface temperature field to output the multi-view compensated temperature field. The multi-view infrared imaging data consists of furnace outer wall temperature images collected simultaneously by at least four different-view infrared thermal imagers. The material parameters of the furnace bricks include: thermal conductivity, density, and specific heat capacity; The key components are the copper port area and the redox port area; S2. Smoke Occlusion Recognition and Temperature Field Purification: Based on the multi-view compensated surface temperature field, the motion estimation of two adjacent frames is performed by using a dense optical flow algorithm, and the optical flow amplitude, root mean square contrast and luminance variance are calculated. After inputting into the infrared occlusion detection network to obtain the output feature map, the median interpolation of the adjacent frames within the time window is used by setting interference judgment to output the purified real temperature field. S3. Constructing a deep regression network for the physical consistency constraint module: Based on the purified real temperature field, a deep regression network is constructed to extract multi-scale thermal features. After embedding the physical consistency constraint module, the theoretical heat flux density is calculated based on the steady-state heat conduction relationship. The total loss function is then constructed for training, and the predicted furnace brick thickness distribution map is output, thus completing the construction of the furnace brick thickness inversion model based on the physical consistency constraint module. The total loss function includes the physical residual loss function, the data fitting loss function, and the spatial smoothing loss function; S4. Furnace body zoning modeling and fusion based on structural differences: Based on the predicted furnace brick thickness distribution map and furnace body region division map, thickness inversion sub-models are trained according to the furnace body region division map. By cropping the temperature field image according to the preset region and inputting it into the corresponding sub-model, the optimized thickness map of each region is obtained. Then, the outputs of multiple models are weighted and fused by defining the region sensitivity and prediction confidence to output the thickness distribution map of the whole furnace. The furnace body area division diagram includes: copper tapping area, redox tapping area, sidewall area, and furnace bottom area; S5. Spatiotemporal evolution modeling and remaining life prediction: Based on the thickness inversion model, the thickness distribution map of the whole furnace and the minimum safe thickness, the thickness change sequence is constructed and input into the time series prediction model to predict the thickness at the next moment. Then, the thickness loss rate of the furnace bricks is calculated to predict the remaining life of the furnace body. S6. Based on the thickness inversion model and time series prediction model, a method and system for online thickness detection and lifetime prediction are completed through online drift monitoring and self-correction mechanisms, online deployment and dynamic updates.
2. The method for online detection of furnace brick thickness and life prediction of copper smelting anode furnace based on multi-source infrared radiation according to claim 1, characterized in that, In S1, the expression for constructing the dynamic emissivity estimation model is: in, The reference emissivity is determined by the material parameters of the furnace bricks. These are the material parameters for the furnace bricks. This indicates the brightness difference of the corresponding pixels between this viewpoint and adjacent viewpoints after registration. The texture roughness index is calculated from local gray-level variance, gradient, etc. The first empirical coefficient, The second empirical coefficient, It represents three-dimensional space and time coordinates; After obtaining the dynamic emissivity estimation model, the background radiation temperature is used. The original infrared thermometry is corrected to obtain a multi-view compensated surface temperature field. The expression is: The multi-view compensated surface temperature field is obtained by weighted fusion. The expression is: in, The field of view number is assigned to the infrared camera. These are the weighting coefficients. This is the original temperature measurement field.
3. The method for online detection of furnace brick thickness and life prediction based on multi-source infrared radiation in a copper smelting anode furnace according to claim 1, characterized in that, In S2, the infrared occlusion detection network adopts an encoder-decoder structure. The encoder contains three convolutional modules, and each convolutional module consists of two layers. Convolution + ReLU activation + The maximum pooling composition has 32, 64, and 128 channels respectively. The decoder employs three upsampling modules, through a single layer... Convolution yields a 4-channel output feature map. , representing the pixel category probabilities of background, smoke, flame, and highly reflective areas, respectively, and the output is a probability map; During training, pixel-level interference masks with manual annotations are used as labels, and cross-entropy loss is employed for supervised learning.
4. The method for online detection of furnace brick thickness and life prediction based on multi-source infrared radiation in a copper smelting anode furnace according to claim 1, characterized in that, In S2, the interference determination is specifically set as follows: when the probability of the smoke pixel category... hour, The preset smoke detection threshold is used to determine if the smoke is blocked. When the probability of flame pixel category hour, A preset flame detection threshold is used to classify areas as flame regions; when the pixel category probability of a high-reflectivity area is... hour, It was determined to be a high-reflectivity point.
5. The method for online detection of furnace brick thickness and life prediction of a copper smelting anode furnace based on multi-source infrared radiation according to claim 1, characterized in that, In S3, the expression for calculating the theoretical heat flux density based on the steady-state heat conduction relationship is as follows: in, The thermal conductivity of the furnace brick material. The surface temperature field after purification. The temperature of the gas inside the furnace. Preliminary thickness prediction feature map; The total loss function is expressed as follows: in, The physical residual loss function, The loss function is used to fit the data. Let the spatial smoothing loss function be... The preset spatial smoothing loss weight coefficients are used. This is the preset weighting coefficient for physical residual loss.
6. The method for online detection of furnace brick thickness and life prediction based on multi-source infrared radiation in a copper smelting anode furnace according to claim 1, characterized in that, In S4, the sub-model includes a copper tapping area sub-model, a redox tapping area sub-model, a sidewall area sub-model, and a furnace bottom area sub-model; The input scale of the copper mouth area sub-model is The network employs a 6-layer convolutional structure to characterize local erosion features dominated by small-scale, strong scouring. The input scale of the redox sub-model is The network uses a 7-layer convolutional structure; The input scales for both the sidewall sub-model and the furnace bottom sub-model are... The network employs an 8-layer convolutional structure to characterize the thickness evolution features dominated by large-scale, slow erosion.
7. The method for online detection of furnace brick thickness and life prediction of a copper smelting anode furnace based on multi-source infrared radiation according to claim 1, characterized in that, In step S5, the expression for outputting the thickness distribution map of the entire furnace is: in, The thickness prediction result corresponding to the sub-model. The weighted fusion function has the following specific expression: in, For regional indexes, For the summation index, For the preset area sensitivity coefficient, For prediction confidence plot.
8. The method for online detection of furnace brick thickness and life prediction of copper smelting anode furnace based on multi-source infrared radiation according to claim 1, characterized in that, In step S6, the calculation of the furnace brick thickness loss rate and the prediction of the remaining furnace life are specifically as follows: Calculate the brick thickness loss rate based on the prediction results: in, For differential operators, For time The predicted thickness; Assume the minimum safe thickness of the furnace bricks is The formula for predicting the remaining life of the furnace body is: in, This represents the average loss rate due to recent thickness variations.
9. A method for online detection of furnace brick thickness and life prediction based on multi-source infrared radiation in a copper smelting anode furnace, as described in claim 1 or 5, characterized in that... In S6, the expression for the physical residual loss is: in, The total number of pixels in the training samples. For the first Predicted thickness per pixel The temperature of the gas inside the furnace. For the first The outer surface temperature at each pixel The thermal conductivity of the furnace bricks, for index, For reference heat flux density.
10. A system for online detection and lifetime prediction of furnace brick thickness in a copper smelting anode furnace based on multi-source infrared radiation, used to implement the method described in any one of claims 1-9, characterized in that, include: The multi-view infrared acquisition module is used by S1 to simultaneously acquire multi-source temperature images of the outer wall of the furnace. The emissivity estimation and radiation compensation module is used by S1 to dynamically correct the emissivity of the acquired images; The occlusion detection and temperature purification module is used by S2 to eliminate interference from smoke, flames, reflected light and other factors, and to perform smoke occlusion identification and thermal radiation interference suppression. The thickness inversion module is used by S3 to construct a furnace brick thickness inversion model based on physical consistency constraints, including a deep mapping network with physical consistency constraints. The partition fusion module is used by S4 to uniformly fuse the inversion results of multiple regions; The spatiotemporal evolution prediction module is used to calculate the S5 thickness variation trend and furnace brick life. The model self-calibration module is used to automatically update model parameters during long-term operation of S6; The visualization and alarm module is used to output and display thickness cloud maps throughout the entire process and to perform lifespan warnings.